Evolutionary multiobjective optimization for automatic agent-based model calibration: A comparative study
MetadataShow full item record
Institute of Electrical and Electronics Engineers
Model calibrationAgent-Based ModelingEvolutionary multiobjective optimization
I. Moya et al.: EMO for Automatic ABM Calibration, IEEE ACCESS, VOLUME 9, 2021, 55284-55299. DOI: [10.1109/ACCESS.2021.3070071]
SponsorshipSpanish Agencia Estatal de Investigacion; Andalusian Government; University of Granada; European Commission PGC2018-101216-B-I00 P18-TP-4475 A-TIC-284-UGR18; Spanish Government RYC-2016-19800
Complex problems can be analyzed by using model simulation but its use is not straight-forward since modelers must carefully calibrate and validate their models before using them. This is specially relevant for models considering multiple outputs as its calibration requires handling different criteria jointly. This can be achieved using automated calibration and evolutionary multiobjective optimization methods which are the state of the art in multiobjective optimization as they can find a set of representative Pareto solutions under these restrictions and in a single run. However, selecting the best algorithm for performing automated calibration can be overwhelming. We propose to deal with this issue by conducting an exhaustive analysis of the performance of several evolutionary multiobjective optimization algorithms when calibrating several instances of an agent-based model for marketing with multiple outputs. We analyze the calibration results using multiobjective performance indicators and attainment surfaces, including a statistical test for studying the significance of the indicator values, and benchmarking their performance with respect to a classical mathematical method. The results of our experimentation reflect that those algorithms based on decomposition perform significantly better than the remaining methods in most instances. Besides, we also identify how different properties of the problem instances (i.e., the shape of the feasible region, the shape of the Pareto front, and the increased dimensionality) erode the behavior of the algorithms to different degrees.