Modelli di diffusione latente per la generazione di dati sintetici in modelli ad agenti

Agent-Based Models (ABMs) are widely used to study complex systems. By defining a (often stochastic) microscopic dynamic, they allow to draw macroscopic implications on the system as a whole. However, since ABMs follow a model-based approach and are often not differentiable, it is not possible to use micro-level data to estimate the set of rules, latent variables or parameters of an ABM. Representing a given arbitrary ABM as a differentiable program, akin to deep learning models, could provide a way to estimate their parameters, by improving their connection with real-world data. In this work, we propose a first step to tackle this limitation by developing a surrogate model that can approximate an unknown ABM's micro-level dynamic. As a surrogate model, we consider a generative model consisting of a conditional Latent Diffusion Model coupled with a Graph Neural Network. To train the model, we generated a synthetic dataset consisting of micro-level data from realizations of a target ABM, namely Schelling's Segregation. To check the quality of results, we constructed the conditional probability distributions from the samples generated by the trained model. By measuring the Earth Mover's Distance, we find that these distributions resemble the ones from the original model. Furthermore, typical system-level patterns of Schelling's Segregation could be observed when the generation process was iteratively run as a simulation on the entire system. The results of this work showcase that it is indeed possible to use machine learning techniques to reproduce an ABM dynamic, with little to no prior assumptions. In addition, this surrogate model is differentiable and could possibly be used in the future to estimate ABMs' parameters.