Back to Home

Adaptive Grid-Based Thompson Sampling for Efficient Trajectory Discovery

Arindam Fadikar, Abby Stevens, Mickael Binois, Nicholson Collier, and Jonathan Ozik

arXiv:2510.18099

arXiv preprint, October 2025

PDF arXiv
Adaptive trajectory search in stochastic simulations
Figure: Illustration of adaptive trajectory search using Thompson sampling over an evolving candidate grid.

Abstract

This paper studies trajectory-oriented Bayesian optimization for calibrating and exploring stochastic simulation models where a full trajectory, not just a single scalar target, drives inference quality.

The authors propose an adaptive grid-based Thompson sampling strategy that balances global exploration and local refinement while reducing unnecessary evaluations in low-value regions. The method is designed to improve sample efficiency compared with fixed-grid and less adaptive search approaches.

Across benchmark and applied settings, the approach is shown to discover higher-quality trajectories with fewer expensive simulations, supporting faster and more practical uncertainty-aware model calibration workflows.

Key Contributions

Adaptive Grid Construction

Introduces a dynamically refined candidate set that allocates computation where the posterior suggests higher potential for trajectory improvement.

Thompson Sampling Integration

Extends Thompson sampling to trajectory discovery, enabling principled exploration-exploitation tradeoffs in high-cost stochastic simulation studies.

Evaluation Efficiency

Demonstrates reduced simulation budgets for comparable or better trajectory quality, which is critical when each model run is computationally expensive.

Practical Calibration Utility

Provides a practical optimization strategy for uncertainty-aware trajectory matching in scientific and epidemiological model calibration pipelines.

Citation

BibTeX
@article{fadikar2025adaptivets,
  title={Adaptive Grid-Based Thompson Sampling for Efficient Trajectory Discovery},
  author={Fadikar, Arindam and Stevens, Abby and Binois, Mickael and Collier, Nicholson and Ozik, Jonathan},
  journal={arXiv preprint arXiv:2510.18099},
  year={2025},
  month={oct},
  doi={10.48550/arXiv.2510.18099}
}

Related Publications