원문정보
초록
영어
Unlike other games such as chess, draughts and backgammon, computers are currently quite weak at the game of go ( baduk). Brute force is difficult due to the higher branching factor and game length. Human made algorithms become very complex before even approaching human strength on a subproblem of the game. One possible approach to this challenging problem is to use machine learning to let the program learn and improve without increased human effort. Machine learning has been successful in other games (e.g. draughts, backgammon). In this paper we give an overview of existing techniques. We discuss different aspects of learning, and propose some directions of research. In particular we believe that a first order representation language combined with a multistrategy learning system can achieve much more than what currently exists.
목차
2. Computer game playing
2.1 Min-max search and Alpha-beta pruning
2.2 Endgame theory
3. What can be learned?
3.1 Global approaches
3.2 Learning in search
3.3 Learning in the endgame
3.4 Learning in the opening
3.5 Higher level concepts
4. The representation language
5. Experiments
5.1 Learning a heuristic for Tsume-Go problems
5.2 Learning an opening book with temperatures
6. Conclusions and further work
Acknowledements
References