An experimental design framework for evolutionary robotics

Based on the failures of work in the area of machine intelligence in the past, a new paradigm has been proposed: for a machine to develop intelligence it should be able to interact with and survive within a hostile dynamic environment. It should therefore be able to display adaptive behaviour and respond correctly to changes in its situation. This means that before higher cognitive properties can be modeled, the modeling of the lower levels of intelligence would be achieved first. Only by building on this platform of physical and mental abilities may it be possible to develop true intelligence. One train of thought for implementing this is to control and design a robot by modeling the neuroethology of simpler animals such as insects. This thesis outlines one approach to the design and development of such a robot, controlled by a neural network, by combining the work of a number of researchers in the areas of machine intelligence and artificial life. It involves Rodney Brooks’ subsumption architecture, Randall D. Beer’s work in the area of computational neuroethology, Richard Dawkins’ work in the area of biomorphs and computational embryology and finally the work of John Holland and David Goldberg in genetic algorithms. This thesis will demonstrate the method and reasoning behind the combination of the work of the above named researchers. It will also detail and analyse the results obtained by their application