De Lellis, Francesco (2023) Reinforcement Learning for Control. [Tesi di dottorato]

Anteprima

Testo De_Lellis_Francesco_35.pdf Download (10MB) | Anteprima

Abstract

In this thesis, we go through the current state of the art reinforcement learning for control applications. We analyze the pros and cons of the methods provided in the literature. Then we establish a common frame�work to describe both reinforcement learning and control problems and we present four benchmark problems to analyze and compare reinforce�ment learning algorithms. Also, we propose a novel solution, the minimal performance Q-learning, capable of searching and guaranteeing a solution that meets a desired level of performance in terms of steady-state error and settling time. Moreover, we also present the control tutored reinforce�ment learning, an architecture where a feedback controller derived from an approximate model of the environment assists the learning process to enhance its data efficiency. We apply this idea to the classical Q-learning, in the form of a deterministic Control-Tutored Q-Learning (CTQL), that defines the reward function so that a Boolean condition can be used to determine when the control tutor policy is adopted. We also introduce a probabilistic CTQL (pCTQL) that is instead based on executing calls to the tutor with a certain probability during learning. Moreover, we also develop a control tutored deep reinforcement learning the CT-DQN. All the strategies proposed in this thesis are thoroughly analyzed and com�pared with the literature via the set of metrics to evaluate learning and control performances. Eventually, we discuss how control techniques can be applied to face the global COVID-19 pandemic.

Downloads

Actions (login required)

Modifica documento