From Toys to Tools - Calibrating Agent-Based Simulations for Business Decision Support

• Applicant: Professorin Dr. Catherine Cleophas
• Subject Area: Accounting and Finance ; Security and Dependability, Operating-, Communication- and Distributed Systems
• Term: from 2016 to 2022
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 313089115

Increasingly, decision-making in business confronts problems that are determined by heterogeneous groups of actors. One example is optimizing inventory and price controls for maximum revenue. The success of such an optimization depends on market demand, which is determined by the heterogeneous actors involved in the market, i.e., the firm's customers and competitors.A standard approach to supporting business decisions is to use simplified mathematical models to calculate optimal solutions. To evaluate these, agent-based simulations can create more realistic conditions. They can model the individual behavior of heterogeneous actors and the phenomena emerging, evaluate the consequences of interventions ceteris paribus, and compute outcome of a system's long-term evolution. To be applicable in practice, however, simulation models have to be validated rigorously on empirical data. Successful validation is achieved by calibrating the structure and parameters of the simulation model until the fit between simulation results and empirical observations is acceptable. When the available data does not support the straightforward calibration of input parameters or when validation yields ambiguous results, a dilemma arises. This is frequently true for complex, agent-driven systems, and it renders calibration difficult and laborious. To overcome this dilemma, in this project, we aim to create a methodological framework to effectively and efficiently calibrate agent-based simulations that can support business decisions. The overarching objective is to move the methodology of agent-based simulations in business research one step closer to that of simulations in the natural sciences: By enabling more rigorous, quantitative validation, new methods of calibration will make such simulations more credible and thereby make them a more useful research tool. The proposed project will combine methods from information systems engineering, business analytics, and machine learning: We will first survey the state of agent-based simulations in business decision support. Subsequently, we will generate novel calibration methods through mathematical modeling, data-mining, and machine-learning. For evaluation in different application domains, selected methods will be implemented and benchmarked in a modularized framework. A set of case-studies will ensure the application-oriented result evaluation.The resulting effective and efficient calibration methods will enable business researchers to better consider agent-driven decision-support problems in the light of complexity and uncertainty. In return, enabling agent-based simulations to fulfil more rigorous standards of validation will increase their credibility as a research tool.

• DFG Programme: Research Grants