| The development of robust autonomous systems capable of learning and adapting to complex novel situations is the ultimate goal for research in machine learning and artificial intelligence. There is perhaps no other technology that holds the promise for enabling such robust autonomous systems as directly or as strongly as Reinforcement Learning (RL). Yet despite RL's promise, significant research challenges remain that must be overcome in order to realize its potential. The research presented in this thesis makes significant contributions towards addressing two of these challenges, representation learning and improved sample efficiency. Representation learning is an ability to automatically derive an effective description for a given problem to enable its solution. This capability is a critical component for adapting to novel and dynamic domains. Sample efficiency is a measure of how effectively a method can learn from a given quantity of experience data. Experience data can be limited and expensive to obtain, as such sample efficiency is an additional requirement of a scalable robust autonomous system. For representation learning we introduce a new automatic feature selection approach that learns concise and effective problem representations during the process of solving a problem. Empirically this approach demonstrates both improved learning performance and quality of the final solution. With regard to improving sample efficiency, we introduce truly novel approaches for exploiting the sequential nature of sample data which greatly improves value estimation and sample efficiency. We show through rigorous statistical, theoretical, and empirical analysis that our proposed methods make more effective use of available sample data resulting in greatly improved learning performance. Finally, we provide insight for future work that will bring the realization of robust autonomous systems even closer. |
