Evolutionary autonomous agents whose behavior is determined by a neurocontroller "brain" are a promising model for studying neural processing. Nevertheless, they are missing an important quality prevalently found in all levels of natural systems, fault-tolerance, the lack of which results in overly simplistic neurocontrollers. We present a way of modifying a given evolutionary process for encouraging the creation of neurocontrollers that manifest high levels of fault-tolerance, using both direct and incremental evolutions. The evolved neurocontrollers are more robust not only against the faults introduced during the evolutionary process, but also against much more extreme ones. This robustness poses a great challenge for an analysis of the workings of the neurocontrollers, the latter being the focus of this paper: We utilize the Multi-perturbation Shapley value Analysis (MSA) to uncover the important neurons, as well as the interactions between them, revealing the mechanisms underlying the evolved fault-tolerance.
展开▼
