Discretized ISO-learning neural network for obstacle avoidance in reactive robot controllers

Isotropic sequence order learning (ISO-learning) and its variations, input correlation only learning (ICO-learning) and ISO three-factor learning (ISO3-learning) are unsupervised neural algorithms to learn temporal differences. As robotic software operates mainly in discrete time domain, a discretization of ISO-learning is needed to apply classical conditioning to reactive robot controllers. Discretization of ISO-learning is achieved by modifications to original rules: weights sign restriction, to adequate ISO-learning devices outputs to the usually predefined kinds of connections (excitatory/inhibitory) used in neural networks, and decay term in learning rate for weights stabilization. Discrete ISO-learning devices are included into neural networks used to learn simple obstacle avoidance in the reactive control of two real robots. (C) 2008 Elsevier B.V. All rights reserved.