Evolution of NN for the Design of Virtual Agents under Limited Resources Constraints

This paper reports findings on a process for evolving neural networks capable of designing virtual agents. In particular, these virtual agents operate on a system of constrained resources, making the allocation of resources among them an important design consideration. The evolution system optimizes both neural network topologies and connection weights. The experimental results included indicate that the problem cannot be satisfactorily solved by independently evolving topologies or weights. In addition, because there is lack of a priori evidence pointing towards an optimal solution, the evolutionary process used here is able to find better solutions than either global (back propagation) or local (Hebbian) neural network learning algorithms.