For a bit of a secondary project I've been doing the past few weeks I made a self-learning Tetris program. To my own surprise this worked extremely well. It 'grows' an expert player in just a few thousand generations (a few days real time) starting from zero knowledge . And it doesn't have any 'a priori' knowledge about the game, so it adjusts when the (scoring) rules or playing conditions change. It works so well, I was thinking this could very well apply to computer-Go somehow.
Tetris is not Go of course. And also it's a single-player game, where Go is a two-player game. So some adjustments about determining the survivors and procreation may have to be made. It's a bit early for me to divulge more details at this point. Also, this is research which my current employer may not want to become public in great detail just yet. But based on my findings so far I can only encourage anyone who thinks about making a self-learning Go program to go ahead and try. Mark On Mon, Apr 26, 2010 at 11:53 AM, Hendrik Baier <[email protected]> wrote: > Hello list, > > I am a Master's student currently working on his thesis about certain > aspects of Monte-Carlo Go. I would like to pose a question concerning the > literature - I hope some of you can help me out! > > My problem is that I can't find many papers about learning of MC playout > policies, in particular patterns. A lot of programs seem to be using Mogo's > 3x3 patterns, which have been handcoded, or some variation thereof. A lot of > people have tried some form of pattern learning, but mostly to directly > predict expert moves it seems, not explicitly optimizing the patterns for > their function in an MC playout policy. Actually, I am only aware of > "Computing Elo Ratings of Move Patterns in the Game of Go", where patterns > have been learned from pro moves, but then also successfully used in an MC > playout policy; and "Monte Carlo Simulation Balancing". > > Considering the huge impact local patterns have had on the success of MC > programs, I would have expected more attention towards automatically > learning and weighting them specifically for MC playouts. There is no reason > why patterns which are good for predicting experts should also be good for > guaranteeing diverse, balanced playout distributions. Have I missed > something? > > Or how did your program come to its patterns? I'd be interested. Did you > maybe even try learning something else than patterns for your playout > policy? > > cheers, > Hendrik > _______________________________________________ > Computer-go mailing list > [email protected] > http://dvandva.org/cgi-bin/mailman/listinfo/computer-go > _______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
