https://www.hosaka.fi/wiki/api.php?action=feedcontributions&feedformat=atom&user=CslrHosaka/WhiteICE wiki - User contributions [en]2026-10-24T14:49:50ZUser contributionsMediaWiki 1.43.11https://www.hosaka.fi/wiki/index.php?title=Talk:Main_Page&diff=58Talk:Main Page2026-10-12T22:37:23Z<p>Cslr: </p>
<hr />
<div>Next version ideas:<br />
<br />
* try to find/do parameterized music generator and try to learn to optimize music to alter brainwaves and increase intelligence (+ try to stop remove control of brain)<br />
* ken spieker dancefloor euphoria was released to support my development (I also release music and released on track which tried to improve brain + digital sleeping pill ambient track)</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=Talk:Main_Page&diff=57Talk:Main Page2026-10-12T22:36:06Z<p>Cslr: </p>
<hr />
<div>Next version ideas:<br />
<br />
* try find parameterized music generator and try to learn to optimize music to alter brainwaves and increase intelligence (+ try to stop remove control of brain)<br />
* ken spieker dancefloor euphoria was released to support my development (I also release music and released on track which tried to improve brain + digital sleeping pill ambient track)</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=Talk:Main_Page&diff=56Talk:Main Page2026-10-12T22:34:27Z<p>Cslr: Created page with "Next version ideas: * try find parameterized music generator and try to learn to optimize music to alter brainwaves and increase intelligence (+ try to stop remove control of brain) * ken spieker dancefloor euphoria was released to support my development"</p>
<hr />
<div>Next version ideas:<br />
<br />
* try find parameterized music generator and try to learn to optimize music to alter brainwaves and increase intelligence (+ try to stop remove control of brain)<br />
* ken spieker dancefloor euphoria was released to support my development</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=55WhiteICE version information2026-09-17T06:03:50Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.43b (in development)</strong><br />
<br />
TODO: try different settings and configurations to increase performance measurements (need better results).<br />
<br />
Try very deep neural network (deepness=10). Try to DISABLE preprocessing in RIFL dataset creation (preprocessings are not properly handled???).<br />
<br />
Try at least shorter stimulation lenghts again (1000msec/500msec) and longer history (4 step history is SLOW but create complexity of stimulus..)<br />
<br />
<strong>WhiteICE v1.42b (most recent)</strong><br />
<br />
<strong>WhiteICE v1.41b</strong><br />
<br />
DONE: There is a bug in the code that causes windows GUI version of the software to crash in random locations. Lots of debugging needed: fixed many small bugs until error was found. Software should now be more stable.<br />
<br />
DONE: implement —lsl-stream-name=<string> to make it possible to read non-standard stream.<br />
<br />
DONE: fixing bugs created by ChatGPT fixing "bugs".<br />
<br />
DONE: test shorter stimulation lengths again. (500msecs/250msecs) => don't work well, back to (5000msecs/2500msecs).<br />
<br />
DONE: testing (minor fixes in reinforcement learning code done) [lsl concentration channels=4 and lsl intelligence channels=8 (not tested)] => one 149min concentration run gave -0.37% results - need to try to get performance fixed (change tau=0.1 to tau=0.2, deepness=3 to deepness=5 => lsl concentration run gave 0.3% effect and positive average change=4e-5 => seems to work a bit)<br />
<br />
DONE: try to set forgetting time to 30 minutes so only recent response-measurements have effect to reinforcement model????<br />
<br />
<strong>WhiteICE v1.40b</strong><br />
<br />
Added variances of target terms to RL model data. Now results are about same when neural network's width=64 with "lsl fft concentration" target than before 3-5% effect and positive effects otherwise too. This is enough for v1.40b. TODO: but still need test with "lsl fft intelligence eeg_channels=8" target to see if the most important target increases.<br />
<br />
Increased forgetting time from one hour to three hours so more data is available for optimization.<br />
<br />
DONE: fixed "--save-video" to be faster (worse compression) and not to run out of memory for buffering video frames for encoding.<br />
<br />
<strong>WhiteICE v1.39b</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms (2-3% effect). However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it) AND test 10 sec long stimulus again to see if there is better effect (+3.1% effect vs random and about zero effect to distance).<br />
<br />
<strong>FINAL</strong>: 5000 msecs stimulation time which often gives 3% result. Note that this is worse than in 1.38b which seem to give 5% result but RL algorithms in 1.39 are more sophisticated.<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=54WhiteICE version information2026-09-07T01:12:39Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.42b (in development)</strong><br />
<br />
TODO: try different settings and configurations to increase performance measurements (need better results).<br />
<br />
Try very deep neural network (deepness=10). Try to DISABLE preprocessing in RIFL dataset creation (preprocessings are not properly handled???).<br />
<br />
Try at least shorter stimulation lenghts again (1000msec/500msec) and longer history (4 step history is SLOW but create complexity of stimulus..)<br />
<br />
<strong>WhiteICE v1.41b (most recent)</strong><br />
<br />
DONE: There is a bug in the code that causes windows GUI version of the software to crash in random locations. Lots of debugging needed: fixed many small bugs until error was found. Software should now be more stable.<br />
<br />
DONE: implement —lsl-stream-name=<string> to make it possible to read non-standard stream.<br />
<br />
DONE: fixing bugs created by ChatGPT fixing "bugs".<br />
<br />
DONE: test shorter stimulation lengths again. (500msecs/250msecs) => don't work well, back to (5000msecs/2500msecs).<br />
<br />
DONE: testing (minor fixes in reinforcement learning code done) [lsl concentration channels=4 and lsl intelligence channels=8 (not tested)] => one 149min concentration run gave -0.37% results - need to try to get performance fixed (change tau=0.1 to tau=0.2, deepness=3 to deepness=5 => lsl concentration run gave 0.3% effect and positive average change=4e-5 => seems to work a bit)<br />
<br />
DONE: try to set forgetting time to 30 minutes so only recent response-measurements have effect to reinforcement model????<br />
<br />
<strong>WhiteICE v1.40b</strong><br />
<br />
Added variances of target terms to RL model data. Now results are about same when neural network's width=64 with "lsl fft concentration" target than before 3-5% effect and positive effects otherwise too. This is enough for v1.40b. TODO: but still need test with "lsl fft intelligence eeg_channels=8" target to see if the most important target increases.<br />
<br />
Increased forgetting time from one hour to three hours so more data is available for optimization.<br />
<br />
DONE: fixed "--save-video" to be faster (worse compression) and not to run out of memory for buffering video frames for encoding.<br />
<br />
<strong>WhiteICE v1.39b</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms (2-3% effect). However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it) AND test 10 sec long stimulus again to see if there is better effect (+3.1% effect vs random and about zero effect to distance).<br />
<br />
<strong>FINAL</strong>: 5000 msecs stimulation time which often gives 3% result. Note that this is worse than in 1.38b which seem to give 5% result but RL algorithms in 1.39 are more sophisticated.<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=53WhiteICE version information2026-09-06T23:27:15Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.42b (in development)</strong><br />
<br />
TODO: try different settings and configurations to increase performance measurements (need better results).<br />
<br />
Try very deep neural network (deepness=10). Try to DISABLE preprocessing in RIFL dataset creation (preprocessings are not properly handled???).<br />
<br />
Try at least shorter stimulation lenghts again (1000msec/500msec) and longer history (10 steps long history is SLOW???)<br />
<br />
<strong>WhiteICE v1.41b (most recent)</strong><br />
<br />
DONE: There is a bug in the code that causes windows GUI version of the software to crash in random locations. Lots of debugging needed: fixed many small bugs until error was found. Software should now be more stable.<br />
<br />
DONE: implement —lsl-stream-name=<string> to make it possible to read non-standard stream.<br />
<br />
DONE: fixing bugs created by ChatGPT fixing "bugs".<br />
<br />
DONE: test shorter stimulation lengths again. (500msecs/250msecs) => don't work well, back to (5000msecs/2500msecs).<br />
<br />
DONE: testing (minor fixes in reinforcement learning code done) [lsl concentration channels=4 and lsl intelligence channels=8 (not tested)] => one 149min concentration run gave -0.37% results - need to try to get performance fixed (change tau=0.1 to tau=0.2, deepness=3 to deepness=5 => lsl concentration run gave 0.3% effect and positive average change=4e-5 => seems to work a bit)<br />
<br />
DONE: try to set forgetting time to 30 minutes so only recent response-measurements have effect to reinforcement model????<br />
<br />
<strong>WhiteICE v1.40b</strong><br />
<br />
Added variances of target terms to RL model data. Now results are about same when neural network's width=64 with "lsl fft concentration" target than before 3-5% effect and positive effects otherwise too. This is enough for v1.40b. TODO: but still need test with "lsl fft intelligence eeg_channels=8" target to see if the most important target increases.<br />
<br />
Increased forgetting time from one hour to three hours so more data is available for optimization.<br />
<br />
