Continual Reinforcement Learning for Non-stationary Environments

Developing artificial intelligence (AI) agents that can learn new concepts without forgetting what they’ve already learnt is a pressing challenge. The reason is the absence of natural cognition, unlike humans, that assists in managing and retaining information. This problem is further impaired when they must learn to resolve each new task from scratch, much like how infants learn through trial, error and feedback from their environment. This thesis explores methodologies to improve this continual learning process for AI agents using deep learning techniques. The proposed solutions provide efficient learning methods with improved recall of past learned tasks.