"Search, Opponent Modelling, Cooperation, and State Inference in Complex Imperfect Information Domains."

点击翻译按钮获取中文摘要

Artificial intelligence (AI) research applied to games has a long tradition that reaches back at least 75 years with Alan Turing's work on computer chess. The advantage of studying AI algorithms in this area is that games are precisely defined, relatively small when compared to real-world decision domains, and yet sufficiently complex to pose tough research problems whose solutions can help us create machines of human-level intelligence. AI research has had its successes in games like chess, backgammon, and checkers - where machines now play on par with or better than the best human players. However, in more complex domains machines are still trailing behind human experts. The main differences to the games in which AI research has been very successful are that game state information is hidden from players or the number of move choices is very large. Both properties render complete enumeration, which is a cornerstone of many high-performance AI systems, less effective. The objective of the research proposed here is to create systems that will reach or surpass the performance of human experts in real-time decision domains that feature uncertainty, imperfect information, or complex state and action spaces. The benchmark applications we will be working on are trick-based card games and real-time strategy video games. Improving the state of the art requires us to develop new algorithms that can model opponents, infer hidden game states, cooperate with partners, and look-ahead in abstracted search spaces. In the long term, the results of this project will increase our understanding of fundamental AI problems that need to be solved in the process of creating human-level AI systems. In the short term, we anticipate the computer games industry benefiting from our research, because it is in need of credible computer controlled agents in the domains we study.