PEvoRNN: Extended CUDA-based hybrid parallel evolutionary robot controller

Development and simulation of artificial biological models require a considerable computational complexity which can be surmounted using Modern Graphic Processing Units (GPUs), since these devices adopt parallel multi-core architecture which favors the inherently data-parallel nature. In this paper, we present an extended version of our hybrid model presented in [1], [2], based on a high tuning CUDA parameter (Compute Unified Device Architecture) for the parallel evolutionary strategy (ES) and a recurrent neural network (RNN), called PEvoRNN. This proposal takes the advantages of the GPU memory hierarchy effective use, the best parallelism management by investigating tuning CUDA parameters and the optimal GPU-based coding, to generate optimal trajectories of a humanoid robot using GPU Accelerator at multiple levels. PEvoRNN represents a controller which monitors the movement and the evolution of a 3D humanoid robot simulated on the open dynamic engine simulator (ODE). Moreover, this proposal makes feasible and tractable the integration of multiple complex evolutionary methods for heterogeneous training locomotion process. The effectiveness of the proposed parallel tuned evolutionary training technique was validated for real movements in terms of a promising speedup, since this field requires powerful computational resources.