The More, the Merrier: Combining Heuristic Estimators for Satisficing Planning

Abstract

We empirically examine several ways of exploiting the information of multiple heuristics in a satisficing best-first search algorithm, comparing their performance in terms of coverage, plan quality, speed, and search guidance. Our results indicate that using multiple heuristics for satisficing search is indeed useful. Among the combination methods we consider, the best results are obtained by the alternation method of the "Fast Diagonally Downward" planner.

Learning Heuristic Selection with Dynamic Algorithm Configuration

“…It has been shown, however, that the combination of several heuristic values into one, e.g. by taking the maximum or a (weighted) sum, does not lead to informative heuristic estimates (Röger and Helmert 2010). This can be explained by the fact that if one or more heuristics provide very inaccurate values, the whole expansion process is affected.…”

“…Although there is a large number of heuristics, it is known that the performance, i.e., the informativeness, of a heuristic varies from instance to instance (Wolpert and Macready 1995). While in optimal planning it is easy to combine multiple admissible heuristic estimates using the maximum, in satisficing planning the estimates of inadmissible heuristics are difficult to combine in general (Röger and Helmert 2010). The reason for this is that highly inaccurate and uninformative estimates of a heuristic can have a negative effect on the entire search process when aggregating all estimates.…”

“…Therefore, an important task in satisficing planning is to utilize multiple heuristics within one heuristic search. Röger and Helmert (2010) showed the promise of searching with multiple heuristics, maintaining a set of heuristics, each associated with a separate open list to allow switching between such heuristics. This bypasses the problem of aggregating different heuristic estimates, while the proposed alternating procedure uses each heuristic to the same extent.…”

“…Recent works have investigated switching between different search strategies depending on the internal search dynamics of a planner (Gomoluch, Alrajeh, and Russo 2019;Gomoluch et al 2020). One approach that shares the search progress is to maintain multiple heuristics as separate open lists (Röger and Helmert 2010). Furthermore, it has been shown that boosting, i.e., giving preference to heuristics that have recently made progress, can improve search performance (Richter and Helmert 2009).…”

“…In this work, we introduce and define dynamic algorithm configuration (Biedenkapp et al 2020) for planning, by learning a policy that dynamically selects a heuristic within a search with multiple open lists (Röger and Helmert 2010) based on the current search dynamics. We prove that a dynamic adjustment of heuristic selection during the search can exponentially improve the search performance of a heuristic search compared to a static heuristic selection or a non-adaptive policy like alternating.…”