Reinforcing RCTs with Multiple Priors while Learning about External Validity

원문 초록

This paper introduces a framework for incorporating prior information into the design of sequential experiments. These sources may include past experiments, expert opinions, or the experimenter's intuition. We model the problem using a multi-prior Bayesian approach, mapping each source to a Bayesian model and aggregating them based on posterior probabilities. Policies are evaluated on three criteria: learning the parameters of payoff distributions, the probability of choosing the wrong treatment, and average rewards. Our framework demonstrates several desirable properties, including robustness to sources lacking external validity, while maintaining strong finite sample performance.