DONE: fixed "--save-video" to be faster (worse compression) and not to run out of memory for buffering video frames for encoding.<br />
<br />
<strong>WhiteICE v1.39b</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms (2-3% effect). However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it) AND test 10 sec long stimulus again to see if there is better effect (+3.1% effect vs random and about zero effect to distance).<br />
<br />
<strong>FINAL</strong>: 5000 msecs stimulation time which often gives 3% result. Note that this is worse than in 1.38b which seem to give 5% result but RL algorithms in 1.39 are more sophisticated.<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=52WhiteICE version information2026-09-06T22:46:42Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.42b (in development)</strong><br />
<br />
TODO: try different settings and configurations to increase performance measurements (need better results).<br />
<br />
Try very deep neural network (deepness=10).<br />
<br />
Try at least shorter stimulation lenghts again (1000msec/500msec) and longer history (10 steps long history is SLOW???)<br />
<br />
<strong>WhiteICE v1.41b (most recent)</strong><br />
<br />
DONE: There is a bug in the code that causes windows GUI version of the software to crash in random locations. Lots of debugging needed: fixed many small bugs until error was found. Software should now be more stable.<br />
<br />
DONE: implement —lsl-stream-name=<string> to make it possible to read non-standard stream.<br />
<br />
DONE: fixing bugs created by ChatGPT fixing "bugs".<br />
<br />
DONE: test shorter stimulation lengths again. (500msecs/250msecs) => don't work well, back to (5000msecs/2500msecs).<br />
<br />
DONE: testing (minor fixes in reinforcement learning code done) [lsl concentration channels=4 and lsl intelligence channels=8 (not tested)] => one 149min concentration run gave -0.37% results - need to try to get performance fixed (change tau=0.1 to tau=0.2, deepness=3 to deepness=5 => lsl concentration run gave 0.3% effect and positive average change=4e-5 => seems to work a bit)<br />
<br />
DONE: try to set forgetting time to 30 minutes so only recent response-measurements have effect to reinforcement model????<br />
<br />
<strong>WhiteICE v1.40b</strong><br />
<br />
Added variances of target terms to RL model data. Now results are about same when neural network's width=64 with "lsl fft concentration" target than before 3-5% effect and positive effects otherwise too. This is enough for v1.40b. TODO: but still need test with "lsl fft intelligence eeg_channels=8" target to see if the most important target increases.<br />
<br />
Increased forgetting time from one hour to three hours so more data is available for optimization.<br />
<br />
DONE: fixed "--save-video" to be faster (worse compression) and not to run out of memory for buffering video frames for encoding.<br />
<br />
<strong>WhiteICE v1.39b</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms (2-3% effect). However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it) AND test 10 sec long stimulus again to see if there is better effect (+3.1% effect vs random and about zero effect to distance).<br />
<br />
<strong>FINAL</strong>: 5000 msecs stimulation time which often gives 3% result. Note that this is worse than in 1.38b which seem to give 5% result but RL algorithms in 1.39 are more sophisticated.<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=51WhiteICE version information2026-09-06T22:46:22Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.42b (in development)</strong><br />
<br />
TODO: try different settings and configurations to increase performance measurements (need better results).<br />
<br />
Try very deep neural network (deepness=10).<br />
<br />
Try at least shorter stimulation lenghts again (1000msec/500msec) and longer history (10 steps long history is SLOW???)<br />
<br />
<strong>WhiteICE v1.41b (most recent)</strong><br />
<br />
DONE: There is a bug in the code that causes windows GUI version of the software to crash in random locations. Lots of debugging needed: fixed many small bugs until error was found. Software should now be more stable.<br />
<br />
DONE: implement —lsl-stream-name=<string> to make it possible to read non-standard stream.<br />
<br />
DONE: fixing bugs created by ChatGPT fixing "bugs".<br />
<br />
DONE: test shorter stimulation lengths again. (500msecs/250msecs) => don't work well, back to (5000msecs/2500msecs).<br />
<br />
DONE: testing (minor fixes in reinforcement learning code done) [lsl concentration channels=4 and lsl intelligence channels=8 (not tested)] => one 149min concentration run gave -0.37% results - need to try to get performance fixed (change tau=0.1 to tau=0.2, deepness=3 to deepness=5 => lsl concentration run gave 2.7% effect and positive average change=4e-5 => seems to work a bit)<br />
<br />
DONE: try to set forgetting time to 30 minutes so only recent response-measurements have effect to reinforcement model????<br />
<br />
<strong>WhiteICE v1.40b</strong><br />
<br />
Added variances of target terms to RL model data. Now results are about same when neural network's width=64 with "lsl fft concentration" target than before 3-5% effect and positive effects otherwise too. This is enough for v1.40b. TODO: but still need test with "lsl fft intelligence eeg_channels=8" target to see if the most important target increases.<br />
<br />
Increased forgetting time from one hour to three hours so more data is available for optimization.<br />
<br />
DONE: fixed "--save-video" to be faster (worse compression) and not to run out of memory for buffering video frames for encoding.<br />
<br />
<strong>WhiteICE v1.39b</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms (2-3% effect). However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it) AND test 10 sec long stimulus again to see if there is better effect (+3.1% effect vs random and about zero effect to distance).<br />
<br />
<strong>FINAL</strong>: 5000 msecs stimulation time which often gives 3% result. Note that this is worse than in 1.38b which seem to give 5% result but RL algorithms in 1.39 are more sophisticated.<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=50WhiteICE version information2026-09-06T22:42:47Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.42b (in development)</strong><br />
<br />
TODO: try different settings and configurations to increase performance measurements (need better results).<br />
<br />
Try very deep neural network (deepness=10).<br />
<br />
Try at least shorter stimulation lenghts again (1000msec/500msec) and longer history (10 steps long history is SLOW???)<br />
<br />
<strong>WhiteICE v1.41b (most recent)</strong><br />
<br />
DONE: There is a bug in the code that causes windows GUI version of the software to crash in random locations. Lots of debugging needed: fixed many small bugs until error was found. Software should now be more stable.<br />
<br />
DONE: implement —lsl-stream-name=<string> to make it possible to read non-standard stream.<br />
<br />
DONE: fixing bugs created by ChatGPT fixing "bugs".<br />
<br />
DONE: test shorter stimulation lengths again. (500msecs/250msecs) => don't work well, back to (5000msecs/2500msecs).<br />
<br />
DONE: testing (minor fixes in reinforcement learning code done) [lsl concentration channels=4 and lsl intelligence channels=8 (not tested)] => one 149min concentration run gave -0.37% results - need to try to get performance fixed (change tau=0.1 to tau=0.2, deepness=3 to deepness=5 => lsl concentration run gave 2.7% effect and positive average change => seems to work a bit)<br />
<br />
DONE: try to set forgetting time to 30 minutes so only recent response-measurements have effect to reinforcement model????<br />
<br />
<strong>WhiteICE v1.40b</strong><br />
<br />
Added variances of target terms to RL model data. Now results are about same when neural network's width=64 with "lsl fft concentration" target than before 3-5% effect and positive effects otherwise too. This is enough for v1.40b. TODO: but still need test with "lsl fft intelligence eeg_channels=8" target to see if the most important target increases.<br />
<br />
Increased forgetting time from one hour to three hours so more data is available for optimization.<br />
<br />
DONE: fixed "--save-video" to be faster (worse compression) and not to run out of memory for buffering video frames for encoding.<br />
<br />
<strong>WhiteICE v1.39b</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms (2-3% effect). However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it) AND test 10 sec long stimulus again to see if there is better effect (+3.1% effect vs random and about zero effect to distance).<br />
<br />
<strong>FINAL</strong>: 5000 msecs stimulation time which often gives 3% result. Note that this is worse than in 1.38b which seem to give 5% result but RL algorithms in 1.39 are more sophisticated.<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=49WhiteICE version information2026-09-06T22:30:32Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.42b (in development)</strong><br />
<br />
TODO: try different settings and configurations to increase performance measurements (need better results).<br />
<br />
Try at least shorter stimulation lenghts again (1000msec/500msec) and longer history (10 steps long history is SLOW???)<br />
<br />
<strong>WhiteICE v1.41b (most recent)</strong><br />
<br />
DONE: There is a bug in the code that causes windows GUI version of the software to crash in random locations. Lots of debugging needed: fixed many small bugs until error was found. Software should now be more stable.<br />
<br />
DONE: implement —lsl-stream-name=<string> to make it possible to read non-standard stream.<br />
<br />
DONE: fixing bugs created by ChatGPT fixing "bugs".<br />
<br />
DONE: test shorter stimulation lengths again. (500msecs/250msecs) => don't work well, back to (5000msecs/2500msecs).<br />
<br />
