Sarkozy, Gabor N
Selkow, Stanley M.
We have developed the Gomba Testing Framework, a new platform for the comparative evaluation of search algorithms in large adversarial game trees. Gomba is simple, fast, extensible, objective, and can scale to larger trees than previous frameworks have been able to test against. We have implemented and tested a variety of Monte-Carlo based search algorithms using the framework and have analyzed their performance in relation to known metrics in computer Go. Finally, we have taken several solutions to the infinitely-many-armed bandit problem and adapted them to tree search. We have tested these variants in both Gomba and computer Go, and have shown that they can be effective in both cases.
Worcester Polytechnic Institute
Major Qualifying Project
All authors have granted to WPI a nonexclusive royalty-free license to distribute copies of the work, subject to other agreements. Copyright is held by the author or authors, with all rights reserved, unless otherwise noted.