Value-Based Active Sensing for Improving Sampling Efficiency

Traditional control systems often employ periodic sensing without considering whether each acquired sample provides sufficient decision value to justify its sensing cost, leading to inefficient data collection. This paper introduces a decision-theoretic active sensing framework that formulates sensing as an observe-then-plan problem guided by a value-of-information (VoI) criterion. The core of our approach is a sensing utility function that evaluates expected opportunity loss relative to the cost of acquiring a new observation. Recognizing the limitations of classical VoI under model misspecification, we derive a generalized value function that incorporates a risk premium for variability in opportunity loss across belief states and an ignorance penalty for decaying confidence in the predictive model over time. We further present a practical online heuristic algorithm that dynamically updates risk sensitivity and model confidence using posterior feedback. We validate the approach on a real-world temperature-alert case and a CO₂-aware office deployment. Across the evaluated settings, the proposed utility-aware sensing strategy substantially reduces sensing demand while retaining control-relevant decision performance under the fixed one-step downstream decision setting.