Fractal Fitness Landscape and Loss of Robustness in Evolutionary Robot Navigation

An autonomous robot Khepera was simulated with a sensory-motor model, which evolves in the genetic algorithm (GA) framework, with the fitness evaluation in terms of the navigation performance in a maze course. The sensory-motor model is a developed neural network decoded from a graph-represented chromosome, which is evolved in the GA process with several genetic operators.It was found that the fitness landscape is very rugged when it is observed at the starting point of the course. A hypothesis for this ruggedness is proposed, and is supported by the measurement of fractal dimension. It is also observed that the performance is sometimes plagued by Loss of Robustness, after the robot makes major evolutionary jumps. Here, the robustness is quantitatively defined as a ratio of the averaged fitness of the evolved robot navigating in perturbed environments over the fitness of the evolved robot in the referenced environment.Possible explanation of robustness loss is the over-adaptation occurred in the environment where the evolution was taken place. Testing some other possibilities for this loss of robustness, many simulation experiments were conducted which smooth out the discrete factors in the model and environment. It was found that smoothing the discrete factors does not solve the loss of robustness. An effective method for maintaining the robustness is the use of averaged fitness over different navigation conditions.The evolved models in the simulated environment were tested by down-loading the models into the real Khepera robot. It is demonstrated that the tendency of fitness values observed in the simulation were adequately regenerated.