11/11/2024
By Akshay Kolli
The Kennedy College of Sciences, Miner School of Computer & Information Sciences, invites you to attend a master's thesis defense by Akshay Kolli on "Graph attention inference of network topology in multi-agent systems."
Date and Time: Friday, Nov. 22, 2024
Time: Noon to 2 p.m.
Location: 309 Dandeneau Hall, North Campus
Committee Members:
- Kshitij Jerath (Advisor), Mechanical and Industrial Engineering, UMass Lowell
- Reza Azadeh, Ph.D. (Member), Miner School of Computer & Information Sciences, UMass Lowell
- Hadi Amiri, Ph.D. (Member), Miner School of Computer & Information Sciences, UMass Lowell
Abstract:
In multi-agent systems with a group of interacting agents, whether they be humans or robots, the structure of interactions between agents can typically be modeled as a graph. Each agent can be represented as a node in a graph, and the connections or relationships between agents are represented by edges. The ability to accurately infer or reconstruct this underlying graph structure is of paramount importance in understanding the system's dynamics and optimizing control strategies. The graph structure dictates how the state of the system evolves with time, allowing us to study a system's properties by understanding the graph. In many, however, determining these graph structures, particularly in systems with complex dynamics, remains a significant challenge in the field of systems theory, control, and network science.
This thesis introduces a machine learning-based approach that uses attention mechanisms to address this challenge. Attention mechanisms, initially developed in natural language processing, have proven effective in selectively focusing on different parts of the input data, making them ideal for applications where relationships between individual elements need to be understood. By applying this mechanism to multi-agent systems, we propose a method that learns to predict the future states of the system by learning good representations of each agent present in the form of a n-dimensional vector, where n is determined heuristically. These learned representations provide insight into the underlying graph structure of the system, with the strength of the attention values serving as an implicit measure of the connectivity between agents.
The approach is evaluated in the context of two well-known dynamic models in multi-agent systems: linear consensus dynamics and the non-linear dynamics of Kuramoto oscillators. Both models are fundamental in studying synchronization, agreement, and collective behavior in multi-agent systems. Linear consensus dynamics are widely used in applications like flocking, opinion dynamics, and distributed computing, while Kuramoto oscillators are canonical models for studying synchronization phenomena in fields ranging from neuroscience to power grids. By demonstrating the ability of our attention-based machine learning model to learn the network topology of these systems without explicit knowledge of the underlying dynamics, we establish a new data-driven framework for graph inference in multi-agent systems.
Our experimental results, measured through F1 scores in link prediction tasks, show that this approach can successfully identify network structures, with the model being unaware of dynamics of the data it's being trained on.
This thesis will explore the theoretical foundation of our approach, the design and implementation of the machine learning model, and its application to both linear and non-linear dynamical systems. We will also discuss the broader implications of using machine learning for graph inference, potential real-world applications, and future directions for extending this work. By providing a machine learning framework that implicitly learns network topologies, we open new possibilities for improving our understanding and control of complex multi-agent systems.