Investigation of simply coded evolutionary artificial neural networks on robot control problems

One of the advantages of evolutionary robotics over other approaches in embodied cognitive science would be its parallel population search. Due to the population search, it takes a long time to evaluate all robot in a real environment. Thus, such techniques as to shorten the time are required for real robots to evolve in a real environment. This paper proposes to use simply coded evolutionary artificial neural networks for robot control to make genetic search space as small as possible and investigates the performance of them using simulated robots. Two types of genetic algorithm (GAs) are employed, one is the standard GA and the other is an extended GA, to achieve higher final fitnesses as well as achieve high fitnesses faster. The results suggest the benefits of the proposed method.

• Type: Conference
• Year: 2008
• Where: CEC
• Authors: Yoshiaki Katada, Jun Nakazawa
• Keywords: Artificial Intelligence, CEC 2008, Parallel Population Search, Population Search, Real Environment