Swarm optimization with instinct-driven particles

In particle swarm optimization (PSO), each particle stores a candidate solution, and stochastically modifies its candidate over time, based on the best solution found by neighboring particles, and based on the best solution found by the particle itself. We present an enhancement of PSO in which each particle's behavior is also influenced by a third component which is meant to represent the particle's innate instinct-level intelligence. The instinct component is a function of the intrinsic "goodness" of each dimension of the particle's candidate solution and has similarity to the goodness measure used in ant colony methods. We apply our modified-PSO to several 100-variable 900-clause instances of weighted max-sat, comparing our performance to standard PSO and to the Walk-Sat algorithms. We use an aging scheme in which the weight of a clause increases gradually if it is not satisfied. We find that our modified-PSO produces significant improvements over standard PSO and yields performance comparable to Walk-Sat.