Reinforcement learning for sequential decision making with constraints

Reinforcement learning is a widely used approach to tackle problems in sequential decision making where an agent learns from rewards or penalties. However, in decision-making problems that involve safety or limited resources, the agent's exploration is often limited by constraints. To model suc...

Full description

Saved in:
Bibliographic Details
Main Author: LING, Jiajing
Format: text
Language:English
Published: Institutional Knowledge at Singapore Management University 2023
Subjects:
Online Access:https://ink.library.smu.edu.sg/etd_coll/513
https://ink.library.smu.edu.sg/context/etd_coll/article/1511/viewcontent/GPIS_AY2018_PhD_LING_Jiajing.pdf
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Singapore Management University
Language: English
Description
Summary:Reinforcement learning is a widely used approach to tackle problems in sequential decision making where an agent learns from rewards or penalties. However, in decision-making problems that involve safety or limited resources, the agent's exploration is often limited by constraints. To model such problems, constrained Markov decision processes and constrained decentralized partially observable Markov decision processes have been proposed for single-agent and multi-agent settings, respectively. A significant challenge in solving constrained Dec-POMDP is determining the contribution of each agent to the primary objective and constraint violations. To address this issue, we propose a fictitious play-based method that uses Lagrangian Relaxation to perform credit assignment for both primary objectives and constraints in large-scale multi-agent systems. Another major challenge in solving both CMDP and constrained Dec-POMDP is the sample inefficiency issue, mainly resulting from finding valid actions that satisfy all constraints, which becomes even more difficult in large state and action spaces. Recent works in RL have attempted to incorporate domain knowledge from experts into the learning process through neuro-symbolic methods to address the sample inefficiency issue. We propose a knowledge compilation framework using decision diagrams by treating constraints as domain knowledge and introducing neuro-symbolic methods to support effective learning in constrained RL. Firstly, we propose a zone-based multi-agent pathfinding (ZBPF) framework that is motivated by drone delivery applications. We propose a neuro-symbolic method to efficiently solve the ZBPF problem with several domain constraints, such as simple path constraint and landmark constraint in ZBPF. Secondly, we propose another neuro-symbolic method to solve action constrained RL where the action space is discrete and combinatorial. Empirical results show that our proposed approaches achieve better performance than standard constrained RL algorithms in several real-world applications.