Georgia Institute of Technology
Optimality-preserving Integration of Follower Intent Estimation and Control in Stackelberg Dynamic Games
Abstract
dc:description.abstractThis thesis studies two-player Stackelberg dynamic games in which the leader has uncertainty about the follower’s intent. In many interactive multi-agent scenarios, such as autonomous vehicle encounters, and human-robot interaction, the leader must act strategically while lacking complete knowledge of how the follower will respond. An implicit assumption in such settings is that if the leader can estimate the follower’s intent, then solving the corresponding Stackelberg game will produce the best possible interaction outcome for the leader. This thesis examines that assumption directly. To do so, we formulate a Stackelberg dynamic game in which the leader updates its strategy within the finite horizon. This strategy-update formulation captures the practical setting in which the leader has an initial belief about the follower, then it observes its behavior, and then re-optimizes based on an updated belief. The central question is whether, under this update structure, knowing the follower’s true best response (BR) still yields the lowest leader cost. The answer is, in general, no. Under open loop (OL) information, this thesis shows that the OL Stackelberg equilibrium is not generally optimal once a strategy update is introduced. This loss of optimality is tied directly to the time-inconsistency of the OL Stackelberg equilibrium: although the equilibrium is optimal among strategies announced and committed to at the initial time, it need not remain optimal when the leader is allowed to re-optimize later in the horizon. Thus, under a strategy update, the OL Stackelberg equilibrium does not preserve global optimality. To address this issue, the thesis introduces a novel commitment-breaking Stackelberg strategy. In this approach, the leader explicitly accounts for the fact that it will update its strategy later in the horizon. Rather than naively committing to the initial OL Stackelberg equilibrium, the leader optimizes over both the strategy it announces initially and the strategy it intends to play after the update. The resulting solution preserves global optimality under the assumed update structure. This reveals a key feature of Stackelberg games with updates: to act optimally, the leader may need to commit initially to a strategy that it does not plan to carry through over the entire horizon. The thesis also analyzes the feedback (FB) Stackelberg equilibrium under strategy update. In contrast to the OL case, the FB Stackelberg equilibrium is time-consistent. However, time-consistency alone is not sufficient to guarantee full optimality. Because the FB Stackelberg equilibrium is not, in general, Markov perfect, there may still exist alternative FB strategies that achieve lower cost. Thus, although the FB Stackelberg equilibrium is robust to re-optimization in a way that the OL equilibrium is not, it still does not guarantee the lowest cost over the full class of admissible FB strategies. These theoretical results are then connected to the setting of belief updates, where the leader’s uncertainty about the follower is represented explicitly through changing BR beliefs. This establishes that the lack of optimality under strategy updates translates directly into a lack of guaranteed optimality under intent estimation updates. In particular, it is possible for an incorrect follower belief to obtain a lower cost than the true BR belief. This finding challenges a common implicit assumption in inverse learning and intent estimation for interactive control. Finally, the thesis validates the aforementioned theoretical contributions through simulations of a linear quadratic (LQ) collision-avoidance game with Bayesian BR estimation. The numerical results show that the predicted phenomena occur frequently in practice: under both OL and FB information structures, the true BR belief does not always achieve the lowest cost. In addition, the simulations show that the commitment-breaking Stackelberg strategy consistently outperforms OLSE baselines both with and without update, and that its relative advantage is most pronounced when the update occurs earlier in the horizon. Overall, this thesis identifies a fundamental limitation of separating estimation and control in Stackelberg dynamic games. It shows that knowing the follower’s intent is not, by itself, sufficient to guarantee optimal behavior once strategic updates are allowed within the horizon. By introducing the commitment-breaking Stackelberg strategy, the thesis provides a new framework for integrating prediction and control while preserving optimality, and suggests broader connections to security games, deception games, and contingency planning.
Degree
thesis:*- Name thesis:degree_name
- Aerospace Engineering, MS
- Department thesis:degree_department
- Aerospace Engineering
- Grantor
- Georgia Institute of Technology
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Salinas-Rodriguez, Cayetana
- Committee members dc:contributor.committeemember
-
- Dr. Samuel Coogan
- Dr. Chih-Yuan Chiu
- Dr. Jonathan Rogers
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/81652