An Event-based Reinforcement Learning Method For Navigation Control Of Mobile Robots

Event-based optimization and learning methods are a kind of effective approach to speed up the learning process by using the special feature of the problem to be solved. In this paper, a novel event-based probabilistic Q-learning method is presented. The navigation control of mobile robots in a complex environment is selected in order to verify the validity of this method. First, the basic idea of event-based reinforcement learning (ERL) is introduced and a probabilistic Q-learning (PQL) method is extended to its event-based version as an example. Then the proposed event-based probabilistic Q-learning (EPQL) algorithm is applied to a mobile-robot navigation problem that uses a wireless sensor network system for observing the events, including Dead reckoning and RFID methods. A fuzzy controller is used to improve the probability of event trigger. On the basis of the simulation, physical experiment is designed to further illustrate the feasibility of the event-based methodThe results show that the EPQL algorithm is effective for navigation problems in large complex environments and the ERL approach can improve the learning performance by using events that characterize the structure of the problem.