DONE: testing (minor fixes in reinforcement learning code done) [lsl concentration channels=4 and lsl intelligence channels=8 (not tested)] => one 149min concentration run gave -0.37% results - need to try to get performance fixed (change tau=0.1 to tau=0.2, deepness=3 to deepness=5 => lsl concentration run gave 2.7% effect and positive average change => seems to work a bit)<br />
<br />
DONE: try to set forgetting time to 30 minutes so only recent response-measurements have effect to reinforcement model????<br />
<br />
<strong>WhiteICE v1.40b</strong><br />
<br />
Added variances of target terms to RL model data. Now results are about same when neural network's width=64 with "lsl fft concentration" target than before 3-5% effect and positive effects otherwise too. This is enough for v1.40b. TODO: but still need test with "lsl fft intelligence eeg_channels=8" target to see if the most important target increases.<br />
<br />
Increased forgetting time from one hour to three hours so more data is available for optimization.<br />
<br />
DONE: fixed "--save-video" to be faster (worse compression) and not to run out of memory for buffering video frames for encoding.<br />
<br />
<strong>WhiteICE v1.39b</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms (2-3% effect). However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it) AND test 10 sec long stimulus again to see if there is better effect (+3.1% effect vs random and about zero effect to distance).<br />
<br />
<strong>FINAL</strong>: 5000 msecs stimulation time which often gives 3% result. Note that this is worse than in 1.38b which seem to give 5% result but RL algorithms in 1.39 are more sophisticated.<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=48WhiteICE version information2026-09-06T22:26:05Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.42b (in development)</strong><br />
<br />
TODO: try different settings and configurations to increase performance measurements (need better results).<br />
<br />
<strong>WhiteICE v1.41b (most recent)</strong><br />
<br />
DONE: There is a bug in the code that causes windows GUI version of the software to crash in random locations. Lots of debugging needed: fixed many small bugs until error was found. Software should now be more stable.<br />
<br />
DONE: implement —lsl-stream-name=<string> to make it possible to read non-standard stream.<br />
<br />
DONE: fixing bugs created by ChatGPT fixing "bugs".<br />
<br />
DONE: test shorter stimulation lengths again. (500msecs/250msecs) => don't work well, back to (5000msecs/2500msecs).<br />
<br />
DONE: testing (minor fixes in reinforcement learning code done) [lsl concentration channels=4 and lsl intelligence channels=8 (not tested)] => one 149min concentration run gave -0.37% results - need to try to get performance fixed (change tau=0.1 to tau=0.2, deepness=3 to deepness=5 => lsl concentration run gave 2.7% effect and positive average change => seems to work a bit)<br />
<br />
DONE: try to set forgetting time to 30 minutes so only recent response-measurements have effect to reinforcement model????<br />
<br />
<strong>WhiteICE v1.40b</strong><br />
<br />
Added variances of target terms to RL model data. Now results are about same when neural network's width=64 with "lsl fft concentration" target than before 3-5% effect and positive effects otherwise too. This is enough for v1.40b. TODO: but still need test with "lsl fft intelligence eeg_channels=8" target to see if the most important target increases.<br />
<br />
Increased forgetting time from one hour to three hours so more data is available for optimization.<br />
<br />
DONE: fixed "--save-video" to be faster (worse compression) and not to run out of memory for buffering video frames for encoding.<br />
<br />
<strong>WhiteICE v1.39b</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms (2-3% effect). However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it) AND test 10 sec long stimulus again to see if there is better effect (+3.1% effect vs random and about zero effect to distance).<br />
<br />
<strong>FINAL</strong>: 5000 msecs stimulation time which often gives 3% result. Note that this is worse than in 1.38b which seem to give 5% result but RL algorithms in 1.39 are more sophisticated.<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=47WhiteICE version information2026-09-06T21:47:38Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.41b (in development)</strong><br />
<br />
DONE: There is a bug in the code that causes windows GUI version of the software to crash in random locations. Lots of debugging needed: fixed many small bugs until error was found. Software should now be more stable.<br />
<br />
DONE: implement —lsl-stream-name=<string> to make it possible to read non-standard stream.<br />
<br />
DONE: fixing bugs created by ChatGPT fixing "bugs".<br />
<br />
DONE: test shorter stimulation lengths again. (500msecs/250msecs) => don't work well, back to (5000msecs/2500msecs).<br />
<br />
DONE: testing (minor fixes in reinforcement learning code done) [lsl concentration channels=4 and lsl intelligence channels=8 (not tested)] => one 149min concentration run gave -0.37% results - need to try to get performance fixed (change tau=0.1 to tau=0.2, deepness=3 to deepness=5 => lsl concentration run gave 2.7% effect and positive average change => seems to work a bit)<br />
<br />
DONE: try to set forgetting time to 30 minutes so only recent response-measurements have effect to reinforcement model????<br />
<br />
<strong>WhiteICE v1.40b (most recent)</strong><br />
<br />
Added variances of target terms to RL model data. Now results are about same when neural network's width=64 with "lsl fft concentration" target than before 3-5% effect and positive effects otherwise too. This is enough for v1.40b. TODO: but still need test with "lsl fft intelligence eeg_channels=8" target to see if the most important target increases.<br />
<br />
Increased forgetting time from one hour to three hours so more data is available for optimization.<br />
<br />
DONE: fixed "--save-video" to be faster (worse compression) and not to run out of memory for buffering video frames for encoding.<br />
<br />
<strong>WhiteICE v1.39b</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms (2-3% effect). However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it) AND test 10 sec long stimulus again to see if there is better effect (+3.1% effect vs random and about zero effect to distance).<br />
<br />
<strong>FINAL</strong>: 5000 msecs stimulation time which often gives 3% result. Note that this is worse than in 1.38b which seem to give 5% result but RL algorithms in 1.39 are more sophisticated.<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=46WhiteICE version information2026-09-06T20:50:07Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.41b (in development)</strong><br />
<br />
DONE: There is a bug in the code that causes windows GUI version of the software to crash in random locations. Lots of debugging needed: fixed many small bugs until error was found. Software should now be more stable.<br />
<br />
DONE: implement —lsl-stream-name=<string> to make it possible to read non-standard stream.<br />
<br />
DONE: fixing bugs created by ChatGPT fixing "bugs".<br />
<br />
DONE: test shorter stimulation lengths again. (500msecs/250msecs) => don't work well, back to (5000msecs/2500msecs).<br />
<br />
TODO: testing (minor fixes in reinforcement learning code done) [lsl concentration channels=4 and lsl intelligence channels=8] => one 149min concentration run gave -0.37% results - need to try to get performance fixed (change tau=0.1 to tau=0.2, deepness=3 to deepness=5 ????)<br />
<br />
TODO: try to set forgetting time to 30 minutes so only recent response-measurements have effect to reinforcement model????<br />
<br />
<strong>WhiteICE v1.40b (most recent)</strong><br />
<br />
Added variances of target terms to RL model data. Now results are about same when neural network's width=64 with "lsl fft concentration" target than before 3-5% effect and positive effects otherwise too. This is enough for v1.40b. TODO: but still need test with "lsl fft intelligence eeg_channels=8" target to see if the most important target increases.<br />
<br />
Increased forgetting time from one hour to three hours so more data is available for optimization.<br />
<br />
DONE: fixed "--save-video" to be faster (worse compression) and not to run out of memory for buffering video frames for encoding.<br />
<br />
<strong>WhiteICE v1.39b</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms (2-3% effect). However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it) AND test 10 sec long stimulus again to see if there is better effect (+3.1% effect vs random and about zero effect to distance).<br />
<br />
<strong>FINAL</strong>: 5000 msecs stimulation time which often gives 3% result. Note that this is worse than in 1.38b which seem to give 5% result but RL algorithms in 1.39 are more sophisticated.<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=45WhiteICE version information2026-09-06T20:05:35Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.41b (in development)</strong><br />
<br />
DONE: There is a bug in the code that causes windows GUI version of the software to crash in random locations. Lots of debugging needed: fixed many small bugs until error was found. Software should now be more stable.<br />
<br />
DONE: implement —lsl-stream-name=<string> to make it possible to read non-standard stream.<br />
<br />
DONE: fixing bugs created by ChatGPT fixing "bugs".<br />
<br />
DONE: test shorter stimulation lengths again. (500msecs/250msecs) => don't work well, back to (5000msecs/2500msecs).<br />
<br />
TODO: testing (minor fixes in reinforcement learning code done) [lsl concentration channels=4 and lsl intelligence channels=8] => one 149min concentration run gave -0.37% results - need to try to get performance fixed (change tau=0.1 to tau=0.2, deepness=3 to deepness=5 ????)<br />
<br />
<strong>WhiteICE v1.40b (most recent)</strong><br />
<br />
