An Experimental Comparison of Weight Evolution in Neural Control Architectures for a ‘Garbage- Collecting’ Khepera Robot

Nolfi in [1] presented an experimental comparison of weight evolution in five different neural network architectures (feed-forward vs. recurrent, modular vs. non-modular) for the control of a Khepera robot, which had to pick up objects with its gripper arm and place them outside an arena. Best results were achieved with a so-called emergent modular architecture. This paper extends Nolfi’s experiments by comparing the original five architectures and three variations of the emergent modular architecture on both the original task and a more difficult variation of it.