An Explanation-based Reinforcement Learning Approach

This project proposes a novel reinforcement learning (RL) algorithm which is guided by a knowledge-based user feedback. The project integrates user feedback using structured knowledge bases into the traditional RL models. The main objective of the project is to reduce the barriers for a continuous, knowledge-driven dialogue between end-users and RL models, improving few-shot learning capabilities (the ability to generalize from fewer examples).

The core research questions are:

- Can a Knowledge Base driven reinforcement learning (RL) approach supported by end-user feedback deliver few-shot learning capabilities?
- Which Knowledge Representation models and formalisms can support generalization and user feedback process?
- How semantic parsing methods can be used to support the end-user interaction with the knowledge base?

The research can impact the ability to deliver Artificial Intelligence (AI) systems which are more transparent, and which generalize from fewer examples. These two properties are at the center of the requirements for the application of AI within scenarios which require trust (e.g. health and legal domains).

This research is positioned within the EPSRC `data to knowledge' priority area.