Automatic Learning of Search Control in Heuristic Search

Search plays a central role for solving problems in many fields of computer science and, in particular, artificial intelligence (AI). Recent successes in AI disciplines like planning and scheduling, game playing, and constraint programming are in big part due to the development of effective search techniques for exploring huge search spaces. Often there is much work involved in constructing informed heuristics for guiding the search, and subsequently there is interest in methods that can automatically learn search guidance. Such methods become increasingly efficient at solving tasks over time as more and more problem instances are encountered and solved, that is, they are capable of "learning from experience".

The main objective of this research project is to develop new domain-independent adaptive search techniques. The research focus will primarily be along two directions: to further enhance ongoing research into adversary search techniques for learning search control and, secondly, to generalize these methods such that they can also be applied to non-adversary domains such as single-agent search and automated planning. The new techniques will be integrated into existing heuristic-search solvers and planners for measuring their effectiveness.