Added variances of target terms to RL model data. Now results are about same when neural network's width=64 with "lsl fft concentration" target than before 3-5% effect and positive effects otherwise too. This is enough for v1.40b. TODO: but still need test with "lsl fft intelligence eeg_channels=8" target to see if the most important target increases.<br />
<br />
Increased forgetting time from one hour to three hours so more data is available for optimization.<br />
<br />
DONE: fixed "--save-video" to be faster (worse compression) and not to run out of memory for buffering video frames for encoding.<br />
<br />
<strong>WhiteICE v1.39b</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms (2-3% effect). However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it) AND test 10 sec long stimulus again to see if there is better effect (+3.1% effect vs random and about zero effect to distance).<br />
<br />
<strong>FINAL</strong>: 5000 msecs stimulation time which often gives 3% result. Note that this is worse than in 1.38b which seem to give 5% result but RL algorithms in 1.39 are more sophisticated.<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=44WhiteICE version information2026-09-06T19:38:50Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.41b (in development)</strong><br />
<br />
DONE: There is a bug in the code that causes windows GUI version of the software to crash in random locations. Lots of debugging needed: fixed many small bugs until error was found. Software should now be more stable.<br />
<br />
DONE: implement —lsl-stream-name=<string> to make it possible to read non-standard stream.<br />
<br />
DONE: fixing bugs created by ChatGPT fixing "bugs".<br />
<br />
DONE: test shorter stimulation lengths again. (500msecs/250msecs) => don't work well, back to (5000msecs/2500msecs).<br />
<br />
TODO: testing (minor fixes in reinforcement learning code done) [lsl concentration channels=4 and lsl intelligence channels=8] => one 149min concentration run gave -0.37% results - need to try to get performance fixed (change tau=0.1 to tau=0.2 ????)<br />
<br />
<strong>WhiteICE v1.40b (most recent)</strong><br />
<br />
Added variances of target terms to RL model data. Now results are about same when neural network's width=64 with "lsl fft concentration" target than before 3-5% effect and positive effects otherwise too. This is enough for v1.40b. TODO: but still need test with "lsl fft intelligence eeg_channels=8" target to see if the most important target increases.<br />
<br />
Increased forgetting time from one hour to three hours so more data is available for optimization.<br />
<br />
DONE: fixed "--save-video" to be faster (worse compression) and not to run out of memory for buffering video frames for encoding.<br />
<br />
<strong>WhiteICE v1.39b</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms (2-3% effect). However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it) AND test 10 sec long stimulus again to see if there is better effect (+3.1% effect vs random and about zero effect to distance).<br />
<br />
<strong>FINAL</strong>: 5000 msecs stimulation time which often gives 3% result. Note that this is worse than in 1.38b which seem to give 5% result but RL algorithms in 1.39 are more sophisticated.<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=Measurements&diff=43Measurements2026-09-06T18:51:57Z<p>Cslr: </p>
<hr />
<div>This page describes what different reported WhiteICE performance numbers means.<br />
<br />
* the most important performance number is "average change". It tells how much non-random policy actions on the average change reinforcement reward (change of reward), positive values mean the software is working correctly<br />
* "policy vas random percentage" tells how much better in percentages (%) policy actions (non-random) actions are when compared random stimuli<br />
* finally there is "linear distance change" that tell how much better linear model final value (distance of policy) is than linear predicted value of random final values</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=Main_Page&diff=42Main Page2026-09-06T18:51:10Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE information</strong><br />
<br />
Idea of the software is that you run the sound stimulation in background while using a computer. Stimulation requires many hours of measurements before it is effective. Sounds used are very simple because the learning starts from scratch (TODO: collect a database of measurements and models that can be used without a long training time + generate music instead of plain sounds).<br />
<br />
Command-line version of WhiteICE has more options. You start listening a LSL stream from EEG in a local network by using a command:<br />
<br />
<code><br />
whiteice --measure --device=lsl --concentration --lsl-channels=4<br />
</code><br />
<br />
Or for increasing intelligence (should have 8 channel EEG for good results, maximizing certain approximate small world index value):<br />
<br />
<code><br />
whiteice --measure --device=lsl --intelligence --lsl-channels=8<br />
</code><br />
<br />
The command plays sounds in the background (while using a computer), tries to increase concentration and uses first 4 channels found from the first LSL stream found from the network. If you are using "BrainAccess MINI EEG" for example which has 8 channels, you use command "--lsl-channels=8". The default number of LSL channels is 4. LSL stream listened is currently either "EEG", "Eeg" or "eeg" but this will be become configurable later.<br />
<br />
In practice you need to run software for hours to get small average positive effect to target values.<br />
<br />
If you have Interaxon Muse instead, you must configure MindMonitor app to send OSC stream to your computers IP and port 4545 and use a command (the number of channels is always 4):<br />
<br />
<code><br />
whiteice --measure --device=mindmonitor --concentration<br />
</code><br />
<br />
Also read: [[WhiteICE version information]], [[Measurements]]<br />
<br />
<strong>Hosaka information</strong><br />
<br />
Hosaka is currently not actively developed because it does off-line optimization of larger machine learning model (you must first measure responses for hours, then compute model for hours and finally can use non-adapting model with real data). However, currently Hosaka supports using of different pictures as stimulations which feature vectors will be computed for machine learning prediction of responses to pictures.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=Measurements&diff=41Measurements2026-09-06T18:50:47Z<p>Cslr: Created page with "This page described what different reported performance numbers mean. * the most important performance number is "average change". It tells how much non-random policy actions on the average change reinforcement reward (change of reward), positive values mean the software is working correctly * "policy vas random percentage" tells how much better in percentages (%) policy actions (non-random) actions are when compared random stimuli * finally there is "linear distance ch..."</p>
<hr />
<div>This page described what different reported performance numbers mean.<br />
<br />
* the most important performance number is "average change". It tells how much non-random policy actions on the average change reinforcement reward (change of reward), positive values mean the software is working correctly<br />
* "policy vas random percentage" tells how much better in percentages (%) policy actions (non-random) actions are when compared random stimuli<br />
* finally there is "linear distance change" that tell how much better linear model final value (distance of policy) is than linear predicted value of random final values</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=40WhiteICE version information2026-09-06T16:54:04Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.41b (in development)</strong><br />
<br />
DONE: There is a bug in the code that causes windows GUI version of the software to crash in random locations. Lots of debugging needed: fixed many small bugs until error was found. Software should now be more stable.<br />
<br />
DONE: implement —lsl-stream-name=<string> to make it possible to read non-standard stream.<br />
<br />
DONE: fixing bugs created by ChatGPT fixing "bugs".<br />
<br />
DONE: test shorter stimulation lengths again. (500msecs/250msecs) => don't work well, back to (5000msecs/2500msecs).<br />
<br />
TODO: testing (minor fixes in reinforcement learning code done) [lsl concentration channels=4 and lsl intelligence channels=8]<br />
<br />
<strong>WhiteICE v1.40b (most recent)</strong><br />
<br />
Added variances of target terms to RL model data. Now results are about same when neural network's width=64 with "lsl fft concentration" target than before 3-5% effect and positive effects otherwise too. This is enough for v1.40b. TODO: but still need test with "lsl fft intelligence eeg_channels=8" target to see if the most important target increases.<br />
<br />
Increased forgetting time from one hour to three hours so more data is available for optimization.<br />
<br />
DONE: fixed "--save-video" to be faster (worse compression) and not to run out of memory for buffering video frames for encoding.<br />
<br />
<strong>WhiteICE v1.39b</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms (2-3% effect). However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it) AND test 10 sec long stimulus again to see if there is better effect (+3.1% effect vs random and about zero effect to distance).<br />
<br />
<strong>FINAL</strong>: 5000 msecs stimulation time which often gives 3% result. Note that this is worse than in 1.38b which seem to give 5% result but RL algorithms in 1.39 are more sophisticated.<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=39WhiteICE version information2026-09-06T16:02:59Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.41b (in development)</strong><br />
<br />
