Reducing the Planning Horizon Through Reinforcement Learning

Planning is a computationally expensive process, which can limit the reactivity of autonomous agents. Planning problems are usually solved in isolation, independently of similar, previously solved problems. The depth of search that a planner requires to find a solution, known as the planning horizon, is a critical factor when integrating planners into reactive agents. We consider the case of an agent repeatedly carrying out a task from different initial states. We propose a combination of classical planning and model-free reinforcement learning to reduce the planning horizon over time. Control is smoothly transferred from the planner to the model-free policy as the agent compiles the planner's policy into a value function. Local exploration of the model-free policy allows the agent to adapt to the environment and eventually overcome model inaccuracies. We evaluate the efficacy of our framework on symbolic PDDL domains and a stochastic grid world environment and show that we are able to significantly reduce the planning horizon while improving upon model inaccuracies.