The Effect of Evaluation Time Variance on Asynchronous Particle Swarm Optimization

Optimizing computationally intensive models of real-world systems can be challenging, especially when significant wall clock time is required for a single evaluation of a model. Employing multiple CPUs is a common mitigation strategy, but algorithms that rely on synchronous execution of model instances can waste significant CPU cycles if there is variability in the model evaluation time. In this paper, we explore the effect of model run time variance on the behavior of PSO using both synchronous and completely asynchronous particle updates. Results indicate that in most cases, asynchronous updates save considerable time while not significantly impacting the probability of finding a solution.