Making FF Faster in ADL Domains

The fast-forward planning system (FF), which obtains heuristics via a relaxed planning graph to guide the enforced hill-climbing search strategy, has shown excellent performance in most STRIPS domains. When it comes to ADL domains, FF handles actions with conditional effects in a way similar to factored expansion. The result is that enforced hill-climbing guided by the relaxed Graphplan always fails in some ADL domains. We have discovered that the reason behind this issue is the relaxed Graphplan's inability to handle relationships between actions' components. We propose a novel approach called delayed partly reasoning on a naive conditional-effects planning graph (DP-CEPG). We do not ignore action's delete effects and consider restricted induced component mutual exclusions between factored expanded actions. Preliminary results show that enforced hill-climbing while guided by DP-CEPG gains obvious improvements in most ADL problems in terms of both solution length and runtime.