Heuristic Policy Analysis and Efficiency Assessment in Constraint Satisfaction Search

This paper argues that characterizing heuristic performance in terms of adherence to optimal policies can elucidate many differences in search effort associated with different variable ordering heuristics. This framework involves two policies that search must adhere to to be successful: the promise policy is in force when search is on a solution path; the fail-first policy holds when search is in an insoluble subtree. After discussing how adherence to these policies can be measured, the paper shows that many complex patterns of performance can be elucidated by measuring the adherence to each policy. For example, some strategies designed to maximize adherence to one policy are shown to be either unsuccessful in this regard or to affect adherence to the other policy adversely, thus impairing search. In contrast, strategies used by superior heuristics often balance features such as branching and local connectivity so as to enhance adherence to both policies simultaneously. Differences related to problem structure can also be elucidated