Simulation-Based Inference in Agent-Based Models Using Spatio-Temporal Summary Statistics

In agent-based models (ABMs), traditional statistical inference faces challenges due to intractable likelihoods and computational costs. This study evaluates neural posterior estimation (NPE) and neural ratio estimation (NRE) for parameter inference in ABMs and compares them with approximate Bayesian computation (ABC). NPE and NRE are argued to be more efficient than traditional methods such as ABC and circumvent some of their limitations. The assessment of the methods focuses on the satisfaction threshold in Schelling’s model of residential segregation, including regions of high variance and non-equilibrium dynamics. As these simulation-based methods still require summary statistics as high-level descriptions of the ABM, we propose a general approach to construct them based on spatial and/or temporal information and evaluate how the different summary statistics affect performance. Both NPE and NRE generally outperform ABC regardless of summary statistics. Most notably, NRE excels when employing the most detailed spatio-temporal information, but adding spatial or temporal information alone is not always beneficial for NPE, NRE and ABC. This holds true for different training budgets and when estimating multiple parameters. Hence, the study underscores the importance of spatio-temporal information for accurate parameter inference in this ABM, but information redundancy can degrade performance as well. Therefore, finding optimal high-level descriptions to capture fundamental emergent patterns in the model through summary statistics might prove crucial in cases where the systems are governed by more complex behaviour.