Humanization of computational learning in strategy games

[摘要] I review and describe 4 popular techniques that computers use to play strategy games: minimax, alpha-beta pruning, Monte Carlo tree search, and neural networks. I then explain why I do not believe that people use any of these techniques to play strategy games. I support this claim by creating a new strategy game, which I call Tarble, that people are able to play at a far higher level than any of the algorithms that I have described. I study how humans with various strategy game backgrounds think about and play Tarble. I then implement 3 players that each emulate how a different level of human players think about and play Tarble.

[发布机构] Massachusetts Institute of Technology