Emulation Strategies for Epidemic Agent-Based Models

Item: Emulation Strategies for Epidemic Agent-Based Models
Rating: 4
Author: PostSingularity Institute

by Gerald Munetsi · View the channel on PSI

Formal & Physical Sciences Medicine & Health MBNPublic health and preventive medicine MBNSEpidemiology and Medical statistics

🎙 Gerald Munetsi 👥 424 📅 September 22, 2026 ⏱ 35 min 👁 0 📄 original study 🧭 2026-09-22

Available in: English (current) Français

Keywords

emulationsurrogate modelsagent-based modelsneural networksnear-threshold sampling

Summary

This seminar, part of the SEMA Africa series, is presented by Gerald Munetsi, a researcher and data scientist. The talk addresses the computational burden of agent-based models (ABMs) in infectious disease epidemiology, which are highly detailed but slow to run. Munetsi introduces emulation (surrogate modelling) as a solution: training a cheap model like a neural network to approximate the input-output behaviour of the ABM, enabling fast parameter exploration for tasks like Bayesian calibration and sensitivity analysis. He then presents his proof-of-concept work at Cambridge, which aimed to improve emulation reliability. The study used a Barabási–Albert network model with 100,000 agents and a target reproduction number between 0 and 10. Two key strategies were proposed: near-threshold sampling (NTS), which concentrates training data near the epidemic threshold where dynamics are complex, and data augmentation, which generates synthetic data by perturbing existing inputs/outputs. Using multilayer perceptrons, the NTS-based emulators outperformed uniform and Latin hypercube sampling in terms of predictive accuracy, with lower relative mean absolute error and tighter confidence intervals. Emulators were 33 times faster than direct ABM evaluations. Data augmentation further improved accuracy. The emulators also performed better in approximating the ABM ground truth and generalised to out-of-distribution data. Limitations include the simplicity of the ABM and the need for expansion to more complex models. Future directions include Bayesian neural networks for uncertainty quantification and application to SEMA’s ongoing ABM projects. The talk concludes with potential benefits for African modelling, including faster calibration and high-resolution spatial/demographic modelling, followed by a Q&A session.

250 words

Critical Evaluation

Value of the Information & Strength of the Argument

The presentation provides a clear and structured argument for using emulation to accelerate agent-based models. The rationale is well-founded: ABMs are computationally expensive, hindering their use in calibration and scenario analysis. The proposed solutions (NTS and data augmentation) are innovative and grounded in existing literature, addressing concrete limitations of current emulation methods. The results are systematically presented with comparisons to baseline methods (uniform and LHS), showing clear improvements in predictive accuracy and computational speed. The speaker supports his claims with figures demonstrating approximation quality and generalisation. The argumentation is logical and acknowledges limitations, including the use of a simplified network model. The interactive Q&A further clarifies methodological details, such as the choice of MLP and the single training run requirement. Overall, the value lies in providing practical strategies to make ABM emulation more reliable, with potential high impact in resource-limited settings.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigour is high: the study follows a proper methodology with parameter space design, sampling, training/validation/test splits, and comparison metrics. The use of neural networks is justified, and the data augmentation concept is borrowed from computer vision. However, the presentation does not cite specific external references beyond the Barabási–Albert model and the general concept of emulation, which limits the transparency of underlying sources. The title accurately reflects the content and the talk delivers on its promise. The speaker is credible as a specialist in infectious disease modelling. No audience comments were provided for analysis, so no trends can be reported.

257 words

Title / Content Match

The title accurately reflects the content, which focuses on strategies for emulating agent-based models in epidemics, including sampling methods and data augmentation.

Quality & Reliability

8/10

The presentation is a proof-of-concept study based on a peer-reviewed methodology, with clear objectives, methods, and results. The speaker is an infectious disease modelling specialist, and the content aligns with established scientific practices. However, the study uses a simplified model and lacks external validation, so a score of 8 reflects strong internal coherence and expertise.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Contribution & Novelties

The presentation introduces a novel sampling strategy (near-threshold sampling) specifically designed for epidemic ABM emulation, addressing the challenge of learning dynamics near the epidemic threshold. Data augmentation, borrowed from computer vision, is applied to reduce the need for expensive simulations. These strategies are demonstrated to improve emulator accuracy and generalisation in a proof-of-concept study. The work contributes to making ABM emulation more reliable and accessible, particularly for African modelling contexts.

Pour aller plus loin :

Agent-based model — Core concept central to the talk.
Gaussian process — Alternative emulator approach mentioned.
Latin hypercube sampling — Baseline sampling method compared in the study.
Multilayer perceptron — The neural network architecture used.

110 words

Radar Profile

The radar profile shows high scores in quantitative information and technical level, reflecting a data-driven, methodologically detailed presentation. Quality and reliability are strong, but slightly lower due to the proof-of-concept nature and lack of external citations, resulting in a balanced profile with a slight dip in global reliability.

Reliability 7/10