Abstract
MORL is a subfield of traditional Reinforcement Learning that solves problems with multiple, often conflicting objectives. MORL algorithms are run in environments that pose certain objectives and produce a set of optimal solutions for the problem that represent different preferences (or balancing of objectives), called the Pareto Front (PF). WE measure the PF quality using multiple metrics. In order to apply a MORL algorithm to solve a real-world problem, one needs to balance the PF quality and the compute time. In this project we test Envelope Q-Learning, GPI-LS, and D3P0 algorithms in the Minecart and MO-Highway environments. We explore how well these algorithms perform in environments with real-world characteristic under different conditions. We design experiments to test generalisation, action-space complexity, and objective dependency. These results will contribute to informing the optimal selection of MORL algorithms for solving real-world problems.
Videos 1
Watch presentations, demos, and related content
MORL Project Presentation Video
Like, comment, and subscribe on YouTube to support the creator!
MORL Project Presentation Video
Gallery 2
Explore the visual story of this exhibit
Title Image
Project Poster