Deepanshu Vasal
Research Scientist, Northwestern University (He/Him/His)
Research Areas: Multi Agent Reinforcement Learning, Machine learning based control, Decentralized Control, Autonomous Systems, Game Theory, Mechanism Design, Mean Field Games, Feedback Communications, Networks
Contact: dvasal AT umich DOT edu
Education
PhD, University of Michigan, Ann Arbor, 2016
MS, Mathematics, University of Michigan, 2013
MS, Electrical Engineering : Systems, University of Michigan, 2011
BTech, Electronics and Communication Engineering, IIT-Guwahati, 2009
In my research, I have developed new tools to analyze problems in decentralized decision making. I endeavor to focus on fundamental, conceptual problems, with a goal of providing real world engineering solutions. It is very rare in research that one sees new ideas (most papers are applications of known ideas to different domains). In my
research, I present at least two new ideas that were not known before (in dynamic games of incomplete information and communication channels with noisy feedback)
Multi-agent Reinforcement Learning with decentralized information
Reinforcement learning has emerged as a powerful tool to learn the optimum policies of the players when the underlying model is not known. In this work, we provide reinforcement learning algorithms for multi-agent problems when users are strategic and cooperative and have decentralized information. We use the idea of particle filters in reinforcement learning to show that the RL policies converge to the optimum/equilibrium strategies
Stochastic Stackelberg (mean-field) games
In such games, there are one or more leaders who commit to dynamic policies and a finite or infinite number of followers best respond to it and to each other. In this work, we propose a sequential algorithm to compute the policies of all the players. This is a new idea in the theory of such games which opens numerous applications in mechanism design and designing policies of the firms or the governments when interacting with a market
Dynamic games of asymmetric information
Dynamic games of asymmetric information are models of interaction of strategic users when they have private information and they interact in a stochastic/dynamic fashion. In this series of papers, we provide a general sequential decomposition framework to study many classes of such games.
Belief Propagation
Belief Propagation (BP) is a fundamental message passing algorithm to compute marginals from joint distribution where the underlying random variables have dependency structure defined through a graph. It is known that BP is exact on trees and on graphs with at most one loop. In this series of papers we look at different aspects of BP such as multi-agent BP with decentralized information, convergence of BP to global optimum of Bethe energy function and more.
Mean field games
Mean field games is a model that approximates Nash equilibria in large player games of incomplete information. It consists of solving a coupled set of forward (Fokker Planck Kolmogorov) and backward (HJB) fixed-point equations that are coupled across time. In this series of papers, we show that such games can be solved by solving a smaller fixed point equation for each time t, which significantly reduces the complexity of finding mean field equilibria. [Video]
Game Theory and Mechanism/Information Design
Nash introduced the concept of Nash equilibrium that provides a mathematical framework to study strategic behavior. We argue that some Nash equilibria are more plausible than the others, the ones that are more robust to irrational players. We define a notion of alpha-robustness such that if at least N-alpha-1 players are playing the equilibrium strategies (and alpha players play arbitrary strategies), then no user wants to unilaterally deviate. We provide sufficient conditions for the existence of such equilibria. [Video]
Mechanism design is an engineering side of game theory that builds systems such that when acted upon by strategic users, they take actions as intentioned by the system designer. Myerson's optimal auction and the VCG mechanism, while conceptually and mathematically being quite simple, provide solutions to real world problems that are implemented in practice and thus have significant impact. In the same spirit, we design a private good mechanism for large scale users with the objective of social welfare which generalizes the VCG mechanism when the number of players is large, such as in a society.
We also study dynamic information design. [Video]
Multi-agent team decision problems
There are many instances when players with different information are trying to control a system. Witsenhausen's counterexample is one of the classic work which shows that the results of complete information don't easily flow through in these problems and such problems are quite challenging . In this line of work, we present structural results to find optimum policies when the players have different information. Specifically we consider the case when players don't have any common information and to the best of our knowledge, is the first such work that provides a concept of state for such systems. The fundamental issue in such problems is that due to lack of common information there is an infinite regress of higher order beliefs. However, by a clever construct I show that for the LQG case, the update and thus the dynamics of the 1st order beliefs can be made same as that of 2nd order and so on and thus it is sufficient to keep track of one layer of beliefs. This addresses a fundamental problem in the control of these systems and a problem that has been discussed many times in the economics literature as well.
