Use of neuroevolution for neural network policies search for robotic arm

Abstract

An approach to using neuroevolution to find neural network policies for the task of positioning a robotic arm is considered. As a rule, robotic problems have relatively large solution spaces, so here neuroevolutionary algorithms are a good alternative to traditional methods of deep machine learning. A neuroevolutionary algorithm automatically develops neural networks for a specific task and environment. The advantage is that it is only necessary to define the desired behavior abstractly, and the algorithm optimizes the artificial neural network as much as possible to fulfill the requirements. The considered NEAT algorithm allows processing multidimensional state and action spaces, providing flexibility to control complex robot arm movements. It is also capable of detecting control policies that exhibit unpredictable behavior that is not clearly programmed by human engineers. Neuroevolution allows multiple neural networks to be evaluated in parallel, providing efficient exploration of the search space. The operation of the algorithm was investigated in an experiment conducted in a two-dimensional environment with a robotic arm for the positioning task.