I'd like to introduce a paper presented a few days ago in CIG2015, Taiwan.
The main idea is to enlarge the KL-divergence between optimal move and others by biasing the threshold of win/loss in MCTS algorithms. Not exactly the same as dynamic komi but similar and I believe it has some (strong) relation. "Enhancements in Monte Carlo Tree Search Algorithms for Biased Game Trees" by Takahisa Imagawa and Tomoyuki Kaneko. http://game.c.u-tokyo.ac.jp/ja/wp-content/uploads/2015/08/ieeecig2015.pdf Hideki Goncalo Mendes Ferreira: <[email protected]>: >I've been wrapping my head about dynamic komi adjustments for MCTS, >namely on the thesis by the Pachi creator, Petr Baudic. > >On value-based situational compensation the author uses the average on >win rates from the previous simulations to decide whether or not to >change the komi. But I don't see how this criteria makes sense, if we're >interested in finding the best play, shouldn't we be trying to have good >sensibility around the best plays? Trying to average the game only >worsens the ability of the search to differentiate the best contenders. > >Am I seeing this wrong? Has this been addressed before? What do other >engines do? > >- Gonçalo F. >_______________________________________________ >Computer-go mailing list >[email protected] >http://computer-go.org/mailman/listinfo/computer-go -- Hideki Kato <mailto:[email protected]> _______________________________________________ Computer-go mailing list [email protected] http://computer-go.org/mailman/listinfo/computer-go
