Symbolic Model Extraction in Homogeneous and Heterogeneous Swarms
Abstract:
In this work, I present a novel approach to symbolic regression tailored specifically for modeling both homogeneous and heterogeneous swarm behaviors. Swarm behaviors, as observed in animals like ants, bees, and birds, are a significant source of inspiration in research and engineering. However, constructing symbolic models from observational swarm data remains challenging due to the complexity of the interactions involved. Existing methods for simplifying collective behaviors often require extensive manual testing and creativity. To address this, I introduce the first symbolic regression approach capable of supporting a wide range of interactions within a swarm. The method is composed of two phases. The first phase employs a modified Graph Neural Network (GNN), which I call the GNN Multiplexer, to learn and capture the unique interactions within the swarm. This phase generates black-box neural networks that approximate the relationships between any two individuals in the swarm. In the second phase, a modified nested evolutionary algorithm, Macro-Micro Evolution, leverages the data generated by the GNN Multiplexer to generate simple symbolic models that approximate these relationships and are more easily interpretable by humans. I validate this method through a series of case studies, including lattice formation experiments, flocking simulations, and real-world fish schooling data, assessing the performance of both the overall approach and its individual components.
Advisor: Professor Carlo Pinciroli, Robotics Engineering, WPI
Committee: Professor William Michalson, Robotics Engineering, WPI; Professor Kevin Leahy, Robotics Engineering, WPI; and Eliseo Ferrante, Ph.D., New York University Abu Dhabi
Virtual Only: https://wpi.zoom.us/j/5123421638