Abstract

Humans and animals solve a difficult problem much more easily when they are presented with a sequence of problems that starts simple and slowly increases in difficulty. We explore this idea in the context of reinforcement learning. Rather than providing the agent with an externally provided curriculum of progressively more difficult tasks, the agent solves a single task utilizing a decreasingly constrained policy space. The algorithm we propose first learns to categorize features into positive and negative before gradually learning a more refined policy. Experimental results in Tetris demonstrate superior learning rate of our approach when compared to existing algorithms.
Original languageEnglish
Publication statusPublished - 2019
EventNeurIPS 2019 Workshop on Biological and Artificial Reinforcement Learning - Vancouver, Canada
Duration: 13 Dec 2019 → …
https://sites.google.com/view/biologicalandartificialrl/

Workshop

WorkshopNeurIPS 2019 Workshop on Biological and Artificial Reinforcement Learning
CountryCanada
CityVancouver
Period13/12/19 → …
Internet address

Keywords

  • Reinforcement learning
  • Human decision making
  • Tetris

Cite this

Lichtenberg, J., & Şimşek, Ö. (2019). Iterative Policy-Space Expansion in Reinforcement Learning. Poster session presented at NeurIPS 2019 Workshop on Biological and Artificial Reinforcement Learning, Vancouver, Canada.

Iterative Policy-Space Expansion in Reinforcement Learning. / Lichtenberg, Jan; Şimşek, Özgür.

2019. Poster session presented at NeurIPS 2019 Workshop on Biological and Artificial Reinforcement Learning, Vancouver, Canada.

Research output: Contribution to conferencePoster

Lichtenberg, J & Şimşek, Ö 2019, 'Iterative Policy-Space Expansion in Reinforcement Learning', NeurIPS 2019 Workshop on Biological and Artificial Reinforcement Learning, Vancouver, Canada, 13/12/19.
Lichtenberg J, Şimşek Ö. Iterative Policy-Space Expansion in Reinforcement Learning. 2019. Poster session presented at NeurIPS 2019 Workshop on Biological and Artificial Reinforcement Learning, Vancouver, Canada.
Lichtenberg, Jan ; Şimşek, Özgür. / Iterative Policy-Space Expansion in Reinforcement Learning. Poster session presented at NeurIPS 2019 Workshop on Biological and Artificial Reinforcement Learning, Vancouver, Canada.
@conference{40ddf1c4e3bb4fa4aae9c1911c35165e,
title = "Iterative Policy-Space Expansion in Reinforcement Learning",
abstract = "Humans and animals solve a difficult problem much more easily when they are presented with a sequence of problems that starts simple and slowly increases in difficulty. We explore this idea in the context of reinforcement learning. Rather than providing the agent with an externally provided curriculum of progressively more difficult tasks, the agent solves a single task utilizing a decreasingly constrained policy space. The algorithm we propose first learns to categorize features into positive and negative before gradually learning a more refined policy. Experimental results in Tetris demonstrate superior learning rate of our approach when compared to existing algorithms.",
keywords = "Reinforcement learning, Human decision making, Tetris",
author = "Jan Lichtenberg and {\"O}zg{\"u}r Şimşek",
year = "2019",
language = "English",
note = "NeurIPS 2019 Workshop on Biological and Artificial Reinforcement Learning ; Conference date: 13-12-2019",
url = "https://sites.google.com/view/biologicalandartificialrl/",

}

TY - CONF

T1 - Iterative Policy-Space Expansion in Reinforcement Learning

AU - Lichtenberg, Jan

AU - Şimşek, Özgür

PY - 2019

Y1 - 2019

N2 - Humans and animals solve a difficult problem much more easily when they are presented with a sequence of problems that starts simple and slowly increases in difficulty. We explore this idea in the context of reinforcement learning. Rather than providing the agent with an externally provided curriculum of progressively more difficult tasks, the agent solves a single task utilizing a decreasingly constrained policy space. The algorithm we propose first learns to categorize features into positive and negative before gradually learning a more refined policy. Experimental results in Tetris demonstrate superior learning rate of our approach when compared to existing algorithms.

AB - Humans and animals solve a difficult problem much more easily when they are presented with a sequence of problems that starts simple and slowly increases in difficulty. We explore this idea in the context of reinforcement learning. Rather than providing the agent with an externally provided curriculum of progressively more difficult tasks, the agent solves a single task utilizing a decreasingly constrained policy space. The algorithm we propose first learns to categorize features into positive and negative before gradually learning a more refined policy. Experimental results in Tetris demonstrate superior learning rate of our approach when compared to existing algorithms.

KW - Reinforcement learning

KW - Human decision making

KW - Tetris

M3 - Poster

ER -