What are Dynamic Games?
The Brief Explanation
Game theory reasons about two or more agents who have (potentially competing) objectives. Traditional game theory considers a one-step decision process: each agent considers how to play given their knowledge of the other agent, and then executes that play, and the game is over. Dynamic games (sometimes called differential games in continuous-time scenarios) are games that are played repeatedly (or continuously) over a time horizon. Some examples include chess, investing, missile interception, and multi-agent collision-avoidance.
General dynamic games are very challenging to solve in a computationally tractable manner. We consider a subset of these games that is a bit more tractable and very relevant to many applications: zero-sum dynamic games. Zero-sum games are adversarial: there is one objective (e.g., distance to collision), one agent is assumed to be maximizing this objective (e.g., avoiding collision), whereas the other agent is assumed to be minimizing the objective (e.g., causing collision). If you can solve for the set of conditions from which the ego-agent can avoid collision despite worst-case efforts from its adversary, then you have a certificate of robustness within that region. This can be useful in cases where the other agent is not actually adversarial. For example, an airplane should operate in a region of airspace where, even if another plane is taken over and suddenly becomes adversarial, the airplane can maneuver away safely.
The Thorough Explanation
The classic book to read would be “Dynamic Noncooperative Game Theory” by Tamer Basar and Geert Jan Olsder.
A more recent work is Smooth Game Theory from David Fridovich-Keil.
Most people think of Rufus Isaacs as the father of differential game theory, developed during his time at the RAND corporation in the 1950’s. His book, Differential Games, may be a nice historical perspective for those interested.
Why Study Zero-Sum Dynamic Games?
In addition to the explicit multi-agent games described above, analyzing zero-sum games is also useful for general robustness. Safety-critical systems operate under uncertainty: models are imperfect, environments are dynamic, and sensors have noise. If this uncertainty can be quantified, we can treat it as an adversary during analysis. This asks, for example, “from where can I safely land my quadcopter even under worst-case wind conditions?” by treating the wind as an adversary.
Our group develops methods that provide safety guarantees under these worst-case conditions. This includes differential game formulations where the controller must ensure safety against an antagonistic disturbance, distributionally robust control barrier functions that account for stochastic and distributional uncertainty directly from sensor measurements, and adaptive safety filters that refine learned value functions online as conditions change. By treating robustness as a first-class design objective rather than an afterthought, our approaches deliver safety guarantees that are meaningful in deployment.
Robustness and Dynamic Game Papers from our Group
We have some work on handling explicit adversaries, for example:
Pursuit-evasion games between a quadcopter and humanoid: MADR: MPC-guided Adversarial DeepReach (ICRA)
Pursuit-evasion game between 5 adversarial aircraft and one evader aircraft: Conservative Linear Envelopes for High-Dimensional, Hamilton-Jacobi Reachability for Nonlinear Systems via the Hopf Formula (TAC)
We also have work on treating model error as an adversary to construct robust control Lyapunov functions and robust control barrier functions. A fun implication of this type of analysis is that we can actually simplify our dynamics model to something more tractable, characterize the resulting model-mismatch error, and then solve a game treating this introduced modeling error as an adversary. This lets us plan during deployment using the simpler and faster model, with robustness guarantees for the full system model. We have two main lines of work for this:
Linearizing the system so that we can solve a very fast linear game (Hopf reachability) online for control synthesis with conservative guarantees on the nonlinear system. We can even do this in a linear latent space that is of a different dimension than the original system.
Fast and Safe Tracking (FaSTrack) solves an offline game between the full-order dynamics of a system and a reduced-order model. The result of the analysis is a tracking error bound and tracking controller for the full-order dynamics to use during deployment. As the simplified model is used for planning, the full-order system deploys the precomputing tracking controller to stay within the guaranteed robust error bound. There are lots of fun extensions of this work; see more on the FaSTrack page.
Another source of uncertainty is in state estimation, or where the system is in its environment. We have worked on:
Adding a layer of robustness to deterministic algorithms that assume perfect state information:
Robust state estimation for energy systems: Simultaneous improvement of control and estimation for battery management systems
Robust state estimation for robotic surgery: SURESTEP: An Uncertainty-Aware Trajectory Optimization Framework to Enhance Visual Tool Tracking for Robust Surgical Automation (IROS)
A related topic is being robust to uncertain or unknown environments, for example:
Uncertainty in tissue properties during robotic surgery: JIGGLE: An Active Sensing Framework for Boundary Parameters Estimation in Deformable Surgical Environments (RSS)
Robustness to sensor noise: Sensor-Based Distributionally Robust Control for Safe Robot Navigation in Dynamic Environments (IJRR)
Designing safe fallbacks for robust deployment: Steering with Contingencies: Combinatorial Stabilization and Reach-Avoid Filters