Papers
Topics
Authors
Recent
Search
2000 character limit reached

Active Learning for Gaussian Process Regression Under Self-Induced Boltzmann Weights

Published 11 May 2026 in cs.LG and cs.AI | (2605.10654v1)

Abstract: We consider the active learning problem where the goal is to learn an unknown function with low prediction error under an unknown Boltzmann distribution induced by the function itself. This self-induced weighting arises naturally in problems such as potential energy surface (PES) modeling in computational chemistry, yet poses unique challenges as the target distribution is unknown and its partition function is intractable. We propose \texttt{AB-SID-iVAR}, a Gaussian Process-based acquisition function that approximates the intractable Bayesian target distribution in closed form while avoiding partition function estimation, and is applicable to both discrete and continuous input domains. We also analyze a Thompson sampling alternative (\texttt{TS-SID-iVAR}) as a higher variance Monte Carlo variant. Despite the unknown target, under mild conditions, we establish that the terminal prediction error vanishes with high probability, and provide a tighter average-case guarantee. We demonstrate consistent improvements over existing approaches in this setting on synthetic benchmarks and real-world PES modeling and drug discovery tasks.

Summary

  • The paper introduces novel acquisition functions, AB-SID-iVAR and TS-SID-iVAR, that address the circular dependency between the unknown function and its induced target distribution.
  • The paper provides theoretical convergence guarantees with sublinear rates and demonstrates empirical superiority in synthetic benchmarks and real-world applications like PES modeling and drug discovery.
  • The paper highlights the trade-offs between computational cost and sample efficiency, using analytic approximations and MCMC integration for robust uncertainty quantification.

Active Learning for Gaussian Process Regression Under Self-Induced Boltzmann Weights: Expert Analysis

Problem Formulation and Motivation

The paper addresses a novel variant of active learning where the target distribution over the input domain is not independent but is itself determined by the unknown function being learned—specifically via a Boltzmann weight. This self-induced distribution is highly relevant in scientific settings such as potential energy surface (PES) modeling, where task-relevance is concentrated in regions of configuration space deemed physically accessible, yet neither the distribution nor its partition function is tractable. Formally, the target is to minimize the weighted mean squared prediction error under this self-induced Boltzmann distribution, a scenario not covered by classical active learning or distributionally robust active learning frameworks.

Methodological Contributions

Acquisition Functions for SIDAL

The main innovation is the development of acquisition functions robust to the circular dependency between unknown function and induced target distribution. Two algorithms are introduced:

  • AB-SID-iVAR (Approximate Bayesian Self-Induced Distribution integrated Variance Reduction): This algorithm leverages analytic approximations (zero-order Taylor expansion) to compute the expected Boltzmann-weighted variance in closed form, bypassing the need for partition function estimation, and admits tractable optimization in both discrete and continuous input spaces.
  • TS-SID-iVAR (Thompson Sampling SID-iVAR): This variant employs Monte Carlo sampling from the GP posterior, yielding an unbiased but higher-variance estimate of the objective, and embodies a bias-variance tradeoff.

Both methods directly target variance reduction under the surrogate of the Bayesian self-induced distribution, conditioned on the GP posterior, and utilize MCMC sampling for tractable integration in continuous domains.

Convergence Analysis

Theoretical guarantees are provided, establishing sublinear convergence rates for the terminal prediction error under both high-probability and average-case settings. The high-probability bounds explicitly account for the additional Monte Carlo error incurred in continuous domains, and demonstrate that the exponentially weighted prefactor, although unavoidable due to the density ratio argument, does not affect the polynomial rate of decay in prediction error. Notably, the average-case bounds are tighter, with improved logarithmic scaling in the prefactor exponents.

Analysis of Existing Heuristics

The paper formalizes and analyzes popular heuristic approaches from the molecular modeling literature within the SIDAL framework, showing—under mild conditions—that these methods also admit convergence guarantees, thereby providing an unifying theoretical foundation for SID-aware empirical strategies.

Empirical Evaluation and Results

Synthetic Benchmarks

Across standard synthetic functions (d = 1–6), AB-SID-iVAR consistently achieves the lowest weighted MSE relative to baseline methods (Random Sampling, Uncertainty Sampling, Integrated MSE, EPIG). The performance gap widens in higher dimensions and with more concentrated target distributions, reflecting increased inefficiency of approaches that do not exploit the structure of the self-induced Boltzmann weighting. TS-SID-iVAR typically matches AB-SID-iVAR but exhibits greater variance, especially for multimodal targets.

Real-World Applications

  • Potential Energy Surface Modeling: For PES fitting tasks of varying dimensionality, AB-SID-iVAR outperforms both SID-unaware and heuristic SID-aware baselines, demonstrating improved accuracy in regions that dominate Boltzmann statistics. The methodology aligns naturally with the requirement for accurate predictions in physically relevant regions and achieves substantial reduction in weighted MSE.
  • Molecular Drug Discovery: On a large molecular pool, with GP surrogates using Tanimoto kernels, AB-SID-iVAR realizes one to two orders of magnitude lower weighted MSE compared to baselines and maintains positive R2R^2 scores on elite compound subsets, even as baselines degrade to negative values as the Boltzmann density sharpens. This indicates the method's robustness to generalization in high-value regions critical for downstream screening.

Ablation and Robustness

Ablation studies highlight the importance of the Bayesian surrogate and variance-threshold constraint set: naive plug-in surrogates or unconstrained optimization are insufficient in scenarios with multimodal or oscillatory targets, as they fail to capture uncertainty and thus prematurely commit to suboptimal regions.

Computational Considerations

The acquisition optimization and MCMC sampling introduce higher per-iteration costs relative to classical methods, but remain practical for high-value, low-throughput scientific applications where sample efficiency dominates computational cost. Empirical studies indicate modest sensitivity to MCMC sample size and confirm robustness across varying target concentration and bias parameterizations.

Implications and Theoretical Impact

The formalization of active learning under self-induced distributions closes a gap in the theoretical understanding of data acquisition for tasks where accuracy matters only in regions of high probability determined by an unknown function. This is particularly important for applications such as rare-event simulation, quantum chemistry, and drug discovery, where the Boltzmann weighting is physically motivated.

The acquisition functions and convergence analysis provide a rigorous foundation for SID-aware algorithms, advancing both practical and theoretical active learning. The framework offers principled guidance on when and how uncertainty quantification should influence query selection in the presence of function-dependent targets.

Future extensions may focus on relaxing the global density ratio argument, developing input-dependent coverage metrics, bridging to regret bounds in Bayesian optimization, and extending theory to non-Boltzmann or more complex SID formulations. Efficient amortized Boltzmann samplers (e.g., diffusion-based) hold promise for further reducing computational overhead.

Conclusion

This paper introduces a principled framework for active learning under self-induced Boltzmann distributions, motivated by scientific applications where conventional uniform-error objectives are inadequate. The AB-SID-iVAR and TS-SID-iVAR algorithms provide tractable, theoretically justified approaches with superior empirical performance. The work provides both practical tools and a foundation for further exploration of targeted, structure-aware learning protocols in applications where acquisition cost and data relevance are tightly coupled.

Paper to Video (Beta)

No one has generated a video about this paper yet.

Whiteboard

No one has generated a whiteboard explanation for this paper yet.

Open Problems

We haven't generated a list of open problems mentioned in this paper yet.

Collections

Sign up for free to add this paper to one or more collections.

Tweets

Sign up for free to view the 1 tweet with 2 likes about this paper.