Automatic Neural Robot Controller Design using Evolutionary Acquisition of Neural Topologies

In this paper we present an automatic design of neural controllers for robots using a method called Evolutionary Acquisition of Neural Topologies (EANT). The method evolves both the structure and weights of neural networks. It starts with networks of minimal structures determined by the domain expert and increases their complexity along the evolution path. It introduces an efficient and compact genetic encoding of neural networks onto a linear genome that enables one to evaluate the network without decoding it. The method uses a meta-level evolutionary process where new structures are explored at larger time-scale and existing structures are exploited at smaller time-scale. We demonstrate the method by designing a neural controller for a real robot which should be able to move continously in a given environment cluttered with obstacles. We first give an introduction to the evolutionary method and then describe the experiments and results obtained. 1 Evolutionary Acquisition of Neural Topologies Evolutionary Acquisition of Neural Topologies (ENAT) [6,7] is an evolutionary reinforcement learning system that is suitable for learning and adaptation to the environment through interaction. It combines meaningfully the principles of neural networks, reinforcement learning and evolutionary methods. The method introduces a novel genetic encoding that uses a linear genome of genes (nodes) that can take different forms. The forms that can be taken by a gene can either be a neuron, or an input to the neural network, or a jumper connecting two neurons. The jumper genes are introduced by the structural mutation along the evolution path. They encode either forward or recurrent connections. Figure 1 shows an example of encoding a neural network using a linear genome. As can be seen in the figure, a linear genome can be interpreted as a tree based program if one considers all the inputs to the network and all jumper connections as terminals. The linear genome has some interesting properties that makes it useful for evolution of neural controllers. It encodes the topology of the neural controller