A multiobjective optimisation approach for the dynamic inference and refinement of agent-based model specifications

Despite their increasing popularity, agent-based models are hard to test, and so far no established testing technique has been devised for this kind of software applications. Reverse engineering an agent-based model specification from model simulations can help establish a confidence level about the implemented model and in some cases reveal discrepancies between observed and normal or expected behaviour. In this study, a multiobjective optimisation technique based on a simple random search algorithm is deployed to dynamically infer and refine the specification of three agent-based models from their simulations. The multiobjective optimisation technique also incorporates a dynamic invariant detection technique which serves to guide the search towards uncovering new model behaviour that better captures the model specification. The Non-dominated Sorting Genetic Algorithm (NSGA-II) was also deployed to replace the random search algorithm, and the results from both approaches were compared. While both algorithms revealed good potential in capturing the model specifications, the pure exploratory nature of random search was found more suitable for the application at hand, compared to the balanced exploitation/exploration nature of genetic algorithms in general.