Cognition is not computation; Evolution is not optimisation

Abstract

An overview is given of the role and relevance of Artificial Neural Nets (ANNs) as control systems for autonomous agents. Though ANNs can be used as computational input/output devices, cognition requires not this but rather some method of implementing dynamical systems. A wider class of ANNs incorporating temporal dynamics and feedback is presented, as one way to achieve this. These are difficult to design, and evolutionary approaches are a possible approach. Since evolving complex robot controllers inevitably takes a long time, one cannot afford to start afresh with each new problem, and an incremental adaptation approach will be necessary in the long term. This means that standard off-the-shelf optimising genetic algorithms are not appropriate unless adjusted to their new role.