PROJECT TITLE:

Self-Organizing Neural Networks Integrating Domain Knowledge and Reinforcement Learning - 2015

ABSTRACT:

The use of domain knowledge in learning systems is predicted to improve learning efficiency and reduce model complexity. However, thanks to the incompatibility with data structure of the training systems and real-time exploratory nature of reinforcement learning (RL), domain data can't be inserted directly. During this paper, we tend to show how self-organizing neural networks designed for on-line and incremental adaptation will integrate domain knowledge and RL. Specifically, image-based domain data is translated into numeric patterns before inserting into the self-organizing neural networks. To confirm effective use of domain data, we tend to present an analysis of how the inserted knowledge is employed by the self-organizing neural networks during RL. To this finish, we tend to propose a vigilance adaptation and greedy exploitation strategy to maximise exploitation of the inserted domain information whereas retaining the plasticity of learning and using new data. Our experimental results based on the pursuit-evasion and minefield navigation downside domains show that such self-organizing neural network can build effective use of domain knowledge to boost learning potency and cut back model complexity.