Past seminars

TBasar

Prof. Tamer Başar
University of Illinois at Urbana-Champaign
Video | Slides

Multi-Agent Dynamical Systems with Asymmetric Information and with Elements of Learning

Decision making in dynamic uncertain environments with multiple agents arises in many disciplines and application domains, including control, communications, distributed optimization, social networks, and economics. Here a natural framework, and a comprehensive one, for modeling, optimization, and analysis is the one provided by stochastic dynamic games (SDGs), which accommodates different solution concepts depending on how the interactions among the agents are modeled, particularly whether they are in a cooperative mode (with the same objective functions, as in teams) or in a noncooperative mode (with different objective functions) or a mix of the two, such as teams of agents interacting noncooperatively across different teams (and of course cooperatively within each team). What also affects (strategic) interactions among the agents is the asymmetric nature of the information different agents acquire (and do not share or only partially share (selectively) with others, even within teams). What makes such problems even more challenging in a dynamic environment with networked agents is the dependence of the information available to one agent at some point in time on the policies or decisions of other agents who have already acted at earlier instants of time. Such decision problems, initially studied in a team framework, are known as those with nonclassical information where optimal policies of team agents must be designed to balance a tradeoff between contribution to optimality of the team objective function and signaling through their actions useful information to other agents in their neighborhood who would be acting after them. Existence of such a tradeoff between signaling and optimization creates even more challenging issues in SDGs with misaligned objectives among at least a subset of agents, which however can be addressed effectively for a specially structured subclass of such games, namely mean-field games.

This talk will provide an overview of the landscape above, first for a general class of stochastic dynamic teams and games, and then for a subclass where the objective functions are quadratic, and the interaction relationships are linear. The talk will also cover reinforcement learning embedded into policy development when agents do not have precise information on the underlying models.


m_kamgapour

Prof. Maryam Kamgarpour
EPFL

Learning equilibria in games with bandit feedback

A central challenge in large-scale engineering systems, such as energy and transportation networks, is enabling autonomous decision-making among interacting agents. Game theory provides a natural framework to model and analyze such problems. In practice, however, agents often have only partial information about the costs and actions of others. This makes decentralized learning a key tool for developing effective strategies. In this talk, I will discuss recent advances in decentralized learning for static and Markov games under bandit feedback. I will outline algorithms with convergence guarantees and highlight directions for future research.


g_fieni

Giona Fieni
ETH Zürich
Video

Game Theory in Formula One

df_keil

Prof. David Fridovich-Keil
University of Texas at Austin
Video | Slides

Variations on a Theme: Information Structure, Equilibria, and Dynamic Games

This talk reviews a fundamental building block of dynamic game theory—the linear-quadratic game—and discuss how Nash equilibrium solutions differ as a consequence of the information players have access to at different times. In this context, we examine several recent results, aligned to the following questions:

  1. How can we find feedback strategies which closely approximate Nash solutions, but minimize inter-agent communication/sensing?
  2. If agents’ access to information changes during an interaction, are there scenarios in which we can still find equilibria efficiently?
  3. In two-player, zero-sum games, there are classical results about the equivalence of solutions under different information structures for linear-quadratic games. In what sense do these extend beyond the linear-quadratic setting?

b_nortmann

Dr. Benita Nortmann
EMPA - Swiss federal laboratories for material science and technology
Video | Slides

Exploring feedback Nash equilibria in infinite-horizon LQ dynamic games

Many modern processes involve real-time decisions in dynamic environments influenced by multiple decision makers with team-based and individual, potentially competitive objectives. Dynamic game theory captures interactions between such players and allows us to model and design decision strategies. We focus on the class of infinite-horizon, nonzero-sum, linear quadratic (LQ), discrete-time dynamic games, which are relevant in various engineering and economic applications. Linear feedback Nash equilibrium (FNE) solutions to such LQ games are characterised via the solutions of coupled algebraic matrix equations. While reminiscent of the algebraic Riccati equation arising in LQ optimal control, these coupled equations are generally difficult to solve and may admit multiple solutions with different outcomes.

