[AAMAS 2024] Multi-Agent Reinforcement Learning for Real-World Applications (Part 1) - LG AI Research Blog

At AAMAS 2024, the authoritative international conference in the field of multi-agent systems, LG AI Research's Data Intelligence(DI) Lab presented two papers. This post is the first part of a series on real-world applications of MARL, covering agent modeling and algorithms for applying AI technology across various industries.

Reinforcement Learning(RL) is the process by which an agent learns an optimal action policy that maximizes reward through interaction with an environment. This is modeled as a Markov Decision Process(MDP), which consists of State, Action, Reward, and Transition Probability.

Multi-Agent Reinforcement Learning(MARL) is the process in which multiple agents exist within an environment and each agent learns its own policy. Each agent has independently defined states and actions, and finds its optimal policy while interacting with other agents.

As a key example, the post introduces Harvard University's research on 'Surge Routing'. This research proposes a MARL-based On-demand Mobility Routing algorithm to improve the efficiency of autonomous taxi services by responding to sudden demand changes caused by large-scale events and similar factors.