Learning Search Control Strategy

目前最好的双人完全信息博弈程序只玩一个游戏；他们穷追不舍，却不学习。该项目的目标是进一步证明弱理论作为搜索范式的可行性，并促进构建能够学习整个任务类别的计算机，而不是需要个性化指导。这项研究开发了一个程序HOYLE，它可以正确地玩任何这样的游戏，并且通过经验，可以学会玩得非常好。HOYLE并没有进行广泛的搜索，而是在弱域理论的指导下为每款新游戏学习搜索控制策略：关于游戏玩法的程序知识、关于特定游戏的陈述性知识、关于战略元素的语言和框架，以及一组被称为顾问的狭窄但专业的观点的组合。当它玩一个新游戏时，HOYLE会根据自己的游戏经验，选择性地构建、组织和重新制定每个游戏的知识库。HOYLE采用了一种新颖的架构，采用多种控制策略进行移动选择，将这些知识库与它的顾问相结合，为游戏创造出连贯、深刻、稳步改进的策略。这种多面学习的原型已被证明在一个不同但有限的领域非常有效。HOYLE在其设计和实施中都涉及理论形成、专家合作、冲突解决、实验设计和操作方面的重要问题。

The best current programs for two-person perfect information games play only a single game; they search exhaustively and they do not learn. The goals of this project are to demonstrate further the viability of a weak theory as a search paradigm and to facilitate the construction of computers that learn entire categories of tasks, rather than requiring individualized instruction. This research develops a program, HOYLE, that can play any such game correctly and, with experience, can learn to play it extremely well. Instead of extensive search, HOYLE learns a search control strategy for each new game under the guidance of its weak domain theory: a combination of procedural knowledge about game playing, declarative knowledge about specific games, a language and framework for strategic elements, and a set of narrow but expert perspectives called Advisors. As it plays a new game, HOYLE selectively constructs, organizes, and reformulates a knowledge base for each game from its playing experience. Under a novel architecture that employs a variety of control strategies for move selection, HOYLE combines that knowledge base with its Advisors to produce a coherentr, incisive, steadily improving strategy for the game. A prototype of this multifaceted learning has proved itself remarkably effective in a varied but limited domain. HOYLE addresses, both in its design and in its implementation, important questions in theory formation, collaboratin of experts, conflict resolution, experimental design, and operationalization.