Active learning-based model predictive control for multi-agent coordination 

Andrea Carron

In recent years, the combination of learning-based techniques and model predictive control has received considerable attention from the control and learning communities due to the capabilities of handling constrained nonlinear systems and improving performance using data. In this talk, we present an application of learning-based model predictive control to coverage control in unknown environments. In coverage control, a fleet of agents aims 
at "optimally" partitioning an area, also called environment, with respect to environmental demands. Those demands are reflected by a utility function of measurable values of interest. When the utility function is unknown, it is necessary to trade off the exploration of the environment against the exploitation of the acquired knowledge. To tackle this problem, we present two active learning-based model predictive control schemes that are guaranteed to converge to an optimal configuration. The first is based on a hierarchical framework, where references that trade-off 
exploration and exploitation are computed by a reference generator and then tracked by the agents' local model predictive control (MPC) tracking schemes. The second control scheme avoids the hierarchical structure by integrating reference optimization and active learning in the MPC formulation. Finally, we demonstrate the effectiveness of the proposed algorithms on a robotic platform composed of miniature RC race cars.

To the top of the page