Bayesian Inference of Contextual Bandit Policies via Empirical Likelihood

Researchers have developed a new Bayesian inference method using empirical likelihood to analyze multiple contextual bandit policies with unprecedented accuracy in finite sample regimes. The technique is specifically robust to small sample sizes and provides full uncertainty quantification for policy value evaluation and comparison. Demonstrated through Monte Carlo simulations and applied to an adolescent BMI dataset, it addresses a critical limitation in reinforcement learning where data is scarce but decisions must be confident.