To effectively employ dynamic games to model and design multi-player decisions, it is important to understand the existence, number, and properties of equilibria and to develop efficient computation methods. In this talk, we first build intuition for the number and properties of FNE solutions by considering games with scalar dynamics and inputs, using geometric arguments. We then discuss iterative methods to compute FNE solutions for general LQ games, and a data-driven approach, which enables players to jointly converge to a FNE solution without knowledge of each other’s control objectives.


n_lanzetti

Dr. Nicolas Lanzetti
Caltech
Video | Slides

Strategically Robust Game Theory via Optimal Transport

In many game-theoretic settings, agents are challenged with taking decisions against the uncertain behavior exhibited by others. Often, this uncertainty arises from multiple sources, e.g., incomplete information, limited computation, bounded rationality. While it may be possible to guide the agents' decisions by modeling each source, their joint presence makes this task particularly daunting. Toward this goal, it is natural for agents to seek protection against deviations around the emergent behavior itself, which is ultimately impacted by all the above sources of uncertainty. To do so, we propose that each agent takes decisions in face of the worst-case behavior contained in an ambiguity set of tunable size, centered at the emergent behavior implicitly defined. This gives rise to a novel equilibrium notion, which we call strategically robust equilibrium. Building on its definition we show that, when judiciously operationalized via optimal transport, strategically robust equilibria
(i) interpolate between Nash and security strategies;
(ii) come at no additional computational cost compared to Nash equilibria;
(iii) often lead to better decisions and higher payoffs.
Through a variety of experiments including bi-matrix games, congestion games, and Cournot competition, we show that strategic robustness protects against uncertainty in the opponents' behavior and, surprisingly, results in higher equilibrium payoffs – an effect we refer to as coordination via robustification. Joint work with S. Fricker, S. Bolognani, F. Dörfler, and D. Paccagnan.


a_bemporad

Prof. Alberto Bemporad
IMT Lucca
Video | Slides

Solution methods for generalized Nash equilibrium problems and game-theoretic control

Generalized Nash equilibrium problems (GNEPs) arise in non-cooperative multi-agent decision making with shared constraints. This talk focuses on optimization methods for computing generalized Nash equilibria and for addressing game design and game-theoretic control problems. We first present an active-learning approach that identifies equilibria directly from best-response queries, without requiring explicit knowledge of the agents' objective functions. We then introduce a multiparametric solver for linear–quadratic GNEPs with parametric dependence, which yields explicit piecewise-affine equilibrium mappings over polyhedral regions of the parameter space. The talk concludes with an overview of a software library for solving nonlinear and linear–quadratic GNEPs, with applications to game design and game-theoretic linear–quadratic and model predictive control.


s_hall

Sophie Hall
ETH Zürich
Video

System-Theoretic Analysis of Dynamic Generalized Nash Equilibria – Turnpikes and Dissipativity

Generalized Nash equilibria are used in multi-agent control applications to model strategic interactions between agents that are coupled in the cost, dynamics, and constraints, and provide the foundations for game-theoretic MPC (Receding Horizon Games). We study properties of finite-horizon dynamic GNE trajectories from a system-theoretic perspective. We show how strict dissipativity generates the turnpike phenomenon in GNE solutions. Moreover, we establish a converse turnpike result, i.e., the implication from turnpike to strict dissipativity. We derive conditions under which the steady-state GNE is the optimal operating point and, using a game value function, we give a local characterization of the geometry of storage functions. Finally, we design linear terminal penalties that ensure dynamic GNE trajectories applied in open-loop converge to and remain at the steady-state GNE. These connections provide the foundation for future system-theoretic analysis of GNEs similar to those existing in optimal control as well as for recursive feasibility and closed-loop stability results of game-theoretic MPC.


