Learning-Based Robust Bayesian Persuasion with Conformal Prediction Guarantees
Classical Bayesian persuasion assumes that senders fully understand how receivers form beliefs and make decisions--an assumption that rarely holds when receivers possess private information or exhibit non-Bayesian behavior. In this paper, we develop a learning-based framework that integrates neural networks with conformal prediction to achieve robust persuasion under uncertainty about receiver belief formation. The proposed neural architecture learns end-to-end mappings from receiver observations and sender signals to action predictions, eliminating the need to identify belief mechanisms explicitly. Conformal prediction constructs finite-sample valid prediction sets with provable marginal coverage, enabling principled, distribution-free robust optimization. We establish exact coverage guarantees for the data-generating policy and derive bounds on coverage degradation under policy shifts. Furthermore, we provide neural network approximation and estimation error bounds, with sample complexity $O(d \log(|\mathcal{U}||\mathcal{Y}||\mathcal{S}|)/\varepsilon^2)$, where $d$ denotes the effective network dimension, and finite-sample lower bounds on the sender's expected utility. Numerical experiments on smart-grid energy management illustrate the framework's robustness.