Communication channels with noisy feedback
Shannon proved there exists a fundamental capacity of a communication channel such that one can achieve an arbitrarily low probability of error for all rates below this capacity. Communication is a $1.7 trillion (per annum) industry worldwide which includes cell phones, the internet, optical communication, and space communication. However, all communication is one-way. We don’t use feedback in communication in any meaningful way. While it is known that feedback doesn't increase capacity, it can significantly increase the error exponents and could simplify coding and decoding. However, so far there is no mathematical framework to study such channels with noisy feedback, because of which it hasn't been used in practice so far in any meaningful way. In this work, we provide a sequential decomposition framework to study such channels which is the first mathematical treatment of such channels that allows a concept of state and a methodology to find optimal Markovian strategies with respect to this state. [Video]
Capacity of Multiple access channel with feedback
While we know the capacity of a point-to-point channel with and without feedback and have achievable codes to achieve that, Shannon's capacity and capacity-achieving schemes of networked communication systems are not always known. In particular, the capacity of MAC with feedback has been an open problem for 65 years. Using recent developments in decentralized stochastic control, I pose the capacity achieving in a MAC with noiseless feedback as a stochastic optimization problem to provide a single letter capacity expression through a dynamic program, which provides the first analytical method to compute the capacity of this channel, and opens door to studying more such multi-user channels with feedback.
Network Games
In the past few years, information design has emerged as a powerful tool parallel to mechanism design. In this line of work, we study a new dimension we call network design where the principal agent has the ability to design connections among the selfish agents such that the corresponding Nash equilibrium of the networked game coincide with the outcome desired by the principal agent. [Video]
In related work on network games, we combine ideas from random matrix theory (RMT) and linear quadratic games to show a phase transition in such games with respect to perturbation in the adjacency matrix of the network.
To the best of our knowledge, this is the first such line of work that combines ideas from game theory and RMT to characterize such a phenomenon of phase transitions in the real world, and opens door to studying many such phenomena in real life through the intersection of these fields. [Video]
Learning in games with fully rational agents
Learning in games is an important problem of how agents learn to play Nash equilibrium. There are many ways this question has been addressed so far, either using heuristics such as fictitious play, or some notion of the rationality of the agents. However, most of these algorithms are either just heuristics or myopic best responses and there are questions as to why selfish rational agents would play such a learning algorithm. This question is even more important today with technology playing a huge role in decision making and significant advancements have been made at the intersection of machine learning, artificial intelligence with game theory, where machines with high computation power can arguably be fully rational. In this work, we design a decentralized learning algorithm with fully rational selfish agents such that (a) the algorithm converges to the Nash equilibrium of the static game and (b) the algorithm itself is an equilibrium of the dynamic learning process. To the best of our knowledge, this is the first such algorithm with these two properties.
News
Giving a talk at IIM Bangalore on Stackelberg mean field games on March 14, 2024
Presenting my work on sequential linear coding for multi-user Gaussian channels with noisy feedback at EPFL on May 4, 2023
Presented paper on the master equation for discrete-time Stackelberg mean field games at CDC 2022
Presenting my work on multi-user Gaussian channels with noisy feedback at ETH Zurich on Oct 6
Presenting my work on Stackelberg mean field games at IIT Bombay (EE/IEOR) on August 1, UC Berkeley on August 12, UIUC on Sept 2, UT Austin on September 8, Google NYC on Sept 13, Caltech Sept 22
Presenting paper on Master equation for discrete-time Stackelberg mean field games at Stony Brook Game Theory Conference
Presenting paper on Sequential decomposition of stochastic Stackelberg games at ACC 2022
Giving a talk on feedback communication at Purdue neXt-Generation-wireless (XG) center on March 25th (online) [Slides]
Presenting my work on Stackelberg mean field games at Berkeley IEOR on March 11 (Online)
Paper on sequential decomposition of Stackelberg games accepted at ACC 2022
Presenting my work on noisy feedback communication at Stanford Information Theory forum on Feb 11th (online) [Slides] [Video]
Paper on framework for studying decentralized Bayesian learning with strategic agents accepted at Stochastic Systems 2021
Paper on phase transition in Network games was accepted at GameNets 2021
Paper on Gaussian channels with feedback was accepted at ISIT 2021
Three papers accepted to CISS: Paper on multi agent reinforcement learning (invited), large scale network utility maximization, and network design for social welfare
Paper on sequential decomposition of graphon mean field games has been accepted at ACC 2021 (invited)!
Paper on Signaling equilibria in dynamic LQG games accepted at TCNS!
Papers on reinforcement learning in Stackelberg games and Mean field games have been accepted at CDC 2020!
