Exploring Deep Reinforcement Learning for Continuous Action Control

Trantas, Athanasios (2021) Exploring Deep Reinforcement Learning for Continuous Action Control. Master's Thesis / Essay, Artificial Intelligence.

Preview	Text mAI_2021_AthanasiosTrantas.pdf Download (12MB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (130kB)

Abstract

Reinforcement Learning is a learning framework or toolbox with mathematically proven computational tools and methods to understand, calculate and automate goal-directed learning and decision making. Reinforcement Learning has been involved in some of the most remarkable developments in Artificial Intelligence mostly “Deep Reinforcement Learning”; reinforcement learning with function approximation by deep artificial neural networks. In this thesis, Deep Reinforcement Learning for continuous action control is investigated. In contrast with discrete action spaces, having unlimited actions is scary but challenging in parallel. In this study, action selection derives from a deterministic policy. For this reason, a model-free, off-policy, actor-critic algorithm named Sample Policy Gradient is extended and benchmarked. Specifically, a prioritize buffer is used to store the experience and a regularizing term is added to its objective function. Extensive experimentation is analyzed, using two simulators to test and evaluate the performance of the algorithm. To wrap up this part, a performance study and a qualitative comparison with the state-of-the-art algorithms are presented. Additionally, an important aspect of Reinforcement Learning namely Safe Reinforcement Learning is investigated. To unlock the full potential of Reinforcement Learning and apply it to daily life, it is necessary to embed the agents into the data generation distribution, which is the real-world experience.

Item Type:	Thesis (Master's Thesis / Essay)
Supervisor name:	Mohades Kasaei, S.H. and Verheij, H.B.
Degree programme:	Artificial Intelligence
Thesis type:	Master's Thesis / Essay
Language:	English
Date Deposited:	01 Dec 2021 12:18
Last Modified:	01 Dec 2021 12:18
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/26331

Actions (login required)

View Item