Fault-tolerance by regeneration: using development to achieve robust self-healing neural networks

Opposed to the standard paradigm of ´fault-tolerance by redundancy´, ontogeny offers the possibility to engineer artificial organisms which can re-grow faulty components. Similar to what happens in nature, organisms display self-healing: a homeostatic process which allows proper operation while suffering faults. In this paper we present a system which evolves developing spiking neural networks capable of controlling simulated Khepera robots in a wall avoidance task. Development is controlled by a decentralized process executed by each cell´s identical growth program. To test the system´s self-healing capability, networks are (1) subjected to random faults during development and (2) mutilated during operation. Results demonstrate how development can (i) rapidly produce proper neuro-controllers and (ii) re-grow neurons to recover normal operation. These results show that development, originally proposed to increase the evolvability of large phenotypes, also allow the production of artifacts with sustained fault-tolerance. These artifacts would be especially well-suited for tasks that require long periods of operation in absence of external maintenance.