An improved CUDA-based hybrid metaheuristic for fast controller of an evolutionary robot

Abstract

This paper proposes a novel parallel hybrid training approach to conceive an evolutionary robot. The proposed design aims to provide efficient behaviours to perform its tasks in a complex area such as walking toward a hidden destination. Embedded in robot brain, this training and evolution combination is typically accomplished by evolving considerable recurrent neural networks (RNNs) using an evolutionary strategy (ES). The effectiveness of this proposal is improved by employing CUDA technology that executes the evolutionary process of RNNs in a parallel way. The modifications applied are indicating to meet CUDA requirements in terms of CPU/GPU cooperation and memory management. Using a set of experiments performed by GPGPU-based physical simulator named open dynamics engine (ODE) and CUDA-based evolution, the effectiveness of the proposed parallel evolutionary training technique was validated for real movements of humanoid robots. This validation showed a promising speed-up, since this field requires very high powerful computational resources.