Tom Cannon

Task Agnostic Efficient and Adaptive Edge Devices Using Hierarchical Reinforcement Learning, Task Inference and Transfer Learning

Project Summary

My project aims to improve how an artificially intelligent agent infers and transfers skills and knowledge from known environments to unknown environments. Humans are very good at this form of inference and transfer, imagine driving a car in a different country; it may feel strange that the steering wheel, gearstick and road is in a different location – however, we can rapidly adapt using our previous models of the road, car and our own bodies in order to drive safely.

The first part of my project will focus on fundamental algorithms which will enable an agent to explore (skill development) across many different environments whilst also building efficient dynamic hierarchical environmental models which are identifiable from initial dynamic state observations. In the second part of the project, I aim to develop a transfer learning method capable of using known skills in known environments and quickly adapting them to useful skills in unknown environments.

I hope that this research will improve the safety and transparency of powerful AI agents (via a reduction in failed attempts at runtime and accurate model reporting) whilst also helping to bring the advantages of powerful AI processes and insights into the hands of the many without complex set-up and training.

Research Interests

Artificial Intelligence.

Reinforcement learning.

Meta-learning with a focus on meta-inference and transfer learning.

Unsupervised hierarchical skill development.

Unsupervised hierarchical environmental modelling.


MEng Aerospace Engineering with Aerodynamics (University of Southampton).

I have 5 years leadership and management experience as an Officer in the British Army.


Prof Özgür Şimşek

Dr Janina Hoffmann