Learning Generalised Policies for Numeric Planning.

Abstract

We extend Action Schema Networks (ASNets) to learn gen-eralised policies for numeric planning, which features quan-titative numeric state variables, preconditions and effects. Wepropose a neural network architecture that can reason aboutthe numeric variables both directly and in context of othervariables. We also develop a dynamic exploration algorithmfor more efficient training, by better balancing the explo-ration versus learning tradeoff to account for the greater com-putational demand of numeric teacher planners. Experimen-tally, we find that the learned generalised policies are capableof outperforming traditional numeric planners on some do-mains, and the dynamic exploration algorithm to be on aver-age much faster at learning effective generalised policies thanthe original ASNets training algorithm