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

Researchers from the machine learning community have introduced a novel approach to active learning for Gaussian Process Regression, specifically addressing scenarios involving self-induced Boltzmann weights. This problem is prevalent in computational chemistry, such as potential energy surface modeling, where the target distribution is unknown and its partition function is intractable. The team proposes AB-SID-iVAR, a new acquisition function that approximates the Bayesian target distribution in closed form without requiring partition function estimation. Additionally, they analyze TS-SID-iVAR, a Thompson sampling alternative. Theoretical analysis demonstrates that terminal prediction error vanishes with high probability under mild conditions, offering tighter average-case guarantees. Empirical results on synthetic benchmarks and real-world tasks, including drug discovery, show consistent improvements over existing methods. This work significantly advances efficient function learning in complex scientific domains by overcoming previous computational barriers associated with unknown target distributions.