Engineering topics

Christian Muise

Publications and source records attributed to Christian Muise.

3 records

NASA NTRSDate not supplied

Avoiding Selection Bias in Generating Examples of Plans in the Presence of Heuristic Error

It is generally understood that heuristic error hurts the performance of search algorithms, measured in terms of search effort. Hence there is an interest in understanding how to reduce heuristic error. One way to do this is to learn a heuristic from a set of examples of plans generated offline, e.g. bootstrapping methods. In this paper, we consider how some methods for generating examples of plans may skew the training set in the presence of heuristic errors. Initial theoretical results show that duplicate detection is one source of selection bias in the canonical A* algorithm. We introduce a duplicate selection scheme for A* that avoids selection bias in generating cost-optimal examples, without compromising memory efficiency, and develop ideas in the satisficing setting. We evaluate our approach on n x m grids with multiple cost-optimal solutions and synthetic heuristic error. Finally, we attempt to extend these ideas to the problem of generating extreme examples of plans.

Alison S Paredes↗

NASA NTRSDate not supplied

Bias in Planning Algorithms

Does bias exist in planning algorithms? If so, how does bias manifest, and how important is this bias? Answering this question requires a formal, mathematical definition of bias. We formally define bias as the distance between the probability distributions of solutions returned by various algorithms, and the uniform distribution over solutions. We show in this paper that deterministic algorithms are inherently biased, as they don’t return all solutions, and that this property holds even when algorithms return a set of plans instead of just one plan. Exceptions are problem instances or problem classes for which only a single solution exists. We then discuss changing the definition of bias to compare the probability distributions of properties of sets of plans instead of individual plans. We show the property bias is smaller than the bias of actual plans. Finally, we show that entropy is a proxy for the more complex and more expensive distance measurement between pairs of probability distributions. We then describe a roadmap for future investigations of bias in planning.

Planning Scheduling Algorithms↗

NASA NTRSDate not supplied

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.

Planning Scheduling Algorithms↗