Skip to content
Publications

Preprint · Research Square · June 11, 2026

Structural support constraints prevent invalid surrogate predictions in branched agent-based simulations

Alaa Aldin Hajjar, Armen Beklaryan

Abstract

Abstract Branched simulators can impose deterministic output support: some outputs are possible only under specific simulator branches. Standard surrogate regressors may ignore this logic and predict values that the simulator itself could never generate. We evaluate a structural-support validation workflow for branched agent-based simulations, combining branch-aware representation, simulator-derived support masking, grouped validation by simulator invocation, capacity control, and model-family comparison. The testbed is an affordability-constrained residential segregation simulator in which blockage outcomes are structurally zero whenever the economic-feasibility branch is inactive. The dataset contains 400 simulator invocations expanded into 1600 condition-specific rows. The hard mask enforces the known no-economy zero-support rule, reducing invalid blockage predictions to zero while preserving useful economy-active accuracy. In grouped five-fold validation, the selected 137-parameter masked multilayer perceptron achieves root-mean-square errors of 0.0494 for blocked-mobility burden and 0.0746 for final affordability-blocked share, with masked tree and two-stage baselines providing close comparisons. Targeted verification of 40 response-map points with 120 direct simulator replicates gives mean regime agreement of 0.7417. Dense diagnostic mapping is estimate-efficient: surrogate inference takes 0.0566 s, while direct simulation is estimated at 14,209--361,228 s. The response maps are simulator-internal diagnostics, not real-world causal or policy evidence.

Simulation Techniques and ApplicationsAdvanced Multi-Objective Optimization AlgorithmsMulti-Agent Systems and Negotiationsurrogate modelsmultilayer perceptronresponse map

Related publications