Your Name

Daniel Beechey

PhD Student

Department of Computer Science, University of Bath, United Kingdom

djeb20 AT bath DOT ac DOT uk

I am a PhD student at the Bath Reinforcement Learning Laboratory, University of Bath, supervised by Özgür Şimşek. I am also a member of the Centre for Doctoral Training in Accountable, Responsible and Transparent AI (ART-AI). My research interests lie in artificial intelligence, with a focus on reinforcement learning.

Research Summary: I am interested in the capacity of AI systems with limited resources to acquire generalisable behaviours that facilitate continual learning. Specifically, I focus on two main areas: discovering broad and generalised skills that allow for continual adaption to change and explaining the behaviour of these agents by revealing how they are influenced by their observations. For further details, please refer to the research section below.

News

Research

My research aims to (1) uncover general skills for agents operating under contrained resources, enabling continual adaptation to new and diverse tasks, and (2) explain the behaviour of agents in terms of policy and performance.

Explaining Agent Behaviour

An agent's behaviour can be explained by their policy, but this mapping from states to distributions over actions can be arbitraily opaque. How can we explain the behaviour of agents by revealing how they were influenced by the features of their observations?

Explaining Reinforcement Learning with Shapley Values [ICML 2023]

Discovering General Skills

All forms of intelligence are constrained by time and computational resources. Humans minimise planning efforts by learning hierarchical behaviours, enabling quick adaptation to new challenges. How can an agent discover and learn a collection of skills to continually adapt to change?

Publications

ICML SVERL Image

Explaining Reinforcement Learning with Shapley Values

Daniel Beechey, Thomas M. S. Smith, Özgür Şimşek

ICML 2023

Paper | Code | Poster

Chemistry Image

Reformulating Reactivity Design for Data-Efficient Machine Learning

Toby Lewis-Atwell, Daniel Beechey, Özgür Şimşek and Matthew N. Grayson

ACS Catalysis, 13(20), 2023

Paper | Code

Talks

Talk Image

How to Explain Reinforcement Learning with Shapley Values

Bath Doctoral Festival of Ideas, July 2024

Slides

Talk Image

An Introduction to Explainable and Hierarchical Reinforcement Learning

Bath AI Society, April 2024

Slides

Talk Image

Explaining Reinforcement Learning with Shapley Values

Bath Computer Science Conference, July 2023

Alan Turing Institute Student Presentations, June 2023

Slides | Paper | Code

Teaching