RI: Small: A Human-Level, Real-Time, Integrated Agent

该项目正在开发和集成人工智能的统计和符号方法在代理体系结构中，并在竞争领域评估代理，特别是实时战略游戏星际争霸。实时战略(RTS)游戏提供了几个有趣的研究挑战，包括实时决策、巨大的状态空间和不完全信息。星际争霸是一款流行的商业RTS游戏，有几个专业的游戏联盟，因此是评估AI特工表现的理想选择。专业的星际争霸玩家在多个抽象层次上对战略决策进行推理和反应，有时每分钟执行300多个游戏动作，因此开发竞争级别的星际争霸代理是一个非同寻常的挑战。更具体地说，该项目使用新颖的监督和非监督学习算法来自动从专业游戏轨迹集合中学习领域知识；该代理是在反应式规划体系结构ABL(A Behavior Language)中实现的。ABL反应式计划器为在实时执行环境中集成多个不同类型的推理机提供了粘合剂。这项工作有望为理解复杂的实时领域的决策过程做出重大贡献。这一理解将有助于开发可在真实环境中部署的健壮、智能的系统。这项工作将激励人工智能研究人员构建集成的代理体系结构。作为一款拥有非常高水平专业玩法的知名游戏，对星际争霸AI的研究有可能吸引人们对AI研究的极大关注。由我们实验室主办的星际争霸大赛吸引了学术界内外以及高中、本科生和研究生的极大兴趣。因此，这项工作有可能提高公众对人类水平人工智能研究的认识，并将鼓励高中生从事计算机科学和游戏设计方面的职业。

This project is developing and integrating statistical and symbolic methods of Artificial Intelligence in an agent architecture and evaluating the agent in a competitive domain, notably the real-time strategy game StarCraft. Real-time strategy (RTS) games provide several interesting research challenges including real-time decision making, enormous state spaces and imperfect information. StarCraft is a popular commercial RTS game that has several professional gaming leagues, and therefore ideal for evaluating the performance of AI agents. Professional StarCraft players reason about and react to strategic decisions at multiple levels of abstraction, sometimes executing over 300 game actions per minute, so developing competition-level StarCraft agents presents extraordinary challenges. More specifically, the project is using novel supervised and unsupervised learning algorithms to automatically learn domain knowledge from collections of professional gameplay traces; the agent is being implemented within the reactive planning architecture ABL (A Behavior Language). The ABL reactive planner provides the glue for integrating multiple, heterogeneous reasoners within a real-time execution environment. This work is expected to make significant contributions to the understanding of decision making processes in a complex, real-time domain. This understanding will contribute to the development of robust, intelligent systems that can be deployed within real-world environments. This work will motivate AI researchers to build integrated agent architectures. As a well-known game with very high-level professional play, research in StarCraft AI has the potential to attract significant attention to AI research. The StarCraft competition being hosted by our lab has attracted significant interest both within and outside academia, and at the high-school, undergraduate and graduate level. Thus, this work has the potential to raise general public awareness in research in human-level AI, and will encourage high-school students to pursue careers in computer science and game design.