Concurrent Learning and Control of Uncertain Large-Scale Phenomena

Pollution has been a major problem for centuries, leading to rising global temperatures, public health crises, and ecological devastation. Control and monitoring of pollution-related phenomena are of utmost priority for governments, stakeholders, and business. Many of these events exhibit spatio-temporal dynamics (XTD) described by partial differential equations. Compelling examples include wildfires, greenhouse gas propagation, and water contaminant dispersion, where swift and strategic intervention is critical. Effective intervention entails quickly understanding the phenomenon and making timely and strategic decisions. We refer to such problems as "concurrent learning and intervention" (CLI), where timely intervention can be the difference between success and failure, lives saved, and preventing further harm. Effective intervention requires a deep understanding of XTD phenomena and timely, strategic decisions. This project seeks innovative approaches to estimate, predict, and control these phenomena, striking a balance between data acquisition and intervention. While reinforcement learning-based methods can synthesize strategic intervention policies directly from data, they demand extensive data and long training times. Furthermore, they often disregard existing physics-based models that have been well-studied for many phenomena. On the other hand, recent advancements in controlling uncertain environments offer a promising framework for CLI. However, this calls for systematic approaches for computationally efficient data-driven modelling and control techniques. This project aims to formulate innovative identification, estimation, and control methods for networks of autonomous agents and people to achieve an optimal trade-off between (1) acquiring samples from the environment to learn the dynamics, and (2) interacting and modifying the environment to satisfy high-level requirements in a timely fashion. To achieve our aim, we have set the following objectives: Develop novel rapid modelling and estimation techniques based on physics-informed machine learning approaches, in which the nominal physics-based dynamical model is refined by a computationally efficient set membership-based data-driven model. This will help reduce the gap between theoretical models and real-world dynamics. Create a new control approach, called Dual Control for Exploration and Interaction (DCEI), based on state-of-the-art model predictive control methods that integrate known or learned phenomenon models with data-based methods. This method will empower agents to address environmental uncertainties and take actions in a timely manner. Build software tools for practical implementation of the developed methods, such that they aredeployable in real-time embedded hardware. Integrate and test the new methods in two case studies: wildfire emergency response and air pollutant monitoring. The project involves a unique multi-disciplinary collaboration among experts in control sciences, optimization, fire sciences, complex dynamics, and aeronautics, both from the academia and industry. The collaboration with our industrial partners, Andrew Moore & Associates and Flylogix, will ensure that our innovative methodologies can be readily applied in real-world settings. This project addresses a pressing global issue related to pollution-related phenomena. The resulting methods will provide a significant advantage to tactical response teams, enabling better decision-making, faster response times and reduced environmental impacts. By advancing control in unknown environments, we will contribute to machine learning, control sciences, and robotics research communities. This project's framework, initially designed for fire and pollution control, can be adapted for various applications, including social dynamics, communication networks, and drug delivery.