Paper on sequential decomposition of mean field games was presented at ACC 2020
Paper on a new MAC channel protocol was presented at CDC 2019
Paper on repeated asymmetric information games was presented at ACC 2019
Working Papers/Preprints
Deepanshu Vasal, Master equation of discrete-time Stackelberg graphon mean-field games with a single leader
Deepanshu Vasal, Master equation of discrete time Stackelberg graphon mean field games with multiple leaders
Deepanshu Vasal, Stochastic LQG Stackelberg games with single and multiple leaders
Deepanshu Vasal, A policy gradient algorithm for risk-sensitive mean field games and teams
Deepanshu Vasal, A neural networks-based approach to multiple access channel with feedback
Deepanshu Vasal, Homogenous games with local interactions
Deepanshu Vasal, Alpha-tolerant network games
Deepanshu Vasal, Structured equilibria for dynamic games with private monitoring
Deepanshu Vasal, Partially observed graphon mean field games
Deepanshu Vasal, Dynamic program for linear sequential coding for Gaussian (a) point-to-point channel with active noisy feedback, (b) two-way channel, (c) MAC with noisy feedback, (d) broadcast channel with active feedback, (e) relay channel with active feedback, (f) MIMO channel with active feedback, (g) Interference channel with active noisy feedback
Deepanshu Vasal, Equilibrium policies and value function of Aumann Maschler games
Deepanshu Vasal, Partially observable mean field games and teams with common state
Deepanshu Vasal, Alpha-tolerant mean field games
Deepanshu Vasal, A comment on dynamic pivot mechanism
Deepanshu Vasal, Reinforcement learning in graphon mean field teams and games
Deepanshu Vasal, Multi-agent multi-arm bandits with decentralized information
Deepanshu Vasal, Mult arm bandits in non-stationary mean-field games
Deepanshu Vasal, Multi-agent reinforcement learning with hidden actions
Deepanshu Vasal, SPBE for repeated games with static dependent types
Deepanshu Vasal, Existence of SPBE in dynamic games of asymmetric information
Deepanshu Vasal, Dynamic auctions in non-stationary mean-field games
Sameer Mathad, Deepanshu Vasal, David Love, A linear sequential scheme for Gaussian point-to-point channel with noisy feedback
Deepanshu Vasal, Sameer Mathad, David Love, A linear sequential scheme for AWGN MAC with noisy feedback
Deepanshu Vasal, A folk theorem for incomplete information games
Deepanshu Vasal, Decentralized learning in Markov potential games of incomplete information
Deepanshu Vasal, A posterior matching scheme for multiple access channel with noiseless feedback
Deepanshu Vasal, Indirect dynamic pivot mechanism
Deepanshu Vasal, Pivot mechanisms in mean field games
Deepanshu Vasal, Credible double auctions
Deepanshu Vasal, On the existence of Nash equilibrium in large network games
Deepanshu Vasal, Mean field teams and games with correlated types
Deepanshu Vasal and Randall Berry, Fault Tolerant Equilibria in Anonymous Games: best response correspondences and fixed points, [Video]
Deepanshu Vasal, Structural results for monotone policies in dynamic games of incomplete information
Deepanshu Vasal, Dynamic war of attrition
Deepanshu Vasal, Decentralized control of LQG teams and games with dependent types
Deepanshu Vasal, A dynamic program for achieving capacity for multiple access channel with noiseless feedback
Deepanshu Vasal, On zero-error capacity of multiple access channel with noseless feedback
Deepanshu Vasal, A mean-field analysis of cryptocurrencies
Deepanshu Vasal, A framework to study informational cascades with partially observed mean field games
Journal Publications
Published/Accepted
Meng Zhang and Deepanshu Vasal, Mechanism Design for Large-Scale Network Utility Maximization, IEEE Transactions on Mobile Computing, [Video]
Deepanshu Vasal, Sequential decomposition of mean-field games [Video], Dynamic games and applications, 2023
Rajesh Mishra*, Deepanshu Vasal* and Hyeji Kim, Linear Coding for AWGN channels with Noisy Output Feedback via Dynamic Programming, Transactions on Information Theory, 2023
Rajesh Mishra, Deepanshu Vasal and Sriram Vishwanath, Model-free reinforcement learning in mean field games, Transactions of controlled network systems (TCNS), 2023
Deepanshu Vasal and Achilleas Anastasopoulos, A framework for studying decentralized Bayesian learning with strategic agents, Stochastic Systems 2022
Deepanshu Vasal and Achilleas Anastasopoulos, Signaling equilibria for dynamic LQG games with asymmetric information, IEEE Transactions on control of networked systems (TCNS) 2021
