Evolution of Computer Vision Subsystems in Robot Navigation and Image Classification Tasks

Real-time decision making based on visual sensory information is a demanding task for mobile robots. Learning on high-dimensional, highly redundant image data imposes a real problem for most learning algorithms, especially those being based on neural networks. In this paper we investigate the utilization of evolutionary techniques in combination with supervised learning of feedforward nets to automatically construct and improve suitable, task-dependent preprocessing layers helping to reduce the complexity of the original learning problem. Given a number of basic, parameterized low-level computer vision algorithms, the proposed evolutionary algorithm automatically selects and appropriately sets up the parameters of exactly those operators best suited for the imposed supervised learning problem.