Machine Learning for Cooperative Logistics: Multi-Agent Reinforcement Learning to Negotiate Fair Outcomes
Year: 2024
Type: dissertation
Abstract: Effectively managing businesses is perhaps the ultimate testbed for AI. It requires creativity. It requires reasoning. It requires intelligence. Logistics forms the backbone of our economy. From the food we eat, to the computers we work on, to the planes we fly on – logistics has played a central role. Yet, profit margins for logistics providers are thin, and customers have ever-increasing expectations. Worse, humanity is not on track to decarbonise road freight transportation. One effective method to alleviate this pressure is Cooperative Logistics. Here, companies cooperate by pooling their resources together (such as trucks, drivers, and/or warehouses) to improve their individual performance (such as profitability and carbon emissions). But if companies cooperate, how much should everyone be fairly paid? This is known as fair gain sharing. Prior methods do not scale because they require evaluating O(2n) NP-hard vehicle routing problems (VRPs) for n companies – limiting practical cooperation to ~20 companies – forgoing significant economies of scale. Scaling existing methods to 1,000 companies would therefore require solving 1 × 10^301 NP-hard VRPs. We propose MAGS (Multi-Agent Gain Sharing) to enable fair gain sharing at scale. MAGS effectively handles over 1,000 agents using a principled, game-theoretic approach. This is achieved by training deep multi-agent reinforcement learning agents to negotiate fair gain shares. This is challenging due to the n-player, mixed-motive environment with agents learning concurrently. Crucially, we eliminate reliance on the characteristic function. We additionally leverage transformers to design a reward model for coalition structure generation. We show that in a real-world setting with 1,000 carriers, we can potentially reduce carbon emissions by 22.1%1, whilst improving average profitability from £66.71 per day to £259.26 per day. This is particularly important as global freight activity is expected to double by 2050. Furthermore, this thesis lays the principled foundations for Agentic AI and Human-AI natural language negotiation in future work. This thesis therefore contributes to both the artificial intelligence (AI) and supply chain management (SCM) literature.
Source: Apollo (University of Cambridge)
DOI: 10.17863/cam.128509
Author: Mak, Stephen
Language: English
Subfield: Economics and Econometrics
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7147016666