Deepanshu Vasal, Abhinav Sinha and Achilleas Anastasopoulos, A systematic process for evaluating structured perfect Bayesian equilibria in dynamic games with asymmetric information, Transactions on Automatic Control, 2018
Deepanshu Vasal and Achilleas Anastasopoulos, Stochastic Control of Relay channels with cooperative and strategic users, IEEE Transactions on Communications, 2014
Under Review
Deepanshu Vasal, Master equation for discrete-time Stackelberg mean-field games with a single leader, under major revision, Transactions on Automatic Control
Deepanshu Vasal, Master equation of discrete time graphon mean field teams and games
Deepanshu Vasal, Sequential coding for the two-way channel
Yitao Chen and Deepanshu Vasal, Multi-agent decentralized belief propagation on graphs
Abhishek Shende, Deepanshu Vasal and Sriram Vishwanath, Design and analysis of networked LQ games
Rajesh Mishra, Deepanshu Vasal and Sriram Vishwanath, Decentralized multi-agent reinforcement learning with shared actions
Deepanshu Vasal, Sequential decomposition of stochastic Stackelberg games with single leader , [code]
Deepanshu Vasal, Sequential decomposition of stochastic Stackelberg games with multiple leaders
Deepanshu Vasal, Sequential decomposition of partially observable mean field games and teams
Deepanshu Vasal, Master equation of discrete time Stackelberg mean field games with multiple leaders
Deepanshu Vasal Structured equilibria for dynamic games with public monitoring
Deepanshu Vasal, A mean-field model of social herding
Deepanshu Vasal, Sequential decomposition of the point-to-point channel with noisy feedback
Deepanshu Vasal, Learning in quadratic games with fully rational agents
Conference Publications
Deepanshu Vasal and Randall Berry, Master equation for discrete-time Stackelberg mean-field games with a single leader, CDC 2022
Abhishek Shende, Deepanshu Vasal and Sriram Vishwanath, A phase transition in large networked games, GameNets 2021 [Video]
Yitao Chen and Deepanshu Vasal, Convergence of belief propagation to the global minima of Bethe energy function in graphs with motifs, submitted
Abhishek Shende, Deepanshu Vasal and Sriram Vishwanath, Network design for social welfare, CISS 2021 [Video] [Slides]
Rajesh Mishra*, Deepanshu Vasal* and Hyeji Kim, Gaussian channels with feedback: A dynamic programming approach, ISIT 2021
Meng Zhang and Deepanshu Vasal, Mechanism design for large scale systems, CISS 2021 [Video] [Slides]
Rajesh Mishra, Deepanshu Vasal and Sriram Vishwanath, Decentralized multi agent reinforcement learning with shared actions, CISS 2021 (invited)
Deepanshu Vasal, Rajesh Mishra and Sriram Vishwanath, Sequential decomposition of graphon mean field games, ACC 2021 (invited)
Rajesh K Mishra, Deepanshu Vasal and Sriram Vishwanath, Learning Stochastic Stackelberg Security Games, CDC 2020
Rajesh K Mishra, Deepanshu Vasal and Sriram Vishwanath, Reinforcement Learning in Non-stationary Mean Field Games, CDC 2020
Deepanshu Vasal, Learning arrival rates to improve common information based multiple access protocol, Conference on Decision and Control (CDC), 2019
Deepanshu Vasal, Stochastic Stackelberg security games, ACC 2022 (accepted) [code] [Video]
Deepanshu Vasal, Sequential decomposition of mean field games ACC 2020 [Code][Slides] [Video]
Deepanshu Vasal, Sequential decomposition of repeated games with asymmetric information and dependent states, American Control Conference, 2019 [Slides]
Deepanshu Vasal, Impact of social connectivity on herding behavior, NETGCOOP 2018 [Slides]
Deepanshu Vasal and Achilleas Anastasopoulos, Decentralized Bayesian learning in dynamic games, Allerton Conference on Communication, Control, and Computing, 2016 [Slides]
Deepanshu Vasal and Achilleas Anastasopoulos, Signaling equilibria for dynamic LQG games with asymmetric information, Conference on Decision and Control (CDC), 2016 [Slides]
Deepanshu Vasal, Vijay Subramanian and Achilleas Anastasopoulos, Incentive Design for Learning in user-recommendation systems with time-varying state, Asilomar Conference on Signals, Systems, and Computers, 2015 (Invited) [Slides]
Deepanshu Vasal and Achilleas Anastasopoulos, A systematic process for evaluating structured perfect Bayesian equilibria in dynamic games with asymmetric information, American Control Conference (ACC), 2016 (Available on Arxiv) [Slides]
Deepanshu Vasal and Achilleas Anastasopoulos, Achieving socially optimal solution through payments in a dynamic game for the relay channel, Allerton Conference on Communication, Control, and Computing, 2012
*indicates that authors have same contribution