Koopman Representations for Early Outbreak Warning and Minimal Counterfactual Intervention in Multi-Agent Epidemic Simulations

A new AI framework from computer scientist Florin Leon applies Koopman operator theory to multi-agent epidemic simulations for early outbreak warning and minimal intervention design. The model tracks agents with realistic mobility, heterogeneous susceptibility, and immunity-dependent viral progression. By projecting daily aggregated observables from early trajectory windows into a low-dimensional Koopman latent space—where dynamics evolve approximately linearly—the system enables short-horizon forecasting and outbreak risk classification via a random forest. Experiments near the system’s critical tipping points show strong early warning performance, with Koopman-derived features significantly improving class separation between major and minor outbreaks.

The most striking result comes from counterfactual analysis: keeping a single selected agent at home for just one day can reduce final attack rates and often shift the entire epidemic below the outbreak threshold. This suggests that targeted, minimal social distancing—rather than blanket lockdowns—could be highly effective if guided by such predictive models. The work, published on arXiv (2605.01803), demonstrates how combining Koopman representations with machine learning can turn complex multi-agent simulations into actionable public health intelligence, potentially allowing authorities to stop pandemics with surgical precision rather than broad restrictions.