MIMIcking the COmplexity of agent-based models: the MIMICO Project

Abstract : Agent-based models (ABMs) are powerful tools for studying complex phenomena in which numerous heterogeneous entities interact non-linearly, but their systematic exploration is constrained by the computational cost of each run and by the opaque nature of the mapping between parameters and results. The MIMICO project addresses these two obstacles jointly: (1) surrogate models relax the computational bottleneck by learning a fast approximation of the scenario-to-result relationship, while (2) recent advances in Explainable Artificial Intelligence (XAI) make it possible to analyze the behavior of these learned models themselves, beyond the original simulator. This dual perspective leads us to focus on the decision boundaries of the models: the regions of the parameter space where the model’s response becomes unstable. Our working hypothesis is that these boundaries carry information about the dynamics of the target system, independently of the predictive accuracy of the model that reveals them: a stance that departs from the usual accuracy-centric evaluation of surrogate models, since even an imperfect surrogate can be informative about the system’s dynamics if it correctly locates its instabilities. To operationalize this idea, MIMICO proposes a two-stage pipeline designed to optimize the use of a fixed simulation budget. A first, inexpensive model-based screening built on naive sampling or expert knowledge consumes a limited share of the budget to build an initial parameter-result space and discard trivial regions. This space is then used to train a surrogate model, whose fast inference identifies the most unstable or informative sub-regions; the remaining budget is reallocated toward the original model to reinforce and statistically confirm these regions rather than sampling the parameter space uniformly. This approach has already produced concrete results across several case studies. On a stochastic predator-prey ABM, reallocating the simulation budget between an initial screening (3,250 runs) and a refined, focused batch (13,000 runs) reduced the share of trivial outcomes from 60% to 16%, and revealed, through Sobol’ indices, that 69% of the output variance stems from higher-order parameter interactions invisible to linear analysis, exposing sharp tipping points and a "metabolic trap" governing ecosystem coexistence. On the real-world MAELIA agro-ecological platform, a dedicated random-forest-supervised HSIC-ANOVA extended this decision-oriented sensitivity analysis to a genuinely hierarchical, mixed-variable parameter space, an obstacle classical global sensitivity analysis cannot handle. On a spatial double-auction market of adaptive, reinforcement-learning agents, the same exploration principles scaled to learning agents, showing how spatial structure and market design jointly govern the emergence of industrial symbiosis. Finally, on the explainability side, a SHAP-based methodology for characterizing model decision behavior illustrates the broader applicability of explanation-driven diagnostics beyond ABMs, while a state-of-the-art survey structures the emerging XAI-surrogate field and sets a research agenda for embedding explainability throughout the simulation-exploration workflow.