Line Routing Optimization of Branched Pipe Systems Using Multi-Agent Reinforcement Learning
Loading...
Authors
Teoli, Mira Abigail
Date
2024-09-30
Type
thesis
Language
eng
Keyword
Reinforcement Learning , Pipe Routing , Line Routing Optimization
Alternative Title
Abstract
The problem of determining the optimal path between two points is fundamental to the design of interconnected systems. Hydraulic piping, electrical wiring, and engine room layouts are just some examples in which the quality and efficiency of path routing are critical to design success. This process, known as Line Routing Optimization (LRO), is often performed manually by experts, which can be time-consuming and may not fully capture the breadth of available possibilities. There has been extensive research into solving this challenge, with automatic routing methods established based on traditional graph search methods, genetic algorithms, and reinforcement learning. Much of this existing work focuses on the shortest path length and bend-minimizing constraints, disregarding other important objectives, such as branching. Those who do consider these additional objectives determine paths sequentially or by establishing groups, neglecting conflicting objectives of individual routes, resulting in solutions highly dependent on the routing order.
Multi-agent reinforcement learning (MARL) is a subfield of reinforcement learning where multiple agents learn to optimize their behaviour based on individual and collective experiences. This work utilizes MARL to concurrently determine line routes, where each line is represented by its own agent, allowing for optimization of both individual paths and the overall system. The overarching objective of this work is to use RL to perform automatic pipe routing, thereby determining the optimal layout of pipes within a system. By considering design objectives like branching, pipe routes that are more cost-effective and simpler to install can be discovered. The results of this work demonstrate that the use of MARL for LRO shows promise in efficiently designing interconnected systems. A reduction in total pipe length by up to 34% is seen when the branching objective is considered. A discussion of the potential for future work and impact is included in the conclusions, with a mention of refining the MARL model, expanding adaptability to changing obstacle locations and considering additional objectives, such as alignment, head loss, and stress.