Mind the Sim-to-Real Gap & Think Like a Scientist
The paper discusses the challenges of bridging the gap between simulated and real-world decision-making in sequential problems. It introduces a framework for when to use simulations versus real experiments, highlighting the importance of understanding value errors and reachability components. The authors propose a new experimental policy that optimizes decision-making in various contexts, illustrated through case studies in supply chain management and HIV testing.
- ▪The study examines the value error in simulations and how it can be decomposed into identifiable components.
- ▪It highlights the importance of reachability in decision-making, particularly in states not visited by the deployed policy.
- ▪The proposed Fisher-SEP policy aims to minimize predictive variance in target policy values through simulation-aided experimentation.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.21458 (cs) [Submitted on 20 May 2026] Title:Mind the Sim-to-Real Gap & Think Like a Scientist Authors:Harsh Parikh, Gabriel Levin-Konigsberg, Dominique Perrault-Joncas, Alexander Volfovsky View a PDF of the paper titled Mind the Sim-to-Real Gap & Think Like a Scientist, by Harsh Parikh and 3 other authors View PDF HTML (experimental) Abstract:Suppose a planner has a pre-trained simulator of a sequential decision problem and the option to run real experiments in the field. The simulator is cheap to query but inherits confounding and drift from its calibration data. Experimentation is unbiased but consumes one real unit per trial. We study when, and how, the planner should supplement the simulator with experiments. We give three results.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
