Evolutionary Re-Adaptation of Neurocontrollers in Changing Environments

Evolutionary robotics is an interesting novel approach to shape the control system of autonomous robots. This explores issues related to readaptation in changed environments of a population of evolved individuals. Experimental studies are reported for genetic evolution of neurocontrollers that have developed the ability to perform homing navigation for battery recharge of miniature mobile robot. It is shown that re-adaptation to important changes in the environment is very rapid and does not disrupt previously acquired knowledge. The results are discussed in relation to the internal representation of the neurocontroller and to the variability within the population. 1 Evolutionary Shaping of Autonomous Robots An autonomous robot can be seen as an arti cial organism capable of selforganising its own behaviour according to environmental constraints in order to maintain its own viability without human intervention. In recent years, the analogy between autonomous robots and biological organisms has generated a novel approach, also known as \Behaviour{Based Robotics" [10], to program and understand robot behaviour in unknown and unpredictable environments. This approach is quite di erent from the classical AI approach that attempted to emulate human reasoning by building large and complex planning systems which failed to deliver the expected results [1]. In Behaviour{Based Robotics emphasis is put on speed, robustness, low-cost, and incremental development of modular controllers. Here, researchers seek inspiration from biology, trying to understand and reproduce the smart and simple mechanisms that allow animals to survive in their own environment [13]. Typically, the control system of such robots is based on parallel and distributed processes interconnected by modi able links. One example is the wellknown subsumption architecture [2]. Several successful results have been also reported using neural networks; e.g., see [15, 14, 3, 11] for an overview of the eld. Neural networks, in particular, o er interesting advantages if one is interested in biological inspiration and adaptation: a) They facilitate knowledge-tranfer