Energy Landscapes and Data-Driven Control for Multi-agent Autonomous Systems
Abstract
Available for Public Release. Multi-agent systems, including swarms of mobile agents, are known to exhibit complex spatio-temporal coherent structures even when agents follow relatively simple rules. Such complexity has been leveraged to design swarms that complete tasks, such as area coverage and surveillance, large-scale sensing and estimation, building complex environmental maps and models, and using one swarm to defend against another. One major problem in both natural and robotic swarms is that heterogeneity among agents, communication latency, and nonlinear interactions create multiple coherent states, many of which are observable. Due to multi-stable dynamics, uncertain environments play a critical role in causing switching between coherent patterns, resulting in random behavior that may hinder state-space controls for maintaining desired coherent states. Such unpredictability requires a deeper understanding of how noise-induced switching occurs, how complex basins of attraction for autonomous systems interact with uncertain environments, and how novel controls may be implemented to stabilize desired multi-agent system behaviors or destabilize undesirable states.Designing and controlling multi-agent systems interacting with uncertain environments requires more than the usual local analytic techniques. For example, to understand how random switching occurs between states of dynamical patterns necessitates a global pictureof an autonomous system s state space coupled with the environment and parametric uncertainty. Since stochastic multi-agent systemsare non-equilibrium systems, their dynamics are often determined by two key factors: the underlying energy landscape and, importantly, a probability flux. The landscape and flux theory has many advantages for multi-agent and autonomous systems. In particular, it:(1) describes switching paths in large noise regimes, (2) generates non-equilibrium transition state theory, (3) provides additional flux contributionsfor maintaining stochastic steady states (defined as almost invariant sets (AIS)), (4) gives a framework for controlling the global stability of multi-agent stochastic dynamics such as colliding and interacting swarms, and (5) can be learned from data generated models. The main focus of this proposal is to develop both theory and machine-learning techniques for energy landscape analysis applied to the prediction and control of complex dynamical patterns of stochastic multi-agent systems.Since the stochastic dynamics of multi-agent systems are very high dimensional, state-space control can be challenging and prohibitively expensive to design. In contrast, an energy landscape theory can provide lower dimensional parametric control strategies to achieve desired coherent states. Moreover, in the stochastic setting, attractors in an energy landscape are rendered as AIS where the dwell time is long compared to the relaxation times. Given the existence of AIS, we can stabilize such sets using control techniques related directlyto the topology of the energy landscape [20]. Recent developments employing data-driven modeling allows us to predict the AIS by leveraging Koopman transfer operator theory, as well as stochastic controls that extend the dwell times of desirable sets. The development of data-driven techniques to predict and control AIS dwell times and switching mechanisms in autonomous multi-agent systems is another important thrust of this proposal. Going beyond characterizing AIS, we will solve the problem of constructing the energy landscape and flux transport from data (including from sparse data), and design network interaction topology and controls for multi-agent systems that enable dynamic prediction and parametric control on an energy landscape. Finally, we will coordinate our research toleverage data from the NRL 6.2 experimental program "Swarm Behavior Prediction and Disruption Utilizing Antagonistic Agents".
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Mar 12, 2025
- Source ID
- N000142512171
Entities
People
- Mong-ying Hsieh
Organizations
- Office of Naval Research
- United States Navy
- University of Pennsylvania