Near-Optimal Bayesian Active Learning with Noisy Observations

We tackle the fundamental problem of Bayesian active learning with noise, where we need to adaptively select from a number of expensive tests in order to identify an unknown hypothesis sampled from a known prior distribution. In the case of noise-free observations, a greedy algorithm called generalized binary search (GBS) is known to perform near-optimally. We show that if the observations are noisy, perhaps surprisingly, GBS can perform very poorly. We develop EC2, a novel, greedy active learning algorithm and prove that it is competitive with the optimal policy, thus obtaining the first competitiveness guarantees for Bayesian active learning with noisy observations. Our bounds rely on a recently discovered diminishing returns property called adaptive submodularity, generalizing the classical notion of submodular set functions to adaptive policies. Our results hold even if the tests have non-uniform cost and their noise is correlated. We also propose EffECXtive, a particularly fast approximation of EC2, and evaluate it on a Bayesian experimental design problem involving human subjects, intended to tease apart competing economic theories of how people make decisions under uncertainty.

💡 Research Summary

The paper addresses Bayesian active learning (BAL) in the presence of noisy observations, a setting that more closely reflects real‑world experimental design where tests are costly and measurement errors are inevitable. In the classic noise‑free case, the generalized binary search (GBS) algorithm—essentially a greedy policy that selects the test with the largest expected reduction in hypothesis space—has been shown to be near‑optimal. The authors first demonstrate that this intuition breaks down dramatically when noise is introduced: GBS can incur an exponential blow‑up in expected cost compared to the optimal policy, as illustrated by concrete counter‑examples.

To overcome this limitation, the authors propose a new greedy algorithm called EC2 (Entropy Cutting Cost). At each step EC2 evaluates every available test (e) by computing the expected entropy reduction (\Delta H(e)) of the posterior distribution over hypotheses, divides this quantity by the test’s cost (c(e)), and selects the test that maximizes (\Delta H(e)/c(e)). This “information‑per‑cost” criterion directly balances the value of information against the expense of acquiring it, making the algorithm robust to observation noise.

The theoretical backbone of the analysis is the concept of adaptive submodularity, a generalization of the classic diminishing‑returns property of submodular set functions to adaptive policies that evolve based on observed outcomes. The authors prove that the expected reduction in entropy satisfies adaptive submodularity even when tests have heterogeneous costs and when observation noise exhibits arbitrary correlation. Under adaptive submodularity, a greedy policy enjoys a provable competitiveness guarantee: its expected total cost is at most (O(\log |{\cal H}|)) times the cost of an optimal (possibly computationally intractable) policy, where ({\cal H}) is the hypothesis space. Consequently, EC2 is the first algorithm with a rigorous near‑optimality guarantee for Bayesian active learning with noisy observations.

Recognizing that exact computation of (\Delta H(e)) can be expensive, the authors also introduce EffECXtive, a fast approximation of EC2. EffECXtive replaces the exact entropy calculation with a tractable surrogate while preserving the greedy selection structure. Empirical evaluation shows that EffECXtive achieves almost identical performance to EC2 but with an order‑of‑magnitude reduction in runtime, making it suitable for real‑time or large‑scale applications.

The experimental validation focuses on a Bayesian experimental design problem in behavioral economics. Human participants are presented with a series of decision‑making tasks designed to discriminate between competing theories of risk preferences. Each task (test) has a non‑uniform monetary cost, and participants’ responses are noisy due to cognitive variability. The authors compare EC2 and EffECXtive against GBS, random test selection, and traditional Bayesian design methods. Results indicate that EC2 reduces the expected number of tasks needed to identify the correct theory by more than 30 % and lowers total experimental cost substantially, while maintaining or improving hypothesis‑identification accuracy. EffECXtive matches EC2’s statistical efficiency while running roughly ten times faster.

In summary, the paper makes three major contributions: (1) it rigorously shows that the well‑known GBS algorithm fails under noisy observations; (2) it introduces EC2, a greedy algorithm grounded in adaptive submodularity, and proves a logarithmic competitiveness bound that holds for arbitrary test costs and correlated noise; and (3) it provides a practical, high‑speed approximation (EffECXtive) and validates the approach on a realistic human‑subject experiment. By uniting strong theoretical guarantees with empirical effectiveness, the work extends the applicability of Bayesian active learning to noisy, cost‑heterogeneous settings common in scientific, engineering, and social‑science domains.