Event-based Reinforcement Learning And Its Implementation In Indoor Navigation

With the development of wide application of intelligent control algorithm in daily life and production,Reinforcement Learning is confronted with the Curse of Dimensionality and ineffectiveness to Non-MDP more than ever.This paper introduces the event-based optimization to state space partition in RL so as to speed up the convergence and get good performance in Path Planning of the mobile robot.First,the event-based optimization theory is introduced to RL,extracting structural information of the environment,aggregating the states with similar properties and controlling the learning process by triggering the events.Decompose the global task to make each sub-task solved in their rather small space and reuse these sub-strategies to speed up the convergence.A simulation is designed with variant-scale grid maps to prove the efficiency.In the mobile robot experiment,the robot SLAMs while labeling the event landmark by visual input.Then the map is transformed to the grid and the problem is modeled by Event-based RL.Finally,the mobile robot succeeds to navigate by the path learned from Event-based Q-learning algorithm.