Multi-Agent Learning Seminar Series at UPenn: Exploring the Frontiers of Coordinated Intelligence

The real world is inherently multi-agent, consisting of numerous entities that must interact and coordinate in complex, dynamic environments. Achieving effective operation in such settings requires coordinated intelligence systems in which agents not only collaborate effectively but also make trustworthy decisions, addressing practical concerns such as scalability, robustness, safety, and fairness. The University of Pennsylvania (UPenn) hosts a seminar series dedicated to exploring the cutting edge of multi-agent learning, optimization, and related fields. This article provides an overview of the seminar series, highlighting key topics, research areas, and speakers.

A Hub for Optimization and Multi-Agent Learning Research

This seminar serves as a university-wide hub to bring together the many optimization communities across UPenn --- in the Departments of Statistics and Data Science, Electrical Engineering, Computer Science, Applied Mathematics, Economics, Wharton OID, etc. The seminar series features leading experts in optimization and adjacent fields.

Seminar Schedule and Format

Seminars will be held on Fridays, from 11:00 AM - 12:30 PM. The format, however, may vary - in-person seminars will be held in CHE 2108. Seminars will be held on Mondays, from 10:00 - 11:00 AM. The format, however, may vary - in-person seminars will be held in CHE 2110. Seminars will be held virtually on Tuesdays at 11:00 am until further notice. Seminars will be Tuesdays at 11:00 am in room 2110 of the Chemical and Nuclear Engineering Building, unless otherwise noted.*Please note: Due to the COVID-19 outbreak, all seminars for the remainder of the semester have been cancelled.

Key Research Areas and Topics

The seminar series covers a broad range of topics within multi-agent learning, optimization, and related areas. Here's a glimpse into some of the key research areas:

Multi-Agent Reinforcement Learning (MARL)

MARL is a central theme, addressing challenges related to safety, efficiency, robustness, and security in embodied AI and cyber-physical systems. Research explores how agents can learn to cooperate and coordinate in complex environments.

Communication and Coordination: A key focus is on enabling effective communication among agents to enhance object detection accuracy and reduce uncertainty. This includes designing methods for uncertainty quantification in collaborative perception, particularly in connected autonomous vehicles (CAVs).

Incentivizing Cooperation: Researchers are developing novel reward reallocation schemes, such as those based on Shapley value, to incentivize agents to communicate and coordinate effectively in MARL settings.

Robustness and Security: Addressing the challenges of state uncertainties and potential adversarial attacks is crucial. Theoretical analyses of robust MARL methods are explored to ensure system resilience.

Optimization Algorithms and Techniques

The seminar series delves into the design and analysis of various optimization algorithms, with a focus on both theoretical foundations and practical applications.

First-Order Methods: Computational tools are presented for analyzing and designing first-order methods in parametric convex optimization. These methods are popular for their low per-iteration cost and warm-starting capabilities.

Worst-Case Performance Analysis: Numerical frameworks are used to verify the worst-case performance of first-order methods in parametric quadratic optimization. This involves formulating mixed-integer linear programs and incorporating bound-tightening techniques.

Data-Driven Optimization: Statistical learning theory is applied to analyze the performance of first-order methods, establishing generalization guarantees and learning accelerated optimizers.

Nonconvex Constrained Optimization: Algorithmic strategies for solving nonconvex constrained optimization problems, which may arise in informed supervised learning, are explored.

Game Theory and Mechanism Design

The intersection of game theory and mechanism design with machine learning is a prominent area of investigation.

Online Learning in Games: Research establishes near-optimal convergence rates for online learning algorithms in normal-form games and develops new algorithms for minimizing swap regret.

Read also: Education for FBI Agents

Mechanism Design with Imperfect Information: The seminar explores scenarios where buyers have partial information about their valuations, and the seller can acquire information about buyers’ valuations. This includes studying how a seller’s ability to condition payments on partial information can alter the structure of the optimal mechanism.

Scalability and Distributed Optimization

Addressing the scalability challenges in multi-agent systems is a critical focus.

Mean-Field Approximation: The mean-field approximation is introduced as a way to simplify interactions among a large population of agents. Theoretical analysis and convergence results are presented for entropy-regularized mean-field games.

Hierarchical Decomposition: The use of hierarchical decomposition is explored to break down large games into smaller, more manageable sub-games.

Distributed Optimization: The problem of resilient distributed multi-agent optimization for cyber-physical systems in the presence of malicious or non-cooperative agents is considered. This involves developing trustworthy computational models that combine tools from statistical learning and distributed consensus-based optimization.

Controlled Generation in Large Language Models (LLMs) and Diffusion Models

The seminar explores methods and theory for controlled generation within LLMs and diffusion models.

Ecosystem-Level Outcomes in Machine Learning

Research focuses on characterizing and steering ecosystem-level outcomes in machine learning, taking an economic and statistical perspective.

Uncertainty Quantification and Conformal Prediction

The seminar investigates methods for quantifying uncertainty in machine learning predictions and providing precise error bounds.

Conditional Calibration: Algorithms are presented that can produce calibration guarantees that hold conditionally on any bounded set of conditioning events.

Conformal Prediction: Advancements in conformal prediction are presented, providing precise error bounds on ML predictions.

