Abstract: | This project involved the replication of work on evolving adaptive robot controllers by Floreano and Mondada, reported at the Fourth International Conference on the Simulation of Adaptive Behavior (1996). They used a fixed topology neural network controller for which parameters were determined by chromosomes evolved using a genetic algorithm, each of which was evaluated on a Khepera robot. The strenghts of the synapses in the network are adjusted during the life of the robot according to evolved parameters.In this project the evolution was carried out in simulation, using a simulator that had been modified for this task. A number of corridor-following controllers were evolved and successfully transferred to reality. However, about 80% of experiments did not produce corridor-following behaviours. Two successful corridor-following controllers were analysed in terms of the dynamics of their synapses and also tested for robustness to changing environments.
|