Mapping Complex Agent-Based Model Parameter Spaces Using Metamodels and Mixed Adaptive Sampling

Large-scale agent-based models (ABM) of complex adaptive systems often have a high number of input parameters and complicated agent interactions affecting the performance of the system. Executing traditional design of experiments to select points from the parameter space could require considerable simulation runs to effectively analyze the impact of input parameters on the system output. Sequential sampling is a promising technique for reducing simulation runs by training a machine learning metamodel to emulate the relationship between input parameters and ABM outputs while leveraging the metamodel to strategically select points to explore next in the parameter space. While previous studies have applied sequential sampling to few small ABMs, this research extends the approach to large-scale ABM and compares mixed adaptive sampling and pure sequential sampling algorithms. These algorithms were used to train a random forest metamodel emulating the Segregation model and a larger stylized ABM of a crowd logistics platform. Results indicate that the pure sequential sampling algorithm can lead to excessive clustering of design points in highly nonlinear regions and insufficient exploration of the parameter space, while the mixed adaptive sampling algorithm ensures broader coverage of the parameter space while still refining complex regions. The mixed adaptive sampling algorithm was then used to generate feature importance and partial dependence plots for the crowd logistics ABM, thereby demonstrating the usefulness of this approach not only in reducing computational requirements for ABM experimentation but also in enhancing ABM output interpretation.