Reinforcement learning approach to multi-stage decision making problems with changes in action sets

Authors:
Takuya Etoh;Hirotaka Takano;Junichi Murata
Affiliations:
Kyushu University, Nishi-ku, Japan;Kyushu University, Nishi-ku, Japan;Kyushu University, Nishi-ku, Japan
Venue:
Artificial Life and Robotics
Year:
2012

Citing 5
Cited 0

Introduction to Reinforcement Learning

Introduction to Reinforcement Learning
Variance-Penalized Reinforcement Learning for Risk-Averse Asset Allocation

IDEAL '00 Proceedings of the Second International Conference on Intelligent Data Engineering and Automated Learning, Data Mining, Financial Engineering, and Intelligent Agents
Handbook of Learning and Approximate Dynamic Programming (IEEE Press Series on Computational Intelligence)

Handbook of Learning and Approximate Dynamic Programming (IEEE Press Series on Computational Intelligence)
Motivated Learning from Interesting Events: Adaptive, Multitask Learning Agents for Complex Environments

Adaptive Behavior - Animals, Animats, Software Agents, Robots, Adaptive Systems
Risk-sensitive reinforcement learning applied to control under constraints

Journal of Artificial Intelligence Research

Quantified Score

Hi-index	0.00

Visualization

Abstract

Multi-stage decision making (MSDM) problems often include changes in practical situations. For example, in the shortest route selection problems in road networks, travelling times of road sections vary depending on traffic conditions. The changes give rise to risks in adopting particular solutions to MSDM problems. Therefore, a method is proposed in this paper for solving MSDM problems considering the risks. Reinforcement learning (RL) is adopted as a method for solving those problems, and stochastic changes of action sets are treated. It is necessary to evaluate risks based on subjective views of decision makers (DMs) because the risk evaluation is by nature subjective and depends on DMs. Therefore, an RL approach is proposed which uses a new method for evaluating risks of the changes that can easily incorporate the DM's subjective view and can be readily imbedded in reinforcement learning algorithms. The effectiveness of the method is illustrated with a road network path selection problem.