Learning/Tuning in SlowChess Blitz Classic

[jonkr]

I wrote a fairly detailed description of my experiences working with automatic tuning/learning from 1.9, 2.0, 2.1, 2.2 at the webpage below :

Learning & Automatic Tuning for Slow Chess Blitz Classic

It isn't meant to be present anything new, but I thought it would be of some interest given automatic learning is the best thing I've found for increasing playing strength fairly quickly. Using tuning directly or as a guide I was able to increase my program by an estimated 200 self-play elo from just evaluation. (The next best thing I've found is actually trying to run significant test matches on changes, but I don't have compute resources to do that very well.)

Despite unexpectedly good results, I do also wonder what the secrets to making the neural nets so amazingly strong are. If it's because choosing eval features by hand is too much of a bottleneck, my lack of computing resources for training/tuning, not being able to make use of the GPU, weakness in my tuning technique, weakness in my search, etc. Probably a combination of all the above.

Also at the bottom of the page are the current named evaluation table values for 2.2

[jswaff]

Great write up, thanks for posting it. It's lead to an evening of enjoyable reading.

[Dann Corbit]

You should change the name of SlowChess to DragonTamer:
http://www.talkchess.com/forum3/viewtop ... 70#p848835