Adaptive Tree Search

Advisor: Stuart M. Shieber Adaptive Tree Search Wheeler Ruml Combinatorial optimization and constraint satisfaction problems are ubiquitous in computer science, arising in areas as diverse as resource allocation, automated design, planning, and logical inference. Finding optimal solutions to such problems often entails searching an intractably large tree of possibilities. Problems beyond the reach of exhaustive enumeration require methods that exploit problem-specific heuristic knowledge to find the best solution possible within the time available. Previous algorithms typically follow a fixed search order, making inefficient use of heuristic information. For a new problem, it is not always clear which predetermined search order will be most appropriate. I propose an adaptive approach in which the search order is adjusted according to heuristic information that becomes available only during the search. By adapting to the current problem, this approach eliminates the need for pilot experiments and enables the use of good search orders that would be too complicated to program explicitly. To demonstrate the feasibility of the approach, I first present a simple but incomplete technique, adaptive probing. Empirical results demonstrate that the method can effectively adapt on-line, surpassing existing methods on several synthetic benchmarks. I then introduce a general framework for complete adaptive tree search, best-leaf-first search, and show how previous work can be viewed as special cases of this technique. Incorporating different sources of information into the framework leads to different search algorithms. Five dif-