Optimal Experimental Designs via Particle Swarm Optimization Methods

Particle swarm optimization (PSO) method is relatively new, simple yet powerful and widely used in applied fields. However PSO does not seem to have made an impact in mainstream statistical applications hitherto. We propose variants of the PSO method to find optimal experimental designs for both linear and nonlinear models in a novel way. We show that the PSO method can simply generate many types of optimal designs very quickly, including optimal designs under a non-differentiable criterion such as minimax optimal designs where effective algorithms to generate such designs have remained elusive to date.