Deep reinforcement learning : a study of the CartPole problem

Budai, Ádám and Csorba, Kristóf: Deep reinforcement learning : a study of the CartPole problem. Conference of PhD Students in Computer Science, (11). pp. 17-20. (2018)

[img] Cikk, tanulmány, mű
cscs_2018_030-033.pdf

Download (454kB)

Abstract

One of the major challenges of artificial intelligence is to learn solving tasks which are considered to be challenging for even a human. Reinforcement learning is the most general learning framework among the three main type of learning methods: supervised, unsupervised and reinforcement learning. Most of the problems can easily fit into this framework. Experience shows that a lot of machine learning methods with non-linear function approximators suffers from instability regarding convergence. Reinforcement learning is more prone to diverge due to its ability to change the structure of its training data by modifying the way how it interacts with the environment. In this paper we investigate the divergence issue of DQN on the CartPole problem in terms of the algorithm’s parameters. Instead of the usual approach we do not focus on the successful trainings but instead we focus on the dark side where the algorithm fails on such an easy problem like CartPole. The motivation is to gain some further insight into the nature of the divergence issues on a specific problem.

Item Type: Article
Event Title: Conference of PhD students in computer science (11.) (2018) (Szeged)
Journal or Publication Title: Conference of PhD Students in Computer Science
Date: 2018
Volume: 11
Page Range: pp. 17-20
Uncontrolled Keywords: Számítástechnika - előadáskivonat, Algoritmus - előadáskivonat
Additional Information: Bibliogr.: 20. p. ; összefoglalás angol nyelven
Date Deposited: 2019. Oct. 28. 10:01
Last Modified: 2019. Oct. 28. 10:01
URI: http://acta.bibl.u-szeged.hu/id/eprint/61754

Actions (login required)

View Item View Item