DONE: There is a bug in the code that causes windows GUI version of the software to crash in random locations. Lots of debugging needed: fixed many small bugs until error was found. Software should now be more stable.<br />
<br />
DONE: implement —lsl-stream-name=<string> to make it possible to read non-standard stream.<br />
<br />
DONE: fixing bugs created by ChatGPT fixing "bugs".<br />
<br />
TODO: testing (minor fixes in reinforcement learning code done)<br />
<br />
DONE: test shorter stimulation lengths again. (500msecs/250msecs) => don't work well, back to (5000msecs/2500msecs).<br />
<br />
<strong>WhiteICE v1.40b (most recent)</strong><br />
<br />
Added variances of target terms to RL model data. Now results are about same when neural network's width=64 with "lsl fft concentration" target than before 3-5% effect and positive effects otherwise too. This is enough for v1.40b. TODO: but still need test with "lsl fft intelligence eeg_channels=8" target to see if the most important target increases.<br />
<br />
Increased forgetting time from one hour to three hours so more data is available for optimization.<br />
<br />
DONE: fixed "--save-video" to be faster (worse compression) and not to run out of memory for buffering video frames for encoding.<br />
<br />
<strong>WhiteICE v1.39b</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms (2-3% effect). However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it) AND test 10 sec long stimulus again to see if there is better effect (+3.1% effect vs random and about zero effect to distance).<br />
<br />
<strong>FINAL</strong>: 5000 msecs stimulation time which often gives 3% result. Note that this is worse than in 1.38b which seem to give 5% result but RL algorithms in 1.39 are more sophisticated.<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=38WhiteICE version information2026-09-06T15:14:36Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.41b (in development)</strong><br />
<br />
DONE: There is a bug in the code that causes windows GUI version of the software to crash in random locations. Lots of debugging needed: fixed many small bugs until error was found. Software should now be more stable.<br />
<br />
DONE: implement —lsl-stream-name=<string> to make it possible to read non-standard stream.<br />
<br />
DONE: fixing bugs created by ChatGPT fixing "bugs".<br />
<br />
TODO: testing (minor fixes in reinforcement learning code done)<br />
<br />
TODO: test shorter stimulation lengths again. (500msecs/250msecs)<br />
<br />
<strong>WhiteICE v1.40b (most recent)</strong><br />
<br />
Added variances of target terms to RL model data. Now results are about same when neural network's width=64 with "lsl fft concentration" target than before 3-5% effect and positive effects otherwise too. This is enough for v1.40b. TODO: but still need test with "lsl fft intelligence eeg_channels=8" target to see if the most important target increases.<br />
<br />
Increased forgetting time from one hour to three hours so more data is available for optimization.<br />
<br />
DONE: fixed "--save-video" to be faster (worse compression) and not to run out of memory for buffering video frames for encoding.<br />
<br />
<strong>WhiteICE v1.39b</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms (2-3% effect). However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it) AND test 10 sec long stimulus again to see if there is better effect (+3.1% effect vs random and about zero effect to distance).<br />
<br />
<strong>FINAL</strong>: 5000 msecs stimulation time which often gives 3% result. Note that this is worse than in 1.38b which seem to give 5% result but RL algorithms in 1.39 are more sophisticated.<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=37WhiteICE version information2026-09-06T11:30:46Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.42b (in development)</strong><br />
<br />
TODO: test shorter stimulation lengths again. (500msecs/250msecs)<br />
<br />
<strong>WhiteICE v1.41b (in development)</strong><br />
<br />
DONE: There is a bug in the code that causes windows GUI version of the software to crash in random locations. Lots of debugging needed: fixed many small bugs until error was found. Software should now be more stable.<br />
<br />
DONE: implement —lsl-stream-name=<string> to make it possible to read non-standard stream.<br />
<br />
TODO: testing (minor fixes in reinforcement learning code done)<br />
<br />
<strong>WhiteICE v1.40b (most recent)</strong><br />
<br />
Added variances of target terms to RL model data. Now results are about same when neural network's width=64 with "lsl fft concentration" target than before 3-5% effect and positive effects otherwise too. This is enough for v1.40b. TODO: but still need test with "lsl fft intelligence eeg_channels=8" target to see if the most important target increases.<br />
<br />
Increased forgetting time from one hour to three hours so more data is available for optimization.<br />
<br />
DONE: fixed "--save-video" to be faster (worse compression) and not to run out of memory for buffering video frames for encoding.<br />
<br />
<strong>WhiteICE v1.39b</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms (2-3% effect). However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it) AND test 10 sec long stimulus again to see if there is better effect (+3.1% effect vs random and about zero effect to distance).<br />
<br />
<strong>FINAL</strong>: 5000 msecs stimulation time which often gives 3% result. Note that this is worse than in 1.38b which seem to give 5% result but RL algorithms in 1.39 are more sophisticated.<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=36WhiteICE version information2026-09-06T02:22:04Z<p>Cslr: bug fixes</p>
<hr />
<div><strong>WhiteICE v1.41b (in development)</strong><br />
<br />
DONE: There is a bug in the code that causes windows GUI version of the software to crash in random locations. Lots of debugging needed: fixed many small bugs until error was found. Software should now be more stable.<br />
<br />
TODO: implement —lsl-stream-name=<string> to make it possible to read non-standard stream.<br />
<br />
<strong>WhiteICE v1.40b (most recent)</strong><br />
<br />
Added variances of target terms to RL model data. Now results are about same when neural network's width=64 with "lsl fft concentration" target than before 3-5% effect and positive effects otherwise too. This is enough for v1.40b. TODO: but still need test with "lsl fft intelligence eeg_channels=8" target to see if the most important target increases.<br />
<br />
Increased forgetting time from one hour to three hours so more data is available for optimization.<br />
<br />
DONE: fixed "--save-video" to be faster (worse compression) and not to run out of memory for buffering video frames for encoding.<br />
<br />
<strong>WhiteICE v1.39b</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms (2-3% effect). However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it) AND test 10 sec long stimulus again to see if there is better effect (+3.1% effect vs random and about zero effect to distance).<br />
<br />
<strong>FINAL</strong>: 5000 msecs stimulation time which often gives 3% result. Note that this is worse than in 1.38b which seem to give 5% result but RL algorithms in 1.39 are more sophisticated.<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=35WhiteICE version information2026-09-04T20:56:18Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.41b (in development)</strong><br />
<br />
TODO: There is a bug in the code that causes windows GUI version of the software to crash in random locations. Lots of debugging needed.<br />
<br />
TODO: implement —lsl-stream-name=<string> to make it possible to read non-standard stream.<br />
<br />
<strong>WhiteICE v1.40b (most recent)</strong><br />
<br />
Added variances of target terms to RL model data. Now results are about same when neural network's width=64 with "lsl fft concentration" target than before 3-5% effect and positive effects otherwise too. This is enough for v1.40b. TODO: but still need test with "lsl fft intelligence eeg_channels=8" target to see if the most important target increases.<br />
<br />
Increased forgetting time from one hour to three hours so more data is available for optimization.<br />
<br />
DONE: fixed "--save-video" to be faster (worse compression) and not to run out of memory for buffering video frames for encoding.<br />
<br />
<strong>WhiteICE v1.39b</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms (2-3% effect). However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it) AND test 10 sec long stimulus again to see if there is better effect (+3.1% effect vs random and about zero effect to distance).<br />
<br />
<strong>FINAL</strong>: 5000 msecs stimulation time which often gives 3% result. Note that this is worse than in 1.38b which seem to give 5% result but RL algorithms in 1.39 are more sophisticated.<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=34WhiteICE version information2026-09-03T13:38:06Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.41b (in development)</strong><br />
<br />
TODO: There is a bug in a code that causes windows GUI version of the software to crash in random locations. Lots of debugging needed.<br />
<br />
TODO: implement —lsl-stream-name=<string> to make it possible to read non-standard stream.<br />
<br />
<strong>WhiteICE v1.40b (most recent)</strong><br />
<br />
Added variances of target terms to RL model data. Now results are about same when neural network's width=64 with "lsl fft concentration" target than before 3-5% effect and positive effects otherwise too. This is enough for v1.40b. TODO: but still need test with "lsl fft intelligence eeg_channels=8" target to see if the most important target increases.<br />
<br />
Increased forgetting time from one hour to three hours so more data is available for optimization.<br />
<br />
DONE: fixed "--save-video" to be faster (worse compression) and not to run out of memory for buffering video frames for encoding.<br />
<br />
