Promoting Transfer of Robot Neuro-Motion-Controllers by Many-Objective Topology and Weight Evolution

Abstract

The ability of robot motion controllers to quickly adapt to new environments is expected to extend the applications of mobile robots. Using the concept of transfer optimization, this study investigates the capabilities of neuro-motion-controllers, which were obtained by simultaneously solving several source problems, to adapt to target problems. In particular, the adaptation comparison is carried out between specialized controllers, which are optimal for a single source motion problem, and nonspecialized controllers that can solve several source motion problems. The compared types of controllers were simultaneously obtained by a many-objective evolution search that is tailored for the optimization of the topology and weights of neural networks. Based on the examined problems, it appears that nonspecialized solutions, which are “good enough” in all the source motion problems, show significantly better transfer capabilities as compared with solutions that were optimized for a single source motion problem. The proposed approach opens up new opportunities to develop controllers that have good enough performances in various environments while also exhibiting efficient adaptation capabilities to changes in the environments.