a_churkin

Andrey Churkin
Imperial College London
Video | Slides

Understanding Aggregated Flexibility in Active Distribution Networks

The increasing integration of distributed energy resources (DERs) offers new opportunities for distribution system operators to improve network operation through flexibility services. Various DER aggregation methods have been proposed in the literature to utilise these flexible resources, including the concept of aggregated P-Q flexibility areas. Yet, flexibility aggregation in real distribution networks remains challenging due to limited DER coordination and physical network constraints. In this talk, we will discuss the principles and intuition behind aggregated flexibility provision, with a focus on optimisation and game-theoretic models. We will then dive into several open challenges: lack of DER coordination, nonlinearity of power system models, impact of voltage constraints, and flexibility provision in unbalanced networks.


s_guan

Scott Yue Guan
Georgia Tech
Video | Slides

Learning Large-Population Competitive Team Behaviors under Mean Field Interactions

Traditional multi-agent reinforcement learning algorithms face significant scalability challenges as the number of agents increases. Mean-field theory addresses this by approximating large populations as a distribution, enabling each agent to respond to aggregate behavior rather than individual interactions. However, extending this framework to complex, heterogeneous multi-agent settings, particularly those involving mixed cooperative and competitive interactions, remains largely unexplored. In this talk, we study how to model and learn team behaviors in two competing large populations. We introduce a reachability-based perspective that provides insight into when simple, symmetric team strategies can effectively capture complex competitive interactions. This viewpoint allows us to reduce the original large-scale game to a more tractable two-player formulation while preserving its essential structure. Building on this foundation, we develop a scalable learning approach, Mean-Field Multi-Agent Proximal Policy Optimization (MF-MAPPO), whose computational complexity is independent of the number of agents. We demonstrate how this framework enables efficient training of decentralized strategies, leading to improved performance and the emergence of heterogeneous behaviors in large-scale competitive multi-agent systems.


c_maheshwari

Prof. Chinmay Maheshwari
Johns Hopkins University
Video | Slides

Markov near-potential functions: A new paradigm for design and analysis of multi-agent system

Learning enabled services are revolutionizing several engineering domains such as robotics, mobility, energy, and online marketplaces. While significant progress has been made in developing autonomous agents that operate in isolation, deploying them in dynamic, multi-agent environments presents new theoretical, algorithmic, and societal challenges. Towards this goal, I will introduce a novel theoretical framework—Markov near-potential functions—to study multi-agent interactions in dynamic environments. Using this framework, I will provide the first characterization of the long-run outcome of interaction between decentralized (multi-agent) reinforcement learning (RL) agents in shared environments, without imposing constraints on their underlying utility structure. Additionally, I will demonstrate how this framework can be leveraged to design competitive, real-time planning and control strategies for autonomous multi-car racing. Finally, I will present a new multi-agent reinforcement learning pipeline – Near-Potential Policy Optimization – which exploits the structure of MNPFs to compute low-regret approximate Nash strategies in general-sum dynamic games.


b_ferguson

Prof. Bryce L. Ferguson
Darthmouth College
Video | Slides

Collaboration and Competition in Multi-Agent Systems: The Consequences of Coalitions and Model Mismatch

With the rapid growth of communication and computation technologies, systems that once operated in isolation are now coupled into large-scale multi-agent systems, including robotic fleets and autonomous transportation networks. While this increased connectivity enables new capabilities, it also introduces fundamental challenges: the collective behavior induced by decentralized decision-making is generally inefficient. This talk develops game-theoretic models for distributed control and multi-agent decision-making that quantify and shape this inefficiency. I will present recent results on the price of anarchy under structured cooperation, showing how limited coalition formation (k-strong equilibria) mediates the tradeoff between coordination complexity and system performance. These results characterize when increasing communication among agents improves efficiency, and when it introduces diminishing or even adverse returns. I will also discuss the robustness of these models to misspecification. In many settings, agents act on incorrect models of others’ objectives or behavior. We formalize this discrepancy as a game-to-real gap and show how it degrades equilibrium performance. This perspective highlights the importance of accounting for modeling error in the design and analysis of multi-agent systems.

