TU Darmstadt / ULB / TUprints

Long-Term Visitation Value for Deep Exploration in Sparse-Reward Reinforcement Learning

Parisi, Simone ; Tateo, Davide ; Hensel, Maximilian ; D’Eramo, Carlo ; Peters, Jan ; Pajarinen, Joni (2022):
Long-Term Visitation Value for Deep Exploration in Sparse-Reward Reinforcement Learning. (Publisher's Version)
In: Algorithms, 15 (3), MDPI, e-ISSN 1999-4893,
DOI: 10.26083/tuprints-00021017,
[Article]

[img] Text
algorithms-15-00081-v3.pdf
Available under: CC BY 4.0 International - Creative Commons, Attribution.

Download (15MB)
Item Type: Article
Origin: Secondary publication DeepGreen
Status: Publisher's Version
Title: Long-Term Visitation Value for Deep Exploration in Sparse-Reward Reinforcement Learning
Language: English
Abstract:

Reinforcement learning with sparse rewards is still an open challenge. Classic methods rely on getting feedback via extrinsic rewards to train the agent, and in situations where this occurs very rarely the agent learns slowly or cannot learn at all. Similarly, if the agent receives also rewards that create suboptimal modes of the objective function, it will likely prematurely stop exploring. More recent methods add auxiliary intrinsic rewards to encourage exploration. However, auxiliary rewards lead to a non-stationary target for the Q-function. In this paper, we present a novel approach that (1) plans exploration actions far into the future by using a long-term visitation count, and (2) decouples exploration and exploitation by learning a separate function assessing the exploration value of the actions. Contrary to existing methods that use models of reward and dynamics, our approach is off-policy and model-free. We further propose new tabular environments for benchmarking exploration in reinforcement learning. Empirical results on classic and novel benchmarks show that the proposed approach outperforms existing methods in environments with sparse rewards, especially in the presence of rewards that create suboptimal modes of the objective function. Results also suggest that our approach scales gracefully with the size of the environment.

Journal or Publication Title: Algorithms
Volume of the journal: 15
Issue Number: 3
Publisher: MDPI
Collation: 44 Seiten
Uncontrolled Keywords: reinforcement learning, sparse reward, exploration, upper confidence bound, off-policy
Classification DDC: 000 Allgemeines, Informatik, Informationswissenschaft > 004 Informatik
600 Technik, Medizin, angewandte Wissenschaften > 620 Ingenieurwissenschaften
Divisions: 20 Department of Computer Science > Intelligent Autonomous Systems
Date Deposited: 11 Apr 2022 11:11
Last Modified: 11 Apr 2022 11:12
DOI: 10.26083/tuprints-00021017
Corresponding Links:
URN: urn:nbn:de:tuda-tuprints-210175
SWORD Depositor: Deep Green
URI: https://tuprints.ulb.tu-darmstadt.de/id/eprint/21017
PPN:
Export:
Actions (login required)
View Item View Item