Dan Beechey

Autonomous Development of Action Hierarchies

Project Summary

My project aims to develop algorithms which allow agents to autonomously learn useful high-level skills and build action hierarchies within the reinforcement learning framework. A series of primitive actions can be collected into a single higher-level action, such as combining the joint movements in a leg to take a step. The higher-level actions can then be further combined, forming natural hierarchies for skills such as running. Agents can continuously build on skills to develop more and more complex behaviour. The simplification of a complex series of primitive actions into fewer higher-level actions allows for explainability through increased transparency. 

Research Interests

My main research interests lie in machine learning, particularly Reinforcement Learning with an emphasis on Hierarchical Reinforcement Learning. 

The application of reinforcement in behavioural psychology to the field of Reinforcement Learning. 

Explainability and Intelligibility in AI. 

Background

BSc Mathematics, University of Bath.

MSc Data Science, University of Bath.

Supervisors

Dr Özgür Şimşek