Reinforcement Learning Approaches in Dynamic Environments
Year: 2018
Type: dissertation
Abstract: Reinforcement learning is learning from interaction with an environment to achieve a goal. It is an efficient framework to solve sequential decision-making problems, using Markov decision processes (MDPs) as a general problem formulation. In this thesis, we apply reinforcement learning to sequential decision-making problems in dynamic environments.We first present an algorithm based on Q-learning with a customized exploration and exploitation strategy to solve a real taxi routing problem. Our algorithm is able to progressively learn optimal actions for routing an autonomous taxi to passenger pick-up points. Then, we address the factored MDP problem in a non-deterministic setting. We propose an algorithm that learns transition functions using the Dynamic Bayesian Network formalism. We demonstrate that factorization methods allow to efficiently learn correct models; through the learned models, the agent can accrue higher cumulative rewards.We extend our work to very large domains. In the focused crawling problem, we propose a new scoring mechanism taking into account long-term effects of selecting a link, and present new feature representations of states for Web pages and actions for next link selection. This approach allowed us to improve on the efficiency of focused crawling. In the influence maximization (IM) problem, we extend the classical IM problem with incomplete knowledge of graph structure and topic-based user interest. Our algorithm finds the most influential seeds to maximize topic-based influence by learning action values for each probed node.
Source: HAL
Author: Han, Miyoung
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2939567823
