Deception in Learning of Leader-Follower Games

• 成果类型: Article
• 署名作者: John, Varkey M.; Vamvoudakis, Kyriakos G.
• 署名单位: University System of Georgia; Georgia Institute of Technology; University System of Georgia; Georgia Institute of Technology
• 刊物名称: IEEE TRANSACTIONS ON AUTOMATIC CONTROL
• ISSN/ISSBN: 0018-9286
• DOI: 10.1109/TAC.2026.3678503
• 发表日期: 2026
• 关键词: ADAPTIVE OPTIMAL-CONTROL; TIME LINEAR-SYSTEMS; tracking control; Stackelberg; strategy

摘要

In this article, we explore deception mechanisms in the context of game-theoretic learning. We first examine a scenario where certain agents, referred to as intelligent players (IPs), given that they have increased intelligence, attempt to discern the scores of the learning mechanisms used by agents of lower intelligence, referred to as agents. To counteract this, agents alter the values of the score function to mislead IPs, thus obstructing the ability to learn the learning parameters of agents. Next, we consider a scenario where the IPs, having acquired the learning parameters of the agents, try to adjust the values of the agent score function to the preferred outcomes. In response, the agents modify these values to negate the influence of IPs' actions. Both scenarios are analyzed as Stackelberg games: the first under an open-loop information structure (access to only the initial state value) and the second under a closed-loop information structure (access to the entire state trajectory). These games are studied in a continuous-time setting with tracking objectives. Finally, we introduce a novel, efficient, and model-free off-policy algorithm to compute the closed-loop Stackelberg game. The effectiveness of our results is demonstrated through two examples, including an exponentially discounted system used in the literature and a honeypot system.