Evolving Simplistic Neural Controllers toward Adaptation to Internal Perturbations
Résumé
This paper presents a study that consists to evolve neural controllers confronted to various internal perturbations of legged robots: transmission breaking, leg loss, and mechanical wear. Methods of evolutionary robotics are applied to the control of dynamic equilibrium of a legged and wheeled robot that could be confronted to transmission breaking on its driving wheel. These methods consist in evolving synaptic weights of neural controllers (recurrent or none) by a genetic algorithm. A least biased fitness, without any specific problem simplifications, is used to allow comparison with further methodologies. We start by evolving non perturbed individuals and obtain periodic gait with simple multilayer perceptron neural model. Also, adding an internal perturbation during the evolution process gave an interesting switching behaviour. We finally discuss the results to show methods limits and elaborate solutions to solve some issues.