COMPARATIVE ANALYSIS OF METAMODELING TECHNIQUES BASED ON AN AGENT-BASED SUPPLY CHAIN MODEL

Abstract

Agent-based models comprise interacting autonomous entities, and generate both individual and emergent system-level outputs. These models generally have a large set of free parameters, whose impact on output needs to be explored. Considering also the need for replication due to stochasticity, a proper analysis requires a very large set of simulation runs. Therefore, obtaining a simpler representation of a simulation model (e.g., metamodel) can prove useful. We primarily focus on the potential utilization of various metamodeling approaches, namely Decision Trees, Random Forests, k-Nearest Neighbor Regression, and Support Vector Regression in predicting the two different types of outputs of an agent-based model. Results show that system-level outputs are predicted with higher accuracy compared to individual-level outputs under equal sample sizes. Although there is no single metamodeling technique performing best in all cases, we observe that support vector regression is more robust to the increase in the dimension of the problem.