Abstract
We here propose a new evolutionary approach with learning to create a variety of behavioral patterns in autonomous robots. The conventional evolution or learning is to optimize a cost function such as.frtness function and error function. In practice, the robot encounters situations where exist multiple so-lutions having quite similar fitness or error values. The optimum solution is generally selected, while oth-ers are eliminated, even if the difference in the fitness or error is very little between the solutions. This causes an essential problem for behavior-based robots. Ideally, the robot should be able to select one of the behaviors by perceiving a slight difference in the sensory information, but the ability is lost. To over-come this problem, we introduced a structural learning during the evolution of neural network ensemble (NNE). Motor outputs were generated by summing outputs of component neural networks of an NNE, and they were trained to segregate each other by negative correlation learning between generations. In ex-periment, each component network exhibited different functionality, producing a variety of behaviors as a whole. The proposed evolution of NNE with negative-correlation learning thus can be a practical solution for the plasticity-stability problem in robotics.
