Learning for Analysis and Calibration in Agent-Based Simulation

For being able to derive valid results and knowledge from an agent-based simulation ABS, great care needs to be put in the model development process. The main problem does not relate to particular specification and implementation languages, but more generally to the (missing) link between macroscopic aggregate behavior and microscopic agent-level model. Although one major advantage of ABS is that the observation element ”individual” equals the modeling element ”agent”, it is not clear what properties and behavior of the real-world agent actually lead to the measurable aggregate values. Additionally, there is the risk that based on ”tuning” on the local level, agent models are able to reproduce the global level outcome, that are not valid at all. The problem is even more aggravated by the fact that empirical data is only partially available, mainly on the aggregated level, and is mostly quite noisy. Using this data for calibration and validation can distort the simulation behavior and thus make it impossible to create a simulation model with the credibility that we want. For example there may be no model structure and parameter setting for producing sufficient correspondence between real-world and simulation-produced data, not due to the model, but due to noise in the empirical data. Knowing the hidden limitations of the available data is highly valuable, especially when automatic optimization techniques for model calibration are used based on this data. In this paper we explore the use of agent learning techniques as a means for analyzing the effect of particular agent decision model structures on simulation model validity for evaluating its limits according to the available data. A special focus is put on the introduction of heterogeneity in the agent decisions as this highly influences the complexity of any optimization-based calibration methods as it exponentiates the number of parameters that have to be dealt with. Our approach will be illustrated using a shopping behavior model that possesses a relatively simple agent decision making structure – basically discrete choice –, yet is quite challenging due the overall project goal of actually predicting overall and individual shop turnovers.