Keele Research Repository

Previous works have used a novel hybrid network architecture to create deliberative behaviours to solve increasingly challenging tasks in two-dimensional and threedimensional artificial worlds. At the foundation of each is a static hand-designed neural network for robust and deliberative motion planning. This paper presents results from replacing the hand-designed motion-planning subnetwork with HyperNEAT. Simulations are run on the original two-dimensional world with, and without, relative position inputs and a multievaluation fitness function, thus assessing the relative performance of each strategy. The focus of this work is on solutions adaptable to general environments; following evolution, each strategy's performance is evaluated on 10,000 world configurations. The results demonstrate that although HyperNEAT was not able to achieve as robust results as a hand-design approach, the best strategy was comparable, with just a 3-4% drop in performance. Relative position inputs and the multievaluation fitness function were both significant in achieving superior general performance, compared to those simulations without.