Problem solving by global optimization: the rolling-stone neural network

The study of neural networks for solving optimization and constraint satisfaction problems has led to the k-out-of-n design rule. This rule allows for a systematic construction of the weight matrix and bias inputs of a recurrent network of Hopfield type. For efficiently finding solutions to the problem given, an appropriate neuron dynamic with optimization property has to be defined. There are several proposals like the Hopfield network or the Boltzmann machine. Some of these models get trapped in local minima or have arbitrary parameters. In this paper, a new neuron model is derived from the rolling-stone scheme, which is a global optimization method. The results of a simulation are compared to a Hopfield network for solving the N-queens problem.