Reconfigurable analogue hardware evolution of adaptive spiking neural network controllers

Abstract

This paper details the hardware evolution of adaptive Spiking Neural Network (SNN) controllers, implemented on a network of cascaded Field Programmable Analogue Arrays (FPAAs). The fixed architecture, feed forward SNNs are trained using a Genetic Algorithm (GA). An obstacle avoidance simulated robotics controller application is chosen to test the FPAA reconfigurable hardware evolution platform. Evolved behaviours, resulting from FPAA-based SNN controllers, are compared with those obtained using software-based SNN implementations. Results presented indicate the emergence of effective behaviours and adaptation to environmental change.