Evolutionary Learning of a Neural Robot Controller

摘要

In our research we use evolutionary algorithms to evolve robot controllers for executing elementary behaviours. This paper focuses on the behaviour of pushing a box between two walls. Successful execution of this behaviour is critically dependent on the robot's ability to distinguish between the walls and the box. In a recent study we already found that a two-layer recurrent network is more suitable for this task than a single-layer feedforward network [8]. In this paper, we conclude that this result still holds if a more general representation of a feedforward network is used. We also conclude that network architecture changing genetic operators may help reducing a starting network configuration to a smaller one, which is easier to train for the task at hand.