Automatic model decomposition and reuse in an evolutionary cognitive mechanism

This paper addresses the problem of automatically obtaining primitives of the models an evolutionary cognitive mechanism is producing for a robot through its real time interaction with the world. The models are instantiated as Artificial Neural Networks (ANNs) and the objective is to obtain ANNs that cooperate in the process of modelling complex functions. An algorithm where the combination of networks takes place at the phenotypic or functional level is proposed. Thus, a population of networks that are automatically classified into different species depending on the performance of their phenotype is evolved, and individuals from each species cooperate forming a group to obtain a complex output. The components that make up the groups are basic ANNs (primitives) and may be reused in other modelling processes as seeds or combined to generate new solutions. The parameter that reflects the difference between ANNs is their affinity vector, the value which is automatically created and modified for each ANN through a competition based clustering process within the evolutionary process. The main objective of this approach is to explore one path to gradually model complex functions similar to those that arise when obtaining environment or internal models within robotic cognitive systems.