<strong>WhiteICE v1.39b</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms (2-3% effect). However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it) AND test 10 sec long stimulus again to see if there is better effect (+3.1% effect vs random and about zero effect to distance).<br />
<br />
<strong>FINAL</strong>: 5000 msecs stimulation time which often gives 3% result. Note that this is worse than in 1.38b which seem to give 5% result but RL algorithms in 1.39 are more sophisticated.<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=Main_Page&diff=33Main Page2026-09-03T08:11:58Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE information</strong><br />
<br />
Idea of the software is that you run the sound stimulation in background while using a computer. Stimulation requires many hours of measurements before it is effective. Sounds used are very simple because the learning starts from scratch (TODO: collect a database of measurements and models that can be used without a long training time + generate music instead of plain sounds).<br />
<br />
Command-line version of WhiteICE has more options. You start listening a LSL stream from EEG in a local network by using a command:<br />
<br />
<code><br />
whiteice --measure --device=lsl --concentration --lsl-channels=4<br />
</code><br />
<br />
Or for increasing intelligence (should have 8 channel EEG for good results, maximizing certain approximate small world index value):<br />
<br />
<code><br />
whiteice --measure --device=lsl --intelligence --lsl-channels=8<br />
</code><br />
<br />
The command plays sounds in the background (while using a computer), tries to increase concentration and uses first 4 channels found from the first LSL stream found from the network. If you are using "BrainAccess MINI EEG" for example which has 8 channels, you use command "--lsl-channels=8". The default number of LSL channels is 4. LSL stream listened is currently either "EEG", "Eeg" or "eeg" but this will be become configurable later.<br />
<br />
In practice you need to run software for hours to get small average positive effect to target values.<br />
<br />
If you have Interaxon Muse instead, you must configure MindMonitor app to send OSC stream to your computers IP and port 4545 and use a command (the number of channels is always 4):<br />
<br />
<code><br />
whiteice --measure --device=mindmonitor --concentration<br />
</code><br />
<br />
Also read: [[WhiteICE version information]]<br />
<br />
<strong>Hosaka information</strong><br />
<br />
Hosaka is currently not actively developed because it does off-line optimization of larger machine learning model (you must first measure responses for hours, then compute model for hours and finally can use non-adapting model with real data). However, currently Hosaka supports using of different pictures as stimulations which feature vectors will be computed for machine learning prediction of responses to pictures.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=Main_Page&diff=32Main Page2026-09-03T08:11:12Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE information</strong><br />
<br />
Idea of the software is that you run the sound stimulation in background while using a computer. Stimulation requires many hours of measurements before it is effective. Sounds used are very simple because the learning starts from scratch (TODO: collect a database of measurements and models that can be used without a long training time + generate music instead of plain sounds).<br />
<br />
Command-line version of WhiteICE has more options. You start listening a LSL stream from EEG in a local network by using a command:<br />
<br />
<code><br />
whiteice --measure --device=lsl --concentration --lsl-channels=4<br />
</code><br />
<br />
Or for increasing intelligence (should have 8 channel EEG for good results, maximizing certain approximate small world index value):<br />
<br />
<code><br />
whiteice --measure --device=lsl --intelligence --lsl-channels=8<br />
</code><br />
<br />
The command plays sounds in the background (while using a computer), tries to increase concentration and uses first 4 channels found from the first LSL stream found from the network. If you are using "BrainAccess MINI EEG" for example which has 8 channels, you use command "--lsl-channels=8". The default number of LSL channels is 4. LSL stream listened is currently either "EEG", "Eeg" or "eeg" but this will be become configurable later.<br />
<br />
In practice you need to run software for hours to get small average positive effect to target values.<br />
<br />
If you have Interaxon Muse instead, you must configure MindMonitor app to send OSC stream to your computers IP and port 4545 and use a command (the number of channels is always 4):<br />
<br />
<code><br />
whiteice --measure --device=mindmonitor --concentration<br />
</code><br />
<br />
Also read: [[WhiteICE version information]]<br />
<br />
<strong>Hosaka information</strong><br />
<br />
Hosaka is currently not actively developed because it does off-line optimization of larger machine learning model (you must first measure responses for hours, then compute model for hours and finally can use non-adapting model with real data). However, currently Hosaka supports using of different pictures are stimulations which feature vector will be computed for machine learning predition of responses to pictures.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=Main_Page&diff=31Main Page2026-09-03T08:10:49Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE information</strong><br />
<br />
Idea of the software is that you run the sound stimulation in background while using a computer. Stimulation requires many hours of measurements before it is effective. Sounds used are very simple because the learning starts from scratch (TODO: collect a database of measurements and models that can be used without a long training time + generate music instead of plain sounds).<br />
<br />
Command-line version of WhiteICE has more options. You start listening a LSL stream from EEG in a local network by using a command:<br />
<br />
<code><br />
whiteice --measure --device=lsl --concentration --lsl-channels=4<br />
</code><br />
<br />
Or for increasing intelligence (should have 8 channel EEG for good results, maximizing certain approximate small world index value):<br />
<br />
<code><br />
whiteice --measure --device=lsl --intelligence --lsl-channels=8<br />
</code><br />
<br />
The command plays sounds in the background (while using a computer), tries to increase concentration and uses first 4 channels found from the first LSL stream found from the network. If you are using "BrainAccess MINI EEG" for example which has 8 channels, you use command "--lsl-channels=8". The default number of LSL channels is 4. LSL stream listened is currently either "EEG", "Eeg" or "eeg" but this will be become configurable later.<br />
<br />
In practice you need to run software for hours to get small average positive effect to target values.<br />
<br />
If you have Interaxon Muse instead, you must configure MindMonitor app to send OSC stream to your computers IP and port 4545 and use a command (the number of channels is always 4):<br />
<br />
<code><br />
whiteice --measure --device=mindmonitor --concentration<br />
</code><br />
<br />
Also read: [[WhiteICE version information]]<br />
<br />
<strong>WhiteICE information</strong><br />
<br />
Hosaka is currently not actively developed because it does off-line optimization of larger machine learning model (you must first measure responses for hours, then compute model for hours and finally can use non-adapting model with real data). However, currently Hosaka supports using of different pictures are stimulations which feature vector will be computed for machine learning predition of responses to pictures.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=Main_Page&diff=30Main Page2026-09-03T08:07:30Z<p>Cslr: </p>
<hr />
<div><strong>Hosaka/WhiteICE information</strong><br />
<br />
Idea of the software is that you run the sound stimulation in background while using a computer. Stimulation requires many hours of measurements before it is effective. Sounds used are very simple because the learning starts from scratch (TODO: collect a database of measurements and models that can be used without a long training time + generate music instead of plain sounds).<br />
<br />
Command-line version of WhiteICE has more options. You start listening a LSL stream from EEG in a local network by using a command:<br />
<br />
<code><br />
whiteice --measure --device=lsl --concentration --lsl-channels=4<br />
</code><br />
<br />
Or for increasing intelligence (should have 8 channel EEG for good results, maximizing certain approximate small world index value):<br />
<br />
<code><br />
whiteice --measure --device=lsl --intelligence --lsl-channels=8<br />
</code><br />
<br />
The command plays sounds in the background (while using a computer), tries to increase concentration and uses first 4 channels found from the first LSL stream found from the network. If you are using "BrainAccess MINI EEG" for example which has 8 channels, you use command "--lsl-channels=8". The default number of LSL channels is 4. LSL stream listened is currently either "EEG", "Eeg" or "eeg" but this will be become configurable later.<br />
<br />
In practice you need to run software for hours to get small average positive effect to target values.<br />
<br />
If you have Interaxon Muse instead, you must configure MindMonitor app to send OSC stream to your computers IP and port 4545 and use a command (the number of channels is always 4):<br />
<br />
<code><br />
whiteice --measure --device=mindmonitor --concentration<br />
</code><br />
<br />
Also read: [[WhiteICE version information]]</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=29WhiteICE version information2026-09-03T07:24:17Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.41b (in development)</strong><br />
<br />
