Chaotic neural network controlled by particle swarm with decaying chaotic inertia weight for pattern recognition

Abstract

This study introduces a new type of chaotic neural network, which is built upon perturbed Duffing oscillator. The neurons in this network behave collectively based on a modified version of Duffing map. The proposed neural processor can act chaotically at some areas of the state space. The network has some parameters, which can be adjusted for the system to behave either chaotically or periodically. This nonlinear network adopts the bifurcating behavior of the chaotic Duffing map for the most covered search in the neuronal search space. The neuron’s search space is controlled by swarming in the parameter space to settle the parameters of the network into the critical parameters. Swarming of the parameters is based on particle swarm optimization heuristic. The modified particle swarm adopts a decaying inertia weight based on chaotic logistic map to fast settle down into the attractors of periodic solutions. At last, the swarm-controlled neurochaotic processor is applied to build three models to control parameters of the network. Each model is trained to recognize a set of binary patterns that are as the form of alphabetic letters as a classical pattern recognition problem. A comparison study is then conducted among these three models, Hopfield network and a modified Hopfield model, which demonstrate all three models outperform Hopfiled model and are competitive and in most cases outperform the modified Hopfield model.