M2W— Model-to-World LabBlanchet Research Group · Stanford MS&E
Menu

The M2W research program

Turn model discrepancy into computable decision evidence.

We study the gap between the mathematical, simulated, or data-derived worlds used to make decisions and the environments in which those decisions must perform.

One connected program

Model. Simulate. Decide. Transfer.

Each stage creates an object that can be examined, challenged, and improved by the next.

01

Model

Construct useful worlds without mistaking them for reality.

We represent stochastic systems, constraints, dependence, and causal structure at the resolution required by a decision.

  • Stochastic-process models
  • Data-calibrated environments
  • Structural and causal representations
02

Simulate

Make plausible, rare, and unobserved worlds computationally accessible.

We develop Monte Carlo and generative methods for events ordinary sampling misses and counterfactual worlds data cannot reveal directly.

  • Rare-event simulation
  • Exact and unbiased methods
  • Counterfactual and generative worlds
03

Decide

Optimize performance without assuming the nominal world is correct.

We compare and learn policies under uncertainty, distribution shift, partial identification, and operational constraints.

  • Distributionally robust optimization
  • Policy learning and evaluation
  • Stochastic control
04

Transfer

Measure what can fail before relying on a decision in practice.

We quantify sensitivity, calibrate uncertainty, test out-of-environment performance, and use new evidence to update the world model.

  • Model-risk measurement
  • Out-of-sample guarantees
  • Validation and recalibration
Cross-cutting questionHow large is the model-to-world gap, and what does it do to the decision?

Audit the gap

One phrase. Six measurable sources of risk.

M2W does not treat “model error” as a single vague quantity. Different gaps require different diagnostics, guarantees, and evidence.

01

Numerical

Approximation, discretization, and Monte Carlo error inside the computational model.

02

Input

Uncertainty created when a simulator is calibrated from finite data.

03

Structural

Missed mechanisms, dependence, or dynamics in the nominal model.

04

Distribution shift

Changes between the population used to learn and the population where a decision operates.

05

Rare events

Consequences that lie beyond the reach of ordinary data or ordinary simulation.

06

Counterfactual

Outcomes that cannot be jointly observed and may only be partially identified.

Selected research foundations

A twenty-year path toward M2W.

Representative work; the complete, automatically maintained scholarly record remains available separately.

Simulate

Efficient rare-event simulation for the maximum of heavy-tailed random walks

Transfer

Quantifying Distributional Model Risk via Optimal Transport

Transfer

Sample Out-of-Sample Inference Based on Wasserstein Distance

Decide

Distributionally Robust Batch Contextual Bandits

Model + Simulate

Consistency of Neural Causal Partial Identification

Transfer

Stability Evaluation through Distributional Perturbation Analysis

Browse all publications

From theorem to prototype

The next unit of research is an end-to-end decision system.

Decision Labs connect a scientific model to an executable decision, an explicit evidence boundary, and a plan for learning from the world. They are not claims of deployment; they make the path to responsible transfer inspectable.

Explore Decision Labs