Discrete Multi-Valued Particle Swarm Optimization

We present a new optimization technique based on the Particle Swarm Optimization algorithm to be used for solving problems with unordered discrete-valued solution elements. The algorithm achieves comparable performance to both the previous modification of PSO for binary optimization and to genetic algorithms on a standard set of benchmark problems. Performance is maintained on these problems re-encoded with ternary solution elements. The algorithm is then tested on a simulated discrete-valued unsupervised robotic learning problem and obtains competitive results. Various potential improvements, modifications, and uses of the algorithm are discussed.