# Resource Abstraction for Reinforcement Learning in Multiagent Congestion Problems

**Authors:** Kleanthis Malialis, Sam Devlin, Daniel Kudenko  
**arXiv:** [1903.05431](https://arxiv.org/abs/1903.05431)  
**Date:** 2019-03-13  
**Summary page:** https://www.summarizepaper.com/en/paper/1903.05431v1/

## Key points

- Real-world congestion problems are complex and require effective solutions
- Multiagent reinforcement learning (MARL) is a promising approach to tackle these challenges
- Deployability of MARL approaches to congestion problems is limited by learning time, scalability, and decentralised coordination
- Resource Abstraction is introduced as an approach that addresses these challenges by allocating available resources into abstract groups
- Resource Abstraction creates new reward functions that provide a more informative signal to the learning agents and aid coordination among them
- Experimental work on two benchmark domains from the literature shows that using Resource Abstraction significantly improves learning speed and scalability while achieving high joint performance/social welfare for both congestion problems in large-scale scenarios involving up to 1000 reinforcement learning agents
- Resource Abstraction overcomes some of the critical limitations of MARL approaches to congestion problems by providing more informative signals to learning agents through abstraction, enabling better coordination among them without requiring communication between them.
- This approach could have significant implications for real-world applications where decentralised coordination is essential but challenging to achieve due to scale or complexity constraints.

## Summary

Real-world congestion problems, such as traffic congestion, are complex and large-scale issues that require effective solutions. Multiagent reinforcement learning (MARL) has emerged as a promising approach to tackle these challenges by providing an autonomous and distributed solution. However, the deployability of MARL approaches to congestion problems is limited by three factors: learning time, scalability, and decentralised coordination (i.e., no communication between the learning agents). In this paper titled "Resource Abstraction for Reinforcement Learning in Multiagent Congestion Problems," Kleanthis Malialis, Sam Devlin, and Daniel Kudenko introduce Resource Abstraction as an approach that addresses these challenges by allocating available resources into abstract groups. Resource Abstraction creates new reward functions that provide a more informative signal to the learning agents and aid coordination among them. The authors conduct experimental work on two benchmark domains from the literature: an abstract congestion problem and a realistic traffic congestion problem. The current state-of-the-art for solving multiagent congestion problems is a form of reward shaping called difference rewards. However, the authors show that using Resource Abstraction significantly improves learning speed and scalability while achieving the highest possible or near-highest joint performance/social welfare for both congestion problems in large-scale scenarios involving up to 1000 reinforcement learning agents. The paper's findings suggest that Resource Abstraction can overcome some of the critical limitations of MARL approaches to congestion problems. By providing more informative signals to learning agents through abstraction, it enables better coordination among them without requiring communication between them. This approach could have significant implications for real-world applications where decentralised coordination is essential but challenging to achieve due to scale or complexity constraints. Overall, this paper presents an innovative solution that could pave the way for more effective MARL approaches in addressing complex real-world problems like traffic congestion.