This paper sets out a conceptual framework for the open-ended artificial evolution of complex behaviour in autonomous agents. If recurrent dynamical neural networks (or similar) are used as phenotypes, then a Genetic Algorithm that employs variable length genotypes, such as Inman Harvey's SAGA, is capable of evolving arbitrary levels of behavioural complexity. Furthermore, with simple restrictions on the encoding system that governs how genotypes develop into phenotypes, it may be guaranteed that IF an increase in fitness requires an increase in behavioural complexity, then it WILL evolve. In order for this process to be practicable as a design alternative, however, the time periods involved must be acceptable. The final part of this paper looks at general ways in which the encoding scheme may be modified to speed up the process. Experiments are reported in which different categories of scheme were tested against each other, and conclusions are offered as to the most promising type of encoding scheme for a viable open-ended Evolutionary Robotics.
Download compressed postscript file