tag:blogger.com,1999:blog-4837627029742618106.post970061351560619912..comments2016-12-23T17:17:02.688-08:00Comments on Computational Backgammon: Player 3: training with supervised learningMark Higginshttp://www.blogger.com/profile/16878070026852488955noreply@blogger.comBlogger2125tag:blogger.com,1999:blog-4837627029742618106.post-14195268283647651402012-02-15T22:26:42.518-08:002012-02-15T22:26:42.518-08:00One interesting point: Joseph noted that when he r...One interesting point: Joseph noted that when he ran supervised learning he was using huge alphas, like 20, and minimum alphas around 0.5 or 1. That's not what I'm seeing: with big alphas it doesn't converge. I started with alpha=1 and went down from there, without cycling back up to large alphas if performance doesn't improve (like Joseph did).Mark Higginshttps://www.blogger.com/profile/16878070026852488955noreply@blogger.comtag:blogger.com,1999:blog-4837627029742618106.post-18623441633709917822012-02-14T13:46:20.583-08:002012-02-14T13:46:20.583-08:00This goes forward!
It is Ian's and mine exper...This goes forward!<br /><br />It is Ian's and mine experience that the best nets are those that are started as TD and then the training is extended with supervised training.<br /><br />"Epoch" is not only Josephs term. It's the common term in the ML literature.<br /><br />-ØysteinØystein Schønning-Johansenhttps://www.blogger.com/profile/13432607600680027833noreply@blogger.com