Research map / Probabilistic and scientific modeling / Molecules and proteins
Interventionally Consistent Surrogates for Complex Simulation Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Frames complex simulators as SCMs and learns interventionally consistent surrogates (latent ODE/RNN families) by minimising a KL-based abstraction error over interventions, using causal abstraction.
Problem
Surrogate models of expensive socio-technical simulators are typically trained on observational data and can misjudge the effects of policy interventions, misguiding decision-makers.
Approach
Treats the simulator and surrogate as SCMs and learns the surrogate and an intervention map by minimising an expected KL-based abstraction error over an interventional distribution (exact/approximate causal transformations), with a theoretical Markov bound on per-intervention error.
Key results
- Interventionally trained surrogates achieve far lower AMSE/ANLL on interventional test sets than observationally trained ones across LRNN, LODE-RNN and LODE families.
- Observationally trained LRNN predicted no lockdown was the best intervention in 1 of 5 repeats and failed to predict it was the worst in all 5 repeats; interventionally trained surrogates did not predict no lockdown as best.
- Surrogates simulate approximately three times faster than the original ABM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train the surrogate on interventional data rather than observational data | SIRS ABM surrogates · SIRS ABM interventional test I′ | AMSE: 18.521.9 17.1 -> 3.353.41 3.18 | Helped |
| Train the surrogate on interventional data rather than observational data | SIRS ABM surrogates · SIRS ABM observational test O′ | AMSE: 2.522.78 2.16 -> 3.593.68 3.54 | Mixed |
| Use a hybrid LODE-RNN surrogate instead of the LODE or LRNN surrogate | SIRS ABM surrogates trained interventionally · SIRS ABM interventional test I′ | AMSE | Helped |
| Train the surrogate on interventional data rather than observational data | Predator-prey simulator surrogates · Predator-prey interventional test I′ | AMSE: 3.304.32 2.37 -> 1.651.80 1.62 | Helped |
| Train the surrogate on interventional data rather than observational data | Predator-prey simulator surrogates · Predator-prey observational test O′ | AMSE: 1.611.64 1.59 -> 1.852.03 1.81 | Mixed |
Surprising findings
- Observationally trained surrogates often mispredict the ranking of interventions, sometimes ranking no lockdown as the best option, and the lockdown effect can be reversed in predictions.
Limitations
- Expectation over interventions rather than maximum means some interventions may be captured poorly despite low average error.
- Requires surrogate families with tractable differentiable densities for the KL divergence.
- No sample complexity analysis of abstraction learning.
Experimental setup
Datasets: simulated trajectories from spatial SIRS ABM under interventional and observational distributions predator-prey ABM simulations (additional case study)
Benchmarks: interventional hold-out test set I' observational hold-out test set O'
Metrics: average mean squared error (AMSE) average negative log-likelihood (ANLL)
Baselines: observationally trained LRNN observationally trained LODE-RNN observationally trained LODE
Builds on: causal abstraction (Rubenstein et al. 2017) approximate causal abstraction (Beckers et al. 2020) latent ODE (Rubanova et al. 2019) universal differential equations (Rackauckas et al. 2021)
Closest papers by meaning
- G-Sim: Generative Simulations with Large Language Models and Gradient-Free Calibration (ICML 2025 · similarity 0.85)
- Model-Free Assessment of Simulator Fidelity via Quantile Curves (NeurIPS 2026 · similarity 0.84)
- SMRS: advocating a unified reporting standard for surrogate models in the artificial intelligence era. (NeurIPS 2025 · similarity 0.84)
- Decomposing Interventional Causality into Synergistic, Redundant, and Unique Components (NeurIPS 2025 · similarity 0.83)
- Intervention and Conditioning in Causal Bayesian Networks (NeurIPS 2024 · similarity 0.83)
- Empirical Analysis of Model Selection for Heterogeneous Causal Effect Estimation (ICLR 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.