Projects

We consider a strategic task (coordination game) distributed over a network where the nodes are pairs of a human and an AI agent acting as an assistant. Due to the diversity in cognitive abilities of humans, and different computational intelligence limits of the AI assistants, the network system is highly heterogeneous, which is captured by individual node rationality parameters. Nevertheless, the agents interact and with the goal of learning the optimal action over the network under strict limits on rationality and connectivity. From a designer’s perspective, the finite network design is a problem with phenomenal complexities. First, game theoretic learning with agents with heterogeneous rationalities is understudied and lack closed form expressions that would lead to specific objective functions and/or constraints for mathematical optimization; second, the optimization over space of networks has combinatorial complexity (likely NP hard). Third, current methods are not scalable and prohibit the analytical design of realistic networks with probabilistic performance guarantees. Based on our existing results, we first propose the optimal design of graphons, a limit graph for a population of agents with a continuous distribution of rationalities. By using a combination graphons, population games, and optimal control techniques, we can obtain optimal designs from which finite networks with probabilistic guarantees can be sampled from, bypassing the difficulties and complexities of finite network design for a given rationality profile for the decision-making agents. Then, we invert the question by fixing the network among the agents and design a network intervention of the following form. Since AI agents can in principle be controlled by investing in better models and computational resources, we can tune the rationality of the nodes. The problem becomes to optimize the distribution of a rationality budget over a graph. Using the notion of game play under bounded rationality we optimize the outcome of the game when the agents adopt a Quantal Response Equilibrium. Our preliminary work shows that when the network is a tree (a graph without cycles), the optimization problem is convex and admits a unique solution. However, the optimal rationality allocation for a general graph remains an open nonconvex problem, which requires the development of yet to be discovered techniques. Our research aims at providing tools and design principles to achieve reliable human-AI collective decision-making behavior that maximizes the chances of mission success. (Joint work with Tony Kwasnica, FSU).

Most information today is produced for consumption on digital social media platforms such as YouTube, Instagram, and TikTok. Because digital platforms have the ability to collect preference data, they can sustain user engagement with content for longer periods, thereby maximizing their profits. However, users do not always share the same objectives as the platform. They may decide to drop out or withhold their preference data, sacrificing a personalized user experience in order to maintain their privacy. Using a game-theoretic model inspired by the pioneering work of economists Milgrom and Grossman, we model the interaction between a social media platform and its user base, obtaining a proof for the existence of a perfect Bayesian equilibrium in a very general stochastic setting. Surprisingly, the equilibrium policies demonstrate the “enrage to engage” phenomenon, in which the platform actively drives user engagement by recommending (i.e., promoting) content that is disturbing or uncomfortable to many of its users. However, many questions remain open, such as the sensitivity of equilibrium policies to probability distributional shifts, and how platforms leverage the bounded rationality of their user base to maximize profits. On the one hand, this project aims to understand and characterize social behavior; on the other, it seeks to design recommendation algorithms that are “no evil,” in the sense of committing to not show disturbing content. Our research goal is to establish the necessary and sufficient conditions for the optimality of “no evil” recommendation policies, which would reduce the psychological stress that burdens our ever-connected society. (Joint work with Odilon Camara, University of Southern California).

When a group of agents (team) distributed over a network make decisions based on uncertainty in observations, the cohesiveness of collective behavior declines which may hinder missions, create systemic vulnerabilities and ultimately decrease of probability of successful operations. Despite the existence of a framework to analyze strategic team coordination with noisy observations using the notion of Global Games (stochastic coordination games with incomplete information), we still do not know how to measure coordination, and characterize its degradation as a function of: (1) the network connectivity (graph sparsity) among the agents; (2) the amount of observation noise in the agents observations; and (3) how delay in acting may cascade into a stale team that fails to react in time to respond to threats and take opportunities. Our preliminary results show that in fully connected networks, using an interpretable probabilistic metric of coordination among agents with an omniscient fictitious agent, we can characterize such degradation of coordination performance. Moreover, using simple information theoretic bounds, we can predict regimes where coordination above a certain level is outright impossible for a given level of observation noise. However, many questions remain open such as how to extend these results to obtain the ultimate limits of networked coordination (coordination capacity). We also propose the analysis of a two-stage networked coordination game, where the agents act in the first state under uncertainty with an undiscounted reward, and with certainty in the second stage and a discounted reward. Therefore, in this problem, the option to delay creates an additional uncertainty resolution mechanism which can be either detrimental or beneficial for coordination success. A research task is to characterize when such option to delay can be beneficial, which would dictate higher order decision-making from the point of view of mechanism design whether allowing agents engaged in mission critical coordinated tasks should be given the option to delay and to resolve uncertainty. Our approaches will be based both on a population level for a given graphon, or a finite system level using graphs. The goal is to predict and quantify team performance in noisy and communication constrained environments that characterize many applications. (Joint work with Behrouz Touri, UIUC)

In an information rich world, we are often confronted with the question of what data are worth transmitting or storing and what to discard. Consider a communication-denied environment, where only a small fraction of sensed data can be communicated to a remote control station that needs to reconstruct the entire dataset prior to making a relevant decision. Instead of using a traditional approach based on compressed sensing (which only is beneficial under the sparsity structure of signals), we propose using game theoretic approach, known as Information Design (ID). Based on ID, we obtained structural results for optimal data transmission originating from independent Gaussian information sources. Surprisingly, these results depend only on the inherent symmetry of the Gaussian density and can be generalized to a wide family of distributions. Closed-form results, however, cannot be obtained when the sources are correlated unless there is symmetry in the covariance matrix structure. We propose expanding the use of ID into settings that are (1) dynamic (e.g. Gauss-Markov sources), and (2) environments where the transmitter and the receiver do not know the distribution at all. This requires learning what to send, store or discard solely based on data. We will pursue both parametric and nonparametric approaches, utilizing neural networks to discover the optimal structure of optimal data transmission policies. The primary innovation of using ID relative to conventional approaches is the ability to leverage signaling, a phenomenon characteristic of intelligent decision-makers (here, the transmitter and receiver). The core idea is that even when something is not transmitted/stored, critical information can be reconstructed by using the policy itself. For example, if we compare two numbers, decide to send the larger and discard the smaller, the receiver inherently knows that the number discarded was smaller than the one it received. This represents valuable information that can be leveraged for decision-making at higher layers of control in communication denied or extremely constrained applications.