Causal Bandits Improve Decision-Making with Non-Manipulable Variables

The study addresses a critical challenge in causal bandits: situations where some influential variables cannot be directly controlled or manipulated, yet still provide valuable information for decision-making. Traditional causal bandit frameworks often struggle with these non-manipulable variables, limiting their effectiveness in complex real-world scenarios. Researchers developed novel algorithms, Causal Thompson Sampling and Causal Information-Directed Sampling (IDS), specifically designed for contextual causal bandits with such variables. These methods operate under the assumption of a known causal graph without latent confounding, using a Bayesian formulation where conditional probability tables represent unknown parameters. This allows observations from one intervention to inform reward estimates for others through their shared causal mechanisms. The proposed methods demonstrated superior performance over both causal and non-causal baselines in synthetic causal bandit tasks. This improvement stems from their ability to more effectively exploit shared information across different interventions, leading to faster and more accurate identification of high-reward decisions. The research also provides theoretical regret bounds for both algorithms, quantifying their efficiency.