資訊新知 Playing Games with Computational Intelligence 許舜欽 2011/2/23
資訊領域 資訊科學 資訊工程 資訊管理 資訊傳播 …….
資訊工程 理論 硬體 軟體---程式設計 系統軟體 應用軟體 離散數學、計算理論… 邏輯電路設計、計算機結構… 作業系統、資料庫系統、編譯程式… 應用軟體 多媒體應用、網際網路應用、人工智慧應用…
Artificial Intelligence Automatic Theorem Proving Heuristic Search---Computer Game Playing Machine Learning Computer Vision Natural Language Processing Robots …….
Computer Game Playing Offering a diverse range of engaging problems and applications For the first few decades Beating expert human players at some of the most challenging board games Over the last decade Investigating the application of AI and CI to video games
Artificial Intelligence vs. Computational Intelligence Artificial Intelligence(AI) Deals with the development of machine intelligence by any means Computational Intelligence(CI) Deals with algorithms and architectures that enables intelligent behavior to emerge via statistical processes
Game Tree Search Conventional techniques Mini-max search with alpha-beta pruning Two features A good evaluation function A low or modest branching factor Lead only to modest levels of play and offered no threat to expert human player for Computer GO
Monte Carlo algorithms Rely on random sampling and simulated annealing Playing random moves until the end of the game. The win/lose statistics are then used to estimate the value of that position
Monte Carlo Tree Search Selectively building up a tree of explored positions Use the Upper Confidence Bounds for Trees method for the selection policy Have made truly astonishing progress in the world of Computer GO More CPU leads to more simulated play which leads to higher quality actual play
General Game Playing A way to make games a true challenge for machine learning Operate in two phases First—the game rules are given to each player Second– play commences and continue until the end of the game Use a logic based game description language Not appropriate for video games
Video Games As an application of computational intelligence As a test-bed for computational intelligence Hand-programmed with a relatively small number of parameters adapted using evolutionary algorithms Still leaves much room for improvement
References Playing Games with Computation Intelligence Monte Carlo GO Simon M. Lucas Monte Carlo GO B. Brugmann Bandit based Monte Carlo planning L. Krocsis and C. Szepesvari General game playing: Overview of the AAAI competition M.R. Genesereth, N. Love, and B. Pell