Planning Bias: Planning as a Source of Sampling Bias
Many data-driven planning methods are trained on data generated by planners. It is well known that many statistical learning methods are sensitive to sampling bias, and yet there has been little or no attention to planning as a sampling method and its role in introducing sampling bias into planner-generated training data. Recently, it has been demonstrated that A**,* in the presence of problems with variable heuristic error, prefers some solutions over other equally cost-optimal solutions. But, as we discuss in this paper, mitigation may not be as simple as resolving arbitrary tie-breaking by sampling from ties uniformly at random. In this paper, we formalize an intuition of planning bias. We focus on problems which output a single solution. Diverse planning only complicates the problem by generalizing it to bias in the set of sets; we show how it is subject to bias in the single solution. We make some useful observations about deterministic algorithms in contrast to non-deterministic algorithms. We explain how information entropy may be a good way to measure planning bias, and discuss some issues in evaluating practical approaches to measurement. We address the intuition that uniform random tiebreaking should mitigate bias; and sketch a novel approach to constructing an appropriate random distribution for duplicate detection during forward search for unbiased A*. Finally, we suggest directions for future work.