New arXiv Paper Introduces Physics-Informed Causal MDP Framework for Engineering Simulation
A new research article, "Physics-Informed Causal MDPs for Sequential Constraint Repair in Engineering Simulation Pipelines," has been made available on arXiv (arXiv:2604.17910v1). The paper presents the PI-CMDP framework, which focuses on constrained Markov Decision Processes (CMDPs) characterized by constraint dependencies that create a layered Directed Acyclic Graph based on a Lifecycle Ordering Assumption. It employs an Identify-Compress-Estimate approach, facilitating the recognition of causal edge weights and compressing state cardinality from 2^(WL) to (W+1)^L. A physics-informed doubly-robust estimator is implemented, ensuring unbiased results and decreased variance when the physics prior is superior to a learned model. The focus of the study is on constraint repair within engineering simulation pipelines.
Key facts
- Paper titled "Physics-Informed Causal MDPs for Sequential Constraint Repair in Engineering Simulation Pipelines" published on arXiv.
- arXiv identifier is arXiv:2604.17910v1.
- Announcement type is 'new'.
- Introduces the PI-CMDP framework for constrained Markov Decision Processes.
- Framework addresses off-policy learning in CMDPs with large binary state spaces.
- Methodology uses an Identify-Compress-Estimate pipeline.
- Lifecycle Ordering Assumption enables causal identification with partial-identification bounds.
- Markov abstraction compresses state cardinality from 2^(WL) to (W+1)^L.
Entities
Institutions
- arXiv