Rule Generation and Generalization by Inductive Decision Tree and Reinforcement Learning

Abstract

In this paper, we attempt to construct a planning mechanism composed of distributed agents for autonomous vehicle navigation in an unknown workspace. Each agent decides a direction that the vehicle should move without any communication to the other agents, by only observing the workspace and the other agents. The agent includes the reinforcement learning mechanism for generating rules for the navigation. However, the rules depend on a state observation method. For generalizing the rules, an inductive decision tree is introduced to the agent. In a new workspace, the agent plans a path efficiently by learning a specific rule to the new workspace, and using the generalized rules. Some computational simulations have been carried out for verifying the proposed agent.