COBRA: Contextual Bandit Algorithm for Ensuring Truthful Strategic Agents

This paper considers a contextual bandit problem involving multiple agents, where a learner sequentially observes the contexts and the agents' reported arms, and then selects the arm that maximizes the system's overall reward. Existing work in contextual bandits assumes that agents always truthfully report their arms, which is unrealistic in many real-life applications. For instance, consider an online platform with multiple sellers; some sellers may misrepresent product features to maximize reward, such as having the platform preferentially recommend their products to its users. To address this challenge, we propose an algorithm, COBRA, for contextual bandit settings involving strategic agents that disincentivize their strategic behavior without using any monetary incentives, while having an approximate Bayesian incentive compatibility and a sub-linear regret guarantee. Experimental results also validate our theoretical results and the different performance aspects of COBRA.