Model Predictive Control and Reinforcement Learning: a unifying approach

Abstract

The aim of this thesis is to propose new insights in the field of Model Predictive Control (MPC) combined with Reinforcement Learning (RL) algorithms, in the following research contexts: • Set-theoretic Receding Horizon Control for unknown dynamical systems. The proposed methodology is based on three elements: Linear Time-Invariant system behaviour; Data-Driven modelling, and Reinforcement Learning technicalities. These are appropriately combined in order to derive an accurate outer convex approximation of the nonlinear evolution of an unknown system. The effectiveness of the obtained uncertain polytopic model has been tested using a probabilistic approach. • Platoons on Urban Road Networks. Here, the main goals are to alleviate traffic congestion and reduce the CO2 emissions of autonomous vehicles. The core of this part of the project is the formal definition of new strategies, comprehensive of routing algorithms and distributed control systems within a theoretical formulation. To efficiently evaluate the performance of the proposed control architecture, a co-simulation between SUMO (Simulation of Urban MObility) and MATLAB has been implemented. • An intelligent multi-layer control architecture for logistics operations of autonomous vehicles in manufacturing systems. Similar to the case of autonomous vehicles on Urban Road Networks, Reinforcement Learning and Model Predictive Control are combined to solve the problem of task scheduling, routing and control. • Autonomous vehicle platoons subject to cyber-attacks. The core of this problem is how to detect and to deal with false data injections in the field of intelligent transportation systems.