A central theme of game theory research in recent decades has focused on deriving bounds on the worst-case performance of equilibria in noncooperative games. These bounds are called ”Price of Anarchy” bounds since they quantify the worst-case harm (the “price”) resulting from uncoordinated decentralized decision making (“anarchy”). The Price of Anarchy has been analyzed in a wide variety of systems (auctions, distributed resource allocation, network routing); in much recent literature, designing agent incentives or payoff functions to optimize the Price of Anarchy has been a central research goal. However, comparatively little work has focused on understanding the practical relevance of worst-case equilibria themselves.
To begin to fill this gap, our recent work suggests that worst-case equilibria may be substantially less “scary” than previously thought. Indeed, many known worst-case equilibria occur only on game instances which are finely-tuned: perturbing agent behavior on these instances often leads to dramatic improvements in efficiency. Our work shows that this feature is generic in many types of game. First, for a large class of games, we prove a fundamental tradeoff between the efficiency of equilibria and their stability: we show that if agents are highly satisfied with their actions at a particular equilibrium (i.e., the equilibrium is highly stable), then that equilibrium must be of relatively high efficiency. Second, we propose a paradigm of distributed algorithm design which exploits this tradeoff to ensure that agents never (or rarely) converge to worst-case equilibria. Finally, we discuss the exciting new research directions which are suggested by these results.
Many interactions between autonomous systems are naturally modeled as Stackelberg games, where a leader commits to a policy anticipating follower responses. While this paradigm is well-understood in static settings, its extension to dynamic environments introduces a fundamental challenge: time inconsistency. Policies that are optimal initially may become suboptimal or even infeasible as the system evolves and new information is revealed. In this talk, we show how time inconsistency and the ability to break commitments fundamentally alters the optimal control framework in hierarchical decision-making. Rather than serving as a “hero” that enables coordination and improved outcomes, dynamics allows the leader to undermine its own credibility, induce strategic deviations, and break classical optimality guarantees. We present a framework to characterize and quantify time inconsistency in dynamic Stackelberg games, drawing connections to commitment mechanisms, active learning, and dynamic programming breakdowns.
Mechanism design offers a principled approach to steering self-interested agents in societal and engineered systems toward socially desirable outcomes. A substantial body of work, anchored in the price of anarchy, has produced tight characterizations of how local utility functions and taxation mechanisms should be designed so that the worst-case Nash equilibrium is as efficient as possible. This line of work has yielded, for example, optimal taxation schemes for atomic congestion games and optimal utility designs for distributed resource allocation. In this static picture, dynamics enter only implicitly: they are the mechanism by which agents are presumed to find equilibrium, and the designer’s job ends once the equilibrium set is well-shaped. The design problem looks quite different when one places dynamics at the center. Recent work on greedy and best-response processes in resource allocation games shows that utility functions can also be designed to optimize transient, short-term efficiency, and that the utilities that are optimal for asymptotic equilibrium guarantees are not, in general, optimal for the dynamics that produce them. This talk will survey both threads of mechanism design in societal and engineered systems, the equilibrium-centric and the dynamics-centric, and argue that their current disjointness is itself a substantive open problem. The talk will close by returning to the workshop’s central question, offering a perspective on the role of dynamics in shaping outcomes in societal systems.
Standard approaches to recommender system optimization often treat user preferences and participant populations as static. This talk argues that such formulations are a fundamental model misspecification, leading to interventions that may inadvertently exacerbate societal harms or stifle platform sustainability. By adopting a control perspective, we analyze how feedback dynamics shift the assessment of algorithmic impact in two distinct contexts. First, we investigate user preference dynamics. We establish that when recommendations influence the evolution of user interests, myopic maximization of click-through rates can lead to downstream harm. We show that algorithms which prioritize long-term harm mitigation essentially avoid ”gateway” topics that are likely to lead users to engage with harmful content. Second, we examine participation dynamics in two-sided platforms of viewers and providers. While fairness-oriented exposure constraints are frequently characterized as being in tension with utility, we demonstrate that under population feedback, such interventions are instrumental in maintaining ecosystem health. Accounting for these dynamics can promote long-term growth and welfare but better than standard myopic approaches. Together, these examples suggest that while feedback loops introduce complex risks, they also provide the necessary perspective for designing better policies for broader societal interests.
