Single multiplicative neuron model as an alternative to multi-layer perceptron neural network

The paper presents an approach to nonlinear control of complex dynamic systems using artificial neural networks (ANN). A novel form of ANN, namely, single multiplicative neuron (SMN) model is proposed in place of more traditional multi-layer perceptron (MLP). SMN derives its inspiration from the single neuron computation model in neuroscience. SMN model is trained off-line, to estimate the network weights and biases, using a population based stochastic optimization technique, namely, particle swarm optimization (PSO). On-line learning of SMN based on gradientdescent has been presented. The development of the control algorithms is illustrated through the hardware-in-the-loop (HIL) implementation of magnetic levitation and a cart position control system in Lab VIEW environment. SMN structure was much simpler than the MLP for similar performance. The simple structure and faster computation of SMN have the potential to make it a preferred candidate for controller implementation for real-life complex dynamic systems.