Structural Bias in Heuristic Search (Student Abstract)

Abstract

In this line of work, we consider the possibility that some fast heuristic search methods introduce structural bias, which can cause problems similar to sampling-bias for downstream statistical learning methods. We seek to understand the source of this kind of bias and to develop efficient alternatives. Here we present some preliminary results in developing a variation of canonical A* that can overcome the structural bias introduced by first-in-first-out duplicate detection, which we observed under the condition of variable heuristic error. These results inspire a model of greedy-best-first-search for this problem in the satisficing setting. We hope to apply our approach in a novel planning application--activity selection for agent-based modeling for epidemiology--where planning technology should avoid introducing structural bias if possible.