Incremental Learning of Action Models for Planning

Anonymous

Keywords: learning action models, reinforcement learning, incremental learning

TL;DR: Introduce an approach to allow agents to learn PPDDL action models incrementally over multiple planning problems under the framework of reinforcement learning.

Abstract: The soundness and optimality of a plan depends on the correctness of the domain model. In real-world applications, specifying complete domain models is difficult as the interactions between the agent and its environment can be quite complex. We propose a framework to learn a PPDDL representation of the model incrementally over multiple planning problems using only experiences from the current planning problem, which suits non-stationary environments. We introduce the novel concept of reliability as an intrinsic motivation for reinforcement learning, and as a means of learning from failure to prevent repeated instances of similar failures. Our motivation is to improve both learning efficiency and goal-directedness. We evaluate our work with experimental results for three planning domains.