TODO: There is a bug in a code that causes windows GUI version of the software to crash in random locations. Lots of debugging needed.<br />
<br />
<strong>WhiteICE v1.40b (most recent)</strong><br />
<br />
Added variances of target terms to RL model data. Now results are about same when neural network's width=64 with "lsl fft concentration" target than before 3-5% effect and positive effects otherwise too. This is enough for v1.40b. TODO: but still need test with "lsl fft intelligence eeg_channels=8" target to see if the most important target increases.<br />
<br />
Increased forgetting time from one hour to three hours so more data is available for optimization.<br />
<br />
DONE: fixed "--save-video" to be faster (worse compression) and not to run out of memory for buffering video frames for encoding.<br />
<br />
<strong>WhiteICE v1.39b</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms (2-3% effect). However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it) AND test 10 sec long stimulus again to see if there is better effect (+3.1% effect vs random and about zero effect to distance).<br />
<br />
<strong>FINAL</strong>: 5000 msecs stimulation time which often gives 3% result. Note that this is worse than in 1.38b which seem to give 5% result but RL algorithms in 1.39 are more sophisticated.<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=28WhiteICE version information2026-09-03T03:00:49Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.40b (in development)</strong><br />
<br />
Added variances of target terms to RL model data. Now results are about same when neural network's width=64 with "lsl fft concentration" target than before 3-5% effect and positive effects otherwise too. This is enough for v1.40b. TODO: but still need test with "lsl fft intelligence eeg_channels=8" target to see if the most important target increases.<br />
<br />
Increased forgetting time from one hour to three hours so more data is available for optimization.<br />
<br />
DONE: fixed "--save-video" to be faster (worse compression) and not to run out of memory for buffering video frames for encoding.<br />
<br />
<strong>WhiteICE v1.39b (most recent)</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms (2-3% effect). However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it) AND test 10 sec long stimulus again to see if there is better effect (+3.1% effect vs random and about zero effect to distance).<br />
<br />
<strong>FINAL</strong>: 5000 msecs stimulation time which often gives 3% result. Note that this is worse than in 1.38b which seem to give 5% result but RL algorithms in 1.39 are more sophisticated.<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=27WhiteICE version information2026-09-02T23:05:11Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.40b (in development)</strong><br />
<br />
Added variances of target terms to RL model data. Now results are about same when neural network's width=64 with "lsl fft concentration" target than before 3-5% effect and positive effects otherwise too. This is enough for v1.40b. TODO: but still need test with "lsl fft intelligence eeg_channels=8" target to see if the most important target increases.<br />
<br />
Increased forgetting time from one hour to three hours so more data is available for optimization.<br />
<br />
TODO: really fix "--save-video" to be faster (worse compression) and not to run out of memory for buffering video frames for encoding.<br />
<br />
<strong>WhiteICE v1.39b (most recent)</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms (2-3% effect). However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it) AND test 10 sec long stimulus again to see if there is better effect (+3.1% effect vs random and about zero effect to distance).<br />
<br />
<strong>FINAL</strong>: 5000 msecs stimulation time which often gives 3% result. Note that this is worse than in 1.38b which seem to give 5% result but RL algorithms in 1.39 are more sophisticated.<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=26WhiteICE version information2026-08-07T00:46:12Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.40b (in development)</strong><br />
<br />
Added variances of target terms to RL model. This makes RL to learn actually something when it can know how accurate EEG-signal estimates of targets are. Still needs more tuning and implementation of variance terms everywhere in the code. However, thought exploration is now better, results are worse????<br />
<br />
Increased forgetting time from one hour to three hours so more data is available for optimization.<br />
<br />
TODO: really fix "--save-video" to be faster (worse optimization) and not to run out of memory for buffering video frames for encoding.<br />
<br />
<strong>WhiteICE v1.39b (most recent)</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms (2-3% effect). However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it) AND test 10 sec long stimulus again to see if there is better effect (+3.1% effect vs random and about zero effect to distance).<br />
<br />
<strong>FINAL</strong>: 5000 msecs stimulation time which often gives 3% result. Note that this is worse than in 1.38b which seem to give 5% result but RL algorithms in 1.39 are more sophisticated.<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=25WhiteICE version information2026-08-05T17:32:11Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.40b (in development)</strong><br />
<br />
Added variances of target terms to RL model. This makes RL to learn actually something when it can know how accurate EEG-signal estimates of targets are. Still needs more tuning and implementation of variance terms everywhere in the code. However, thought exploration is now better, results are worse????<br />
<br />
Increased forgetting time from one hour to three hours so more data is available for optimization.<br />
<br />
<strong>WhiteICE v1.39b (most recent)</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms (2-3% effect). However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it) AND test 10 sec long stimulus again to see if there is better effect (+3.1% effect vs random and about zero effect to distance).<br />
<br />
<strong>FINAL</strong>: 5000 msecs stimulation time which often gives 3% result. Note that this is worse than in 1.38b which seem to give 5% result but RL algorithms in 1.39 are more sophisticated.<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=24WhiteICE version information2026-08-05T13:44:02Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.40b (in development)</strong><br />
<br />
Added variances of target terms to RL model. This makes RL to learn actually something when it can know how accurate EEG-signal estimates of targets are. Still needs more tuning and implementation of variance terms everywhere in the code. However, thought exploration is now better, results are worse????<br />
<br />
Increased forgetting time from one hour to two hours so more data is available for optimization.<br />
<br />
<strong>WhiteICE v1.39b (most recent)</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms (2-3% effect). However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it) AND test 10 sec long stimulus again to see if there is better effect (+3.1% effect vs random and about zero effect to distance).<br />
<br />
<strong>FINAL</strong>: 5000 msecs stimulation time which often gives 3% result. Note that this is worse than in 1.38b which seem to give 5% result but RL algorithms in 1.39 are more sophisticated.<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=23WhiteICE version information2026-08-04T23:23:17Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.40b (in development)</strong><br />
<br />
Added variances of target terms to RL model. This makes RL to learn actually something when it can know how accurate EEG-signal estimates of targets are. Still needs more tuning and implementation of variance terms everywhere in the code...<br />
<br />
Increased forgetting time from one hour to two hours so more data is available for optimization.<br />
<br />
<strong>WhiteICE v1.39b (most recent)</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms (2-3% effect). However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it) AND test 10 sec long stimulus again to see if there is better effect (+3.1% effect vs random and about zero effect to distance).<br />
<br />
<strong>FINAL</strong>: 5000 msecs stimulation time which often gives 3% result. Note that this is worse than in 1.38b which seem to give 5% result but RL algorithms in 1.39 are more sophisticated.<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=22WhiteICE version information2026-08-04T22:19:57Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.40b (in development)</strong><br />
<br />
Added variances of target terms to RL model. This makes RL to learn actually something when it can know how accurate EEG-signal estimates of targets are. Still needs more tuning and implementation of variance terms everywhere in the code...<br />
<br />
<strong>WhiteICE v1.39b (most recent)</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms (2-3% effect). However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it) AND test 10 sec long stimulus again to see if there is better effect (+3.1% effect vs random and about zero effect to distance).<br />
<br />
<strong>FINAL</strong>: 5000 msecs stimulation time which often gives 3% result. Note that this is worse than in 1.38b which seem to give 5% result but RL algorithms in 1.39 are more sophisticated.<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=21WhiteICE version information2026-07-22T00:06:59Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.39b (most recent)</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms (2-3% effect). However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it) AND test 10 sec long stimulus again to see if there is better effect (+3.1% effect vs random and about zero effect to distance).<br />
<br />
