We work on the estimation, decision, and specification problems that arise when a learning-enabled system has to be shown safe before it is deployed.

A deployed learning-enabled system fails at rates too low to observe directly, acts on information it only partially has, and must eventually be justified to someone who was not involved in building it. Those three problems — estimating rare failures, deciding under partial observability, and producing a specification that can be checked rather than asserted — organize the work below.

Research Themes

Safety Validation and Rare-Event Estimation

Rare-Event Simulation & Importance Sampling

Adaptive importance sampling with normalizing-flow and diffusion proposal distributions, for estimating failure probabilities of black-box systems at rates where naive Monte Carlo is intractable.

Failure Synthesis & Falsification

Generating failure trajectories for learning-enabled autonomous systems, with estimators that carry finite-sample guarantees rather than empirical pass rates.

Robust Perception under Distribution Shift

Evaluating and improving perception for autonomous driving when deployment conditions depart from the training distribution, including capability loss under policy compression.

Sequential Decision Making under Partial Observability

Markov Decision Processes & POMDPs

Formulating sequential decision problems under partial observability, from geological uncertainty in critical mineral supply chains to socially-equitable renewable energy allocation.

Online Planning & Reinforcement Learning

Monte Carlo tree search, online planning, and reinforcement learning for high-dimensional decision problems, with provable performance or regret bounds.

Stochastic Programming & Resource Allocation

Shrinking-horizon and stochastic optimization for allocating constrained resources under uncertain demand across industrial and infrastructure systems.

Systems Assurance and Certification

Machine-Readable V&V Specification

Expressing verification and validation requirements in a scriptable, machine-checkable form, so an assurance argument can be re-executed against a changed system rather than re-asserted.

Requirements Traceability & Runtime Monitoring

Linking high-level system requirements to low-level tests and runtime monitors, following the traceability obligations imposed by standards such as DO-178C and ISO 26262.

Interpretable Models for Regulated Decisions

Interpretable and explainable modeling where a decision must be defensible to a clinician, operator, or regulator, including automatic clinical grading systems.

Application Programs

Critical Minerals & Energy Transition

Exploration, processing, and supply decisions for the minerals and energy systems underpinning the transition, including lithium self-sufficiency pathways and phosphogypsum recovery.

Autonomous Driving & Robotics

Safety validation and decision-making for driving policies, navigation, and human-robot collaboration in field and industrial settings.

Clinical Decision Support

Decision and documentation support in clinical workflows, from stroke diagnosis and ventilator control to orthopaedic grading and primary-care records.

Industrial Supply Chains & Logistics

Multi-echelon supply chain modeling, production planning, and transit systems, including bullwhip benchmarking and co-production under coupling constraints.

Key Contributions

Deep Importance Sampling. A technique that finds potential failures in autonomous systems 1000x faster than conventional methods, enabling safety validation in weeks instead of decades.

Adaptive Meta-Learning for Safety. A framework that lets machine learning systems recognize situations beyond their training, creating a sixth sense for uncertainty. Recognized with a CPS Rising Star 2024 award.

Certified Bounds for Neural Supply Chains. The first application of formal neural network verification to multi-echelon inventory decisions, cutting certified worst-case cost exposure by up to 76% versus tier-by-tier verification.

Mineral-X Supply Chain Models. AI systems developed with Stanford Mineral-X that optimize for resilience and equity in critical mineral supply chains, ensuring the green revolution serves all communities.