TOWARDS EVOLUTIONARY DESIGN OF COMPLEX SYSTEMS INSPIRED BY NATURE

This paper presents first steps towards evolutionary design of complex autonomous systems. The approach is inspired in modularity of human brain and principles of evolution. Rather than evolving neural networks or neural-based systems, the approach focuses on evolving hybrid networks composed of heterogeneous sub-systems implementing various algorithms/behaviors. Currently, the evolutionary techniques are used to optimize weights between predefined blocks (so called Neural Modules) in order to find an agent architecture appropriate for given task. The framework, together with the simulator of such systems is presented. Then, examples of agent architectures represented as hybrid networks are presented. One architecture is hand-designed and one is automatically optimized by means of evolutionary algorithm. Even on such a simple experiment, it can be observed how the evolution is able to pick-up unexpected attributes of the task and exploit them when designing new architecture