Adam Jelley

PhD Student in Deep Reinforcement Learning

University of Edinburgh

Hello!

I’m a PhD student in the Bayesian and Neural Systems group at the University of Edinburgh, supervised by Professor Amos Storkey in the School of Informatics and externally by Sam Devlin at Microsoft Research Cambridge.

My research is broadly focused on developing more sample efficient reinforcement learning approaches, via the use of self-supervised signals, offline data, world models and human feedback. I’m generally interested in improving the scientific understanding of deep reinforcement learning, bridging the gap between academia and industry, and the application of reinforcement learning to production software and real-world problems. I’m supported by a Microsoft Research Scholarship.

Interests

Reinforcement Learning
World Models
RL from Human Feedback
Self-Supervised Learning
Deep Learning Generalisation
AI for Science

Education

Currently PhD Candidate in Sample Efficient Deep Reinforcement Learning, 2025 (Expected)
University of Edinburgh
MSci in Theoretical Physics, 2016
University of Cambridge
MA in Natural Sciences - Physics, 2015
University of Cambridge

Selected Publications

Diffusion for World Modeling: Visual Details Matter in Atari

We introduce DIAMOND, an reinforcement learning agent trained in a diffusion world model.

Eloi Alonso, Adam Jelley, Vincent Micheli, Anssi Kanervisto, Amos Storkey, Tim Pearce, François Fleuret

Diffusion for World Modeling: Visual Details Matter in Atari

Aligning Agents like Large Language Models

An investigation into training agents like Large Language Models (LLMs) by unsupervised pre-training, supervised fine-tuning, and finally reinforcement learning from human feedback (RLHF).

Adam Jelley, Yuhan Cao, Dave Bignell, Sam Devlin, Tabish Rashid

Aligning Agents like Large Language Models

Efficient Offline Reinforcement Learning: The Critic is Critical

An approach for efficient offline reinforcement learning by first learning the behaviour policy and values with supervised learning, before improving on this policy with reinforcement learning.

Adam Jelley, Trevor McInroe, Sam Devlin, Amos Storkey

Efficient Offline Reinforcement Learning: The Critic is Critical

Contrastive Meta-Learning for Partially Observable Few-Shot Learning

An approach for meta-learning contrastive representations under partial observability. We demonstrate this approach can be utilised by reinforcement learning agents to learn a representation of their environment.

Adam Jelley, Amos Storkey, Antreas Antoniou, Sam Devlin

Contrastive Meta-Learning for Partially Observable Few-Shot Learning

Deep Ocean Learning of Small Scale Turbulence

We show machine learning can be successfully employed to infer turbulent mixing from quantities measured routinely by global observational programs.

Ali Mashayek, Nick Reynard, Fangming Zhai, Kaushik Srinivasan, Adam Jelley, Alberto Naveira Garabato, Colm-cille P. Caulfield

Deep Ocean Learning of Small Scale Turbulence

Experience

PhD Candidate

University of Edinburgh

May 2021 – Present Edinburgh

PhD Candidate in Machine Learning, with a focus on Deep Reinforcement Learning.

Research Scientist Intern

Microsoft Research Cambridge

Jun 2023 – Sep 2023 Cambridge

Developed a pipeline for aligning agents with preferences on the Xbox game Bleeding Edge, for research into capabilities and limitations of Reinforcement Learning from Human Feedback (RLHF) in this domain.