Measuring the complexity of the real environment with evolutionaryrobot: evolution of a real mobile robot Khepera to have a minimalstructure

摘要

A real mobile robot was evolved in several kinds of realenvironment with various complexities. The fitness function of GAoperations included the complexity measure of the control structure,i.e., individuals with a simpler structure obtained a higher score.Evolution lead to developing a robot with a minimal structure sufficientto live and perform tasks in the given environment. The minimalstructure itself lets one perceive easily the control scheme (orskeleton) of the robot for the given environment, and suggests a designscheme for more economical robots. Moreover, the value of the complexitymeasure of the control structure after evolution in a certainenvironment could be used as an index of complexity of the environment.The authors used a neural network with variable numbers of connectionsfor a control structure of a real mobile robot Khepera. The networksummed up signals from eight proximity sensors to generate outputs totwo motors. The robot was evolved in four different kinds of environmentwith various complexities to perform the task of navigation withobstacle avoidance. The number of connections was used for thecomplexity measure of the control structure, which was included in thefitness function. After evolution, robots with a minimal number ofconnections for a given environment were indeed developed. The number ofconnections obtained was lower in a simpler environment, showing thefeasibility to use the complexity measure as a complexity index of agiven environment