UCT for Games and Beyond

点击翻译按钮获取中文摘要

Artificial Intelligence (AI) research and the development of the multi-billion dollar video games industry have gone hand in hand for many years. Video games are by far the most prevalent way that the public encounter AI techniques on a day to day basis, and the desire for better video games has driven AI research in areas such as move/path planning, decision making, non-player character (NPC) behaviour and the automated generation of game content. A recent development of Monte Carlo methods called the Upper Confidence Bounds for Trees (UCT) method promises to have a profound impact on AI for games. Applications of UCT are not limited to games and have potential benefits for almost any domain where simulation and statistical modelling can be used to forecast outcomes, such as planning, decision support, economic modelling, behavioural analysis, and so on.Since it appeared in 2006/7, UCT has revolutionised the demanding problem of move planning for computer Go to produce artificial players able to beat professional players for the first time this year, a feat previously thought infeasible. UCT has also been successfully applied to the less specialised domain of General Game Playing (GGP) to produce the 2008 and 2009 world champion GGP programs. This success in Go, where substantial problem-specific knowledge is used, and in GGP, where it is impossible to use problem-specific knowledge, points to the tantalising possibility of the broad use of UCT between these two extremes. Game AI researchers are now starting to take such a great interest in UCT that we are seeing the birth of a new research field of Monte Carlo Tree Search (MCTS). However, there has been to date no unified effort to fully understand and exploit the UCT algorithm and related MCTS methods, a state of affairs that we plan to redress.The proposed research will develop and evaluate novel extensions of the UCT method to increase its applicability to a broad range of game-related domains including: its use for move planning and decision making in infinite, continuous real-time environments; its application to situations involving uncertainty and incomplete information; and its application to multi-objective and ensemble planning approaches. We will also investigate its use for more general game-related problems including the detection and optimisation or correction of suboptimal game designs and game content, and the automated generation of new high quality games and game content. Further, we will demonstrate how the techniques we develop can be applied to broader non-game domains by demonstrating their application to robotic control and automated music generation, in particular the creatively challenging task of jazz improvisation.The potential impact of UCT and MCTS cannot be overstated. Landmark events that have driven AI research include the introduction of tree search methods which have been the backstay of AI decision making since the inception of this field in the 1950s, and the formalisation of Monte Carlo methods in the 1970s for simulation-based decision making in a broader range of more general and less well-defined problems. UCT/MCTS promises to be the next major breakthrough in AI methods that combines the power of tree search with the generality of simulation-based search.