Simulation-based optimization of an agent-based simulation

Optimization of an agent-based simulation (ABS) bears specific challenges. It is demonstrated in this paper that mainstream simulation-based optimization (SBO) approaches often do not perform well in such a setting, sometimes hardly outperforming a mere random search. Two new algorithms for SBO which combine superior solution quality with high resource efficiency and reliability for such problems are presented: an evolutionary algorithm called "neighbourhood elite selection" (NELS) with a specific selection mechanism which prevents premature clustering, and a hybrid algorithm which combines NELS with the popular best-inclass algorithm Simultaneous Perturbation Stochastic Approximation (SPSA). Those two algorithms are designed to perform well for problems which show typical properties of an agent-based simulation, a field that has largely been neglected so far, but should structurally also be universally applicable for other simulation-based optimization problems as well. In contrast to present literature, specific emphasis lies on the dynamic control of how many replications of the simulation are required for each solution brought up during the optimization run in order to make efficient use of the scarce simulation resources. The algorithms are benchmarked against the academic best-in-class optimization algorithm SPSA. A sketch of practical case studies is provided, showing how the optimization of an ABS can be used to help solve business decision problems like price optimization for a mobile phone operator.