Chaotic simulated annealing in multilayer feedforward networks

This paper presents a method of chaotic simulated annealing for avoiding and subsequently escaping from local minima in the training of multilayer feedforward neural networks. A modified form of the standard simulated annealing algorithm is implemented using both Gaussian random numbers and various types of strange chaotic attractors for perturbation of network weight parameters. Specifically, the attractor generated by the logistic equation, Henon´s (1976) attractor Rossler´s attractor, and the Lorenz attractor are used at different initial conditions and parametric variations for chaotic perturbations. The variance fractal dimension is used as a quantitative measure of the geometric properties of the strange chaotic attractors. It is shown that, for this application, chaotic simulated annealing using the logistic equation is up to 600 percent faster than conventional simulated annealing with Gaussian random numbers