Research And Application On Relational Reinforcement Learning

Reinforcement Learning has obtained significant progress during the past several-decade development and become one of the most active areas in Machine Learning and Artificial Intelligence. In practice, because of the excessive scales of States Space and the restrictions in hardware condition, the efficiency of algorithms are not well enough, most of which adopt computing of attribute-value and cannot reflect relations among objects. Along with the progress of logic programming, the relation could be described by variable, which made learning tasks abstracted from complicated computation. Relational Reinforcement Learning affords a new method for dealing with big States Space of Reinforcement Learning by combining logic programming and Reinforcement Learning.This paper focus on two works:1. Propose the improved Relational Reinforcement Learning algorithm through the analysis of all kinds of existing algorithms and operating mechanisms. The improved algorithm adopts incremental Logical Decision Trees, deals with every sample and decreases calculation because there are too many repeats of calculations and iterations, and date copies in original algorithms. In order to recuperate the lost of information of leaf in incremental update, the algorithm assigns each of predicator a priority. Select candidate tests in leaf splitting according to priority level in order to increase the convergence rate of algorithm. Contrasts the original algorithm by experiment, the improved algorithm has a clear promotion in efficiency.2. Outline the control algorithm of the existing intelligent cars and establish an Autonomous Driving System base on Relational Reinforcement Learning. The Autonomous Driving System consists of three parts: state analysis model, policy learning model and knowledge base. These models make it easy for vehicles with different characteristics by setting different background knowledge. It takes full advantage of ability of Relational Reinforcement Learning, and improves the adaptability of system. Through simulating different environments, we test the performance of new Autonomous Driving System, and obtain better results.