f_laine

Dr. Forrest Laine
Vanderbilt University
Video | Slides

Mathematical Program Networks

Mathematical Program Networks (MPNs) are introduced. An MPN is a collection of interdependent Mathematical Programs (MPs) which are to be solved simultaneously, while respecting the connectivity pattern of the network defining their relationships. The network structure of an MPN impacts which decision variables each constituent mathematical program can influence, either directly or indirectly via solution graph constraints representing optimal decisions for their decedents. Many existing problem formulations can be formulated as MPNs, including Nash Equilibrium problems, multilevel optimization problems, and Equilibrium Programs with Equilibrium Constraints (EPECs), among others. The equilibrium points of an MPN correspond with the equilibrium points or solutions of these other problems. By thinking of a collection of decision problems as an MPN, a common definition of equilibrium can be used regardless of relationship between problems, and the same algorithms can be used to compute solutions. The presented framework facilitates modeling flexibility and analysis of various equilibrium points in problems involving multiple mathematical programs.


n_mehr

Prof. Negar Mehr
UC Berkeley
Video

Interactive Autonomy: Learning and Control for Multi-agent Interactions

To truly transform our lives, autonomous systems must operate in complex environments shared with other agents. For instance, delivery robots need to move through spaces that are shared with humans, and warehouse robots must coordinate in shared factory floors. Such multi-agent settings demand systematic methods that enable efficient and reliable interactions among agents. In the first part of my talk, I focus on control challenges in such domains and will discuss game-theoretic planning and control for robots. Intelligent interaction requires robots to reason about how their decisions affect and are affected by others. I will present our recent results showing how exploiting structural properties of interactions leads to motion planning algorithms that are both efficient and tractbale for real-time deployment. The second part of the talk will focus on learning in interactive domains, including imitation learning and reinforcement learning. While these approaches have advanced significantly in single-agent settings, multi-agent domains present unique learning challenges because decisions are tightly coupled across agents. I will highlight how some of these challenges can be addressed to make learning feasible in interactive multi-agent domains.


s_grammatico

Prof. Sergio Grammatico
Video | Slides
TU Delft

Stackelberg equilibrium seeking

We discuss Stackelberg equilibrium seeking in hierarchical multi-agent systems with limited information and possibly multiple follower equilibria. We first address the ill-posedness induced by multiple equilibria by introducing an optimal selection mechanism based on Tikhonov regularization. Such a mechanism yields a well-defined induced Stackelberg equilibrium and enables follower-agnostic zeroth-order seeking. Still in the setting of unknown follower models, in the second part of the talk, we present a learning-based approach that integrates zeroth-order optimization with online estimation of the followers response, hereby significantly improving efficiency and reducing interaction requirements.


s_li

Prof. Sarah H. Q. Li
Video | Slides
Georgia Institute of Technology

From Centralized Game-theoretic Coordination to Commitment Breaking in Multi-Agent Aerial Traffic Systems

This talk investigates multi-agent aerial traffic across a spectrum of coordination paradigms, from centralized air traffic planning to decentralized interaction with breakable commitments. I first discuss structured coordination through MDP congestion games and reach-avoid potential games, where agents reason strategically within a shared mathematical framework to improve traffic efficiency or enforce stronger safety objectives. These approaches illustrate how game-theoretic structure can support scalable coordination, while also revealing tradeoffs between computational tractability and guarantee strength. I then consider decentralized airspace settings such as uncontrolled airspaces in which centralized planning is unavailable and agents coordinate only through announced intentions or commitments. In these lower-information environments, coordination depends not only what agents announce, but also on their credibility. This leads to game-theoretic questions of belief mismatch, strategic deviation, and commitment breaking. The talk uses these examples to highlight a broader challenge in aerial autonomy: how to design coordination mechanisms that remain safe, tractable, and robust as centralized structure weakens and strategic behavior becomes more prominent.