Conspiracy beliefs range from the benign to the socially destructive. The latter types occur when individuals form social groups that identify with the beliefs, and cross from mere belief into collective action using radical or violent tactics. I will present preliminary findings of a large-scale multidisciplinary project aiming to tackle this problem using an integrated modelling approach. The first phase involves experiments with real people who discuss conspiracy beliefs in online chatrooms. Statistical modelling reveals how different individual thinking styles combine with the social dynamics of online discussions to reinforce or weaken the support for the conspiracy. In the second phase, Large Language Modelling will be used to analyse the chatlogs, aiming to infer individual thinking styles and develop an accurate predictor of support for the conspiracy. In the third phase, an agent-based model is introduced, and parametrised using the chatlog and experimental data. Simulations will aim to provide insight into how these chatroom dynamics might evolve over longer periods of time, with the ultimate objective of being used to model interventions and counterfactuals that could reduce the likelihood of people becoming radicalised and taking the step towards collective action.
Game-theoretic models provide a natural framework for understanding societal systems, capturing strategic interactions among self-interested agents in settings ranging from resource allocation to social networks. Such models often admit multiple equilibria, leaving open the question of which outcome best models a society of agents. Learning dynamics offer a compelling perspective in that they do more than describe how agents adapt. They also can select among equilibria.
Several results demonstrate the beneficial side of dynamics, showing how various learning rules (e.g., adaptive play, log-linear learning, higher-order reinforcement learning, stochastic satisficing) perform equilibrium selection. In each case, the dynamics resolve the static ambiguity of multiple equilibria, leading to a socially favorable outcome. An important underlying assumption is that all agents follow the prescribed learning rule.
Learning dynamics, however, can cut the other way. Results on dynamic equilibrium selection can be repurposed to show that the same dynamical mechanisms that select desirable equilibria can be exploited to deselect them. Furthermore, such destabilization can be engineered by a single malicious agent.
This talk presents an overview of selected past and recent results on this topic, landing on the conclusion that learning dynamics are neither purely heroic nor villainous. Like the anti-hero, their power cuts both ways.
Many modern socio-technical systems are characterized by the presence of multiple agents who interact and affect each other in complex ways. In such settings, agents typically do not know their best action ahead of time, as this might depend on actions of others, and instead need to learn via multiple rounds of experimentation, dynamically adapting their behavior based on received feedback and rewards. Understanding what is the overall outcome of such learning dynamics is a key question in game theory with important consequences for assessing and regulating long term system performances. While such a question has been intensively studied for static games (i.e., settings in which the agents interact according to the same game at every repetition), obtaining similar results for dynamic games (in which at each repetition the game may be in a different state with different rewards) is significantly more challenging. In this talk, we will present recent results that aim at addressing this gap for dynamic Markov games with potential structure, focusing on decentralized dynamics based on behavioral considerations, such as fictitious play.
A systems approach for evaluating the transportation needs of a specific community with limited access to their destination is presented. It is shown that with appropriate behavioral modeling and a corresponding optimization framework, their needs can be met in an effective manner.
Algorithmic Fairness has recently become an issue of major societal prominence due (primarily) to the widespread deployment of AI assistant systems (see for example the AI Act introduced by European Commission https://artificialintelligenceact.eu). While it is certainly true that AI has acted as a catalyst for a large and emerging volume of research that is centered on detecting and mitigating bias in algorithmic systems, the issue of algorithmic fairness extends well beyond AI, and is in fact a major issue when designing any form of algorithmic system that seeks to enforce a social contract (systems that seek to assist and influence humans in sharing resources). While it is perhaps not well known, Control and Systems Theory has made important contributions to this domain in the past, and is now well positioned to assist in the development of contemporary societal systems that are fair by design. Our objective in this talk is to introduce several problems where the issue of algorithmic fairness has emerged as a specific challenge in a Control and Systems Theory context. A specific focus of the talk will be on problems where instability by design (in non-linear systems) is used as a mechanism for fairness, on pricing algorithms, and on new directions of emergent research in the area of Optimal Transport (arising from the Autofair project). We shall conclude by presenting several open problems in this new and emerging domain.