Adding Exploration to Greedy Best-First Search

While greedy best-first search (GBFS) is a popular algorithm for solving automated planning tasks, it can exhibit poor performance if the heuristic in use mistakenly identifies a region of the search space as promising. In such cases, the way the algorithm greedily trusts the heuristic can cause the algorithm to get stuck performing a lot of unnecessary work. In this report, we consider simple techniques that help GBFS to avoid overly trusting the heuristic. The first technique, heuristic perturbation, will be shown to lead to large increases in coverage in some domains, and large decreases in others. Over all problems tested, this technique does increase the average coverage by up to 9.5% over standard GBFS when it is parameterized effectively. The second technique, -greedy node selection, will be shown to lead to smaller improvements than heuristic perturbation in many domains, though it does so without hurting the algorithm’s performance in any other domains. Over all tested problems, this technique will be shown to increase coverage even when used with a wide range of parameter values, with the best setting leading to a 11.0% increase in coverage when compared to standard GBFS. We will also show that these techniques can be paired together effectively in an algorithm portfolio due to the complementary way they introduce exploration into the search, with our best portfolio having an expected coverage that is 22.5% higher than standard GBFS.