Dissertation Proposal. Julian Bishop. Evolutionary Feature Discovery for Online Reinforcement Learning.

Contact Name: 
Julian Bishop
Sep 28, 2012 11:00am - 1:00pm


This project takes a step towards general Artificial Intelligence (AI) by contributing a versatile online

Reinforcement Learning (RL) agent based on feature discovery. The completed work has produced an

online RL agent that can evaluate and select features from a universe that is far larger than the number

of features used by the agent for determining its policy. The agent has been successfully tested in the

problem domain of Connect-4 using hand-designed features. In the proposed work, this agent will be

extended to construct and improve its features automatically using evolutionary algorithms, and will

then be tested in additional domains. The resulting agent will also be used to investigate feature

representation bias in RL by evaluating a variety of function representations for the features. Such

functions will facilitate representation-agnostic feature discovery, and obviate expensive humanengineered

features. Therefore this project will make RL effective in real-world problems where domain

expertise is scarce, and in doing so helps overcome a significant barrier to general-purpose AI.