Funded by the Australian Research Council Discovery Grant
Advancing stochastic optimisation: highly-correlated restless bandit models,
Professor Peter Taylor, Dr Jing Fu and Professor Jose Niño-Mora would like to advertise two fully-funded PhD projects, one to be held at The Royal Melbourne Institute of Technology (RMIT) and the other to be held at The University of Melbourne, Australia.
Project 1 (RMIT, Australia)
Title: Restless-Bandit-Enhanced Multi-Agent Reinforcement Learning
Project Description: Restless-Bandit-Enhanced Learning (RB-L) is an emerging framework that integrates restless bandit theorems with reinforcement learning to tackle the curse of dimensionality. It is widely applicable in realistic scenarios. This project aims to trade off learning and control in practical scenarios with inevitable high-dimensional state and/or action spaces. It will incorporate advanced control and learning algorithms, such as classic multi-agent reinforcement learning. The expected outcomes are scalable learning and control algorithms, with rational guarantees of overall performance.
The prospective HDR student is expected to have fundamental knowledge of stochastic modeling, reinforcement learning, linear/convex optimization, and good programming skills for large-scale simulations.
Project 2 (University of Melbourne, Australia)
Title: Optimality in highly-correlated restless bandit models
Project Description: Conventional techniques for analyzing restless bandit models such as Whittle relaxation, fluid approximation, and LP-based approximation focus on proving asymptotic optimality by exploring levels of relaxation over the original optimization problem.
A solution of a relaxed problem can reflect intrinsic properties. It can be utilized to propose a heuristic policy for the original problem. Often it is possible to show that such a policy is asymptotically optimal. The objective of this project is to extend these methods to highly-correlated restless bandit models.
The project will also consider how well the solutions perform in the non-asymptotic regime.
.
The prospective HDR student is expected to have fundamental knowledge of stochastic modeling, probability, reinforcement learning, and linear/convex optimization, and good programming skills.
In the first instance, interested applicants should email Dr Jing Fu at jing.fu@rmit.edu.au, Professor Peter Taylor at taylorpg@unimelb.edu.au and Professor Jose Niño-Mora at jnino@est-econ.uc3m.es explaining why they are interested in the projects and providing details of their curriculum vitae and their academic record.