Surrogate Modelling for Efficient Discovery of Emergent Population Dynamics

Abstract

Outcomes of simulating complex systems models, such as emergent properties and desirable system level behaviours, can be discovered via heuristic techniques such as Genetic Algorithms (GAs). Using simulation as the cost function evaluation for a GA (i.e. simulation guided search) is computationally expensive. Additionally the GA search process may require many generations before high quality solutions can be discovered. As such, simulation guided search can be considered high latency with respect to discovery of a range of high quality solutions. In this paper we experimentally demonstrate that the time to discovery of high quality solutions can be reduced through a low latency, hybrid GA search using a machine learning surrogate model trained to approximate simulation via large amounts of batched parallel simulation data generated in a HPC environment. Using a common population dynamics model optimised for GPU simulation by the FLAME GPU framework, we directly compare the hybrid approach with simulation guided search to understand the relationship between computational cost and quality of prediction. Our results indicate that given equivalent levels of simulation investment, results of equivalent quality can be obtained. The hybrid approach is however able to reduce the latency of the GA search process by shifting the computational cost of simulation to a highly parallel pre-search step used to train surrogate models.