<strong>FINAL</strong>: 5000 msecs stimulation time which often gives 3% result. Note that this is worse than in 1.38b which seem to give 5% result but RL algorithms in 1.39 are more sophisticated.<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=20WhiteICE version information2026-07-22T00:06:41Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.39b</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms (2-3% effect). However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it) AND test 10 sec long stimulus again to see if there is better effect (+3.1% effect vs random and about zero effect to distance).<br />
<br />
<strong>FINAL</strong>: 5000 msecs stimulation time which often gives 3% result. Note that this is worse than in 1.38b which seem to give 5% result but RL algorithms in 1.39 are more sophisticated.<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=19WhiteICE version information2026-07-11T23:03:59Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.39b (in development)</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms (2-3% effect). However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it) AND test 10 sec long stimulus again to see if there is better effect (+3.1% effect vs random and about zero effect to distance).<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=Main_Page&diff=18Main Page2026-07-11T08:48:23Z<p>Cslr: </p>
<hr />
<div><strong>Hosaka/WhiteICE information</strong><br />
<br />
Idea of the software is that you run the sound stimulation in background while using a computer. Stimulation requires many hours of measurements before it is effective. Sounds used are very simple because the learning starts from scratch (TODO: collect a database of measurements and models that can be used without a long training time + generate music instead of plain sounds).<br />
<br />
Command-line version of WhiteICE has more options. You start listening a LSL stream from EEG in a local network by using a command:<br />
<br />
<code><br />
whiteice --measure --device=lsl --concentration --lsl-channels=4<br />
</code><br />
<br />
The command plays sounds in the background (while using a computer), tries to increase concentration and uses first 4 channels found from the first LSL stream found from the network. If you are using "BrainAccess MINI EEG" for example which has 8 channels, you use command "--lsl-channels=8". The default number of LSL channels is 4.<br />
<br />
In practice you need to run software for hours to get small average positive effect to target values.<br />
<br />
If you have Interaxon Muse instead, you must configure MindMonitor app to send OSC stream to your computers IP and port 4545 and use a command (the number of channels is always 4):<br />
<br />
<code><br />
whiteice --measure --device=mindmonitor --concentration<br />
</code><br />
<br />
Also read: [[WhiteICE version information]]</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=Main_Page&diff=17Main Page2026-07-11T08:47:37Z<p>Cslr: </p>
<hr />
<div><strong>Hosaka/WhiteICE information</strong><br />
<br />
Idea of the software is that you run the sound stimulation in background while using a computer. Stimulation requires many hours of measurements before it is effective. Sounds used are very simple because the learning starts from scratch (TODO: collect database of measurements and models that can be used without long training + generate music instead of plain sounds).<br />
<br />
Command-line version of WhiteICE has more options. You start listening a LSL stream from EEG in a local network by using a command:<br />
<br />
<code><br />
whiteice --measure --device=lsl --concentration --lsl-channels=4<br />
</code><br />
<br />
The command plays sounds in the background (while using a computer), tries to increase concentration and uses first 4 channels found from the first LSL stream found from the network. If you are using "BrainAccess MINI EEG" for example which has 8 channels, you use command "--lsl-channels=8". The default number of LSL channels is 4.<br />
<br />
In practice you need to run software for hours to get small average positive effect to target values.<br />
<br />
If you have Interaxon Muse instead, you must configure MindMonitor app to send OSC stream to your computers IP and port 4545 and use a command (the number of channels is always 4):<br />
<br />
<code><br />
whiteice --measure --device=mindmonitor --concentration<br />
</code><br />
<br />
Also read: [[WhiteICE version information]]</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=Main_Page&diff=16Main Page2026-07-11T08:29:41Z<p>Cslr: </p>
<hr />
<div><strong>Hosaka/WhiteICE information</strong><br />
<br />
Idea of the software is that you run the sound stimulation in background while using a computer. Stimulation requires many hours of measurements before it is effective.<br />
<br />
Command-line version of WhiteICE has more options. You start listening a LSL stream from EEG in a local network by using a command:<br />
<br />
<code><br />
whiteice --measure --device=lsl --concentration --lsl-channels=4<br />
</code><br />
<br />
The command plays sounds in the background (while using a computer), tries to increase concentration and uses first 4 channels found from the first LSL stream found from the network. If you are using "BrainAccess MINI EEG" for example which has 8 channels, you use command "--lsl-channels=8". The default number of LSL channels is 4.<br />
<br />
In practice you need to run software for hours to get small average positive effect to target values.<br />
<br />
If you have Interaxon Muse instead, you must configure MindMonitor app to send OSC stream to your computers IP and port 4545 and use a command (the number of channels is always 4):<br />
<br />
<code><br />
whiteice --measure --device=mindmonitor --concentration<br />
</code><br />
<br />
Also read: [[WhiteICE version information]]</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=15WhiteICE version information2026-07-11T00:14:29Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.39b (in development)</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms (2-3% effect). However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it) AND test 10 sec long stimulus again to see if there is better effect.<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=14WhiteICE version information2026-07-10T23:46:00Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.39b (in development)</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms (2-3% effect). However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it)<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=13WhiteICE version information2026-07-10T23:15:28Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.39b (in development)</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better.<br />
<br />
DONE: Added additional measurements to RL's state vector.<br />
<br />
NOW: 5000 msecs stimulation time seems to give much better results than 1000 ms. However, forgetting time should be maybe increased from 30minutes to 60minutes? (TODO: test it)<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=12WhiteICE version information2026-07-07T18:54:31Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.39b (in development)</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau from 0.02 to 0.1 (for testing) so model can learn temporal dynamics better. <br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=11WhiteICE version information2026-07-07T18:47:30Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.39b (in development)</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
Added forgetting to reinforcemen learning, now 30 minutes older measurements are not used. TODO: increase tau to 0.1 (for testing) so model can learn temporal dynamics better. <br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=Main_Page&diff=10Main Page2026-07-06T22:08:13Z<p>Cslr: </p>
<hr />
<div><strong>Hosaka/WhiteICE information</strong><br />
<br />
Command-line version of WhiteICE has more options. You start listening a LSL stream from EEG in a local network by using a command:<br />
<br />
<code><br />
whiteice --measure --device=lsl --concentration --lsl-channels=4<br />
</code><br />
<br />
The command plays sounds in the background (while using a computer), tries to increase concentration and uses first 4 channels found from the first LSL stream found from the network. If you are using "BrainAccess MINI EEG" for example which has 8 channels, you use command "--lsl-channels=8". The default number of LSL channels is 4.<br />
<br />
In practice you need to run software for hours to get small average positive effect to target values.<br />
<br />
If you have Interaxon Muse instead, you must configure MindMonitor app to send OSC stream to your computers IP and port 4545 and use a command (the number of channels is always 4):<br />
<br />
<code><br />
whiteice --measure --device=mindmonitor --concentration<br />
</code><br />
<br />
Also read: [[WhiteICE version information]]</div>Cslrhttps://www.hosaka.fi/wiki/index.php?title=WhiteICE_version_information&diff=9WhiteICE version information2026-07-06T21:15:06Z<p>Cslr: </p>
<hr />
<div><strong>WhiteICE v1.39b (in development)</strong><br />
<br />
New version with improved RL algorithm that SHOULD give more often positive results (in many hours long stimulation). Stimulation interval is 1000ms. Effect seem to be only about 1% with this new code.<br />
<br />
Program learns very slowly, 3h 20minutes (300 models) didn’t show big improvements. Try to increase tau from 0.02 to 0.1 and see if it still increases performance a bit. => tau increase seem to give worse results???<br />
<br />
<strong>WhiteICE v1.38b</strong><br />
<br />
This version re-enabled longer stimulation interval 5000ms which seem to have positive effect to "episodes-result.txt" average reinforcement value time-series when stimulatio is run for many hours. Typical effect seem to be 5% (versus random stimulation)<br />
<br />
<strong>WhiteICE v1.37b</strong><br />
<br />
This experimental version used very short stimulation interval 500ms.</div>Cslr