Sequential Statistical Testing: Sequential statistical testing is used to enable test-time training schemes, allowing pre-trained models to adapt online to unfamiliar environments.

Understanding and Eliciting Emergent Coordination

The seminar explores how to understand and elicit emergent coordination/cooperation in large multi-agent systems.

Detection of AI-Generated Content

Approaches to detect AI-generated content via watermarking are discussed.

Speakers and Their Research

The seminar series features a diverse range of speakers from various universities and research institutions. Here are some of the highlighted speakers and their research areas:

Fei Miao (University of Connecticut): Dr. Miao's research focuses on multi-agent reinforcement learning, robust optimization, uncertainty quantification, and game theory, to address safety, efficiency, robustness, and security challenges of Embodied AI and CPS, for systems such as connected autonomous vehicles, sustainable and intelligent transportation systems, and smart cities.

Meena Jagadeesan (UC Berkeley): Ms. Jagadeesan's research investigates multi-agent interactions in machine learning ecosystems from an economic and statistical perspective.

Guillaume Sartoretti (National University of Singapore): Dr. Sartoretti's research lies in understanding and eliciting emergent coordination/cooperation in large multi-agent systems, by identifying what information and mechanisms can help agents reason about their individual role/contribution to each other and to the team.

Noah Golowich (Massachusetts Institute of Technology): Mr. Golowich's research interests lie in theoretical machine learning, with a particular focus on the connections between multi-agent learning, game theory, online learning, and theoretical reinforcement learning.

Specific Seminar Presentations

Here are summaries of some specific seminar presentations:

Communication, Learning, and Control in Multi-Agent Systems: This talk focuses on learning and control with communication capabilities, designing an uncertainty quantification method for collaborative perception in connected autonomous vehicles (CAVs), and developing a safe and scalable deep multi-agent reinforcement learning (MARL) framework.

Ecosystem-Level Outcomes in Machine Learning: This talk discusses research on characterizing and steering ecosystem-level outcomes in machine learning, taking an economic and statistical perspective.

Multi-Agent Scalability by Learning Emergent Coordination: This talk summarizes early work in independent learning and discusses recent advances in convention, communication, and context-based learning within a wide variety of robotic applications.

Resilient Distributed Multi-Agent Optimization for Cyber-Physical Systems: This talk considers the problem of resilient distributed multi-agent optimization for cyber-physical systems in the presence of malicious or non-cooperative agents.

Taming Gradient Descent Ascent: This talk focuses on solving min-max problems and shows that unconventional stepsize schedules are necessary for convergence.

Multi-Agent Learning: Convergence and Watermarking: This talk discusses results which address the question of failure of convergence for iterative algorithms such as gradient descent and an approach to detect such content via watermarking.

Hybrid Reinforcement Learning and Machine Unlearning: This talk discusses work on Hybrid RL, which has led to the development of the first general-purpose, computationally efficient, and theoretically rigorous algorithms for RL, and work on the foundations of machine unlearning.

Computational Tools for Analyzing and Designing First-Order Methods in Parametric Convex Optimization: This talk presents computational tools for analyzing and designing first-order methods in parametric convex optimization.

Reliable Machine Learning in High-Stakes Applications: This talk presents a new advancement in conformal prediction and shows how sequential statistical testing can enable a novel test-time training scheme, allowing a pre-trained model to adapt online to unfamiliar environments.

Data Complexity and Generalization in Neural Networks: This talk studies the generalization properties of neural networks through the lens of data complexity.

Stochastic Gradient-Based Algorithms for Nonconvex Constrained Optimization: This talk presents recent work on the design and analysis of stochastic-gradient-based algorithms for solving nonconvex constrained optimization problems.

Gradient Equilibrium: A New Perspective on Online Learning: This talk presents a new perspective on online learning that we refer to as gradient equilibrium.

Cooperative Autonomous Vehicle Behaviors for Traffic Smoothing at Scale: This talk shows that we can use reinforcement learning to design AV behaviors that operate cooperatively to smooth traffic in large, realistic simulators and performed a large-scale road test, the first of its kind, in which we deployed a hundred of these cruise controllers onto a highway to show traffic smoothing at scale.

Remedying Scalability Challenges in Multi-Agent Decision-Making: This talk will focus on two aspects of the scalability challenge: (i) number of agents, and (ii) large state space.

Expanding Electronic Systems Beyond Conventional Operating Regimes

Emerging computing and sensing applications increasingly demand electronic devices that operate beyond the limits of conventional CMOS scaling, requiring improved energy efficiency, robustness under extreme conditions, and new modes of integration with biological and physical environments. In this seminar, I will first present my Ph.D. research on ferroelectric memory devices based on wurtzite III-nitride ferroelectrics, with a focus on AlScN-gated ferroelectric FETs, and ferroelectric diodes. I will then outline my broader research vision aimed at expanding electronic systems beyond conventional operating regimes. This includes (i) electronics robust to extreme environments such as high temperature and radiation, (ii) low-temperature and BEOL-compatible device integration for monolithic 3D architectures, and (iii) transient and bio-resorbable electronic systems for healthcare monitoring and space applications.

tags: #multi #agent #learning #seminar #upenn #schedule