Research On Behavior Learning For Intelligent Robot Based On Evolutionary Algorithms

The real environment in which robot works is often dynamic, unknowable and unexpected. Traditional sensor-planning-action method requires that people can forecast all possible situations. Face this complex environment, it is very difficult to design robot's behaviors by researches. So it's necessary to decrease people's supervision and evolve robot's adaptive behavior by a self-organization process. Behavior learning is one of pivotal technologies of intelligent robotics.Methods for behavior learning include behavior-based robotics, reinforcement learning and evolutionary algorithms etc. This paper is mainly research robot's behavior learning by means of evolutionary algorithms. Begin with the simplest and most representative behavior—obstacle-avoidance behavior, we evolve obstacle-avoidance behavior by genetic algorithm and genetic programming separately. The following aspects are investigated and discussed.Firstly, this paper introduce the basic concept of evolutionary robotics , then we introduce the main content of evolutionary robotics and some possible problems and corresponding solutions are proposed. We also detailedly introduce development of genetic algorithms and genetic programming at present.Secondly, this paper simply expound obstacle-avoidance behavior and traditional methods for obstacle-avoidance behavior implement. We implement obstacle-avoidance behavior learning by genetic algorithms, and this simulation result shows it's effective.Finally, we implement obstacle-avoidance behavior learning by genetic programming. We implement obstacle-avoidance rules'evolution simply using sensor information and motion. For limitations in genetic programming, such as size explosion in evolutionary process, convergence speed in genetic programming and adaptability of solution. We provide corresponding methods for eliminating these limitations: prevent size explosion in evolutionary process by restricting max-layernumber; by adding leaf node's mutation to make algorithms jump out of...