Sample efficiency, transfer learning and interpretability for deep reinforcement learning

Deep learning has revolutionised artificial intelligence, where the application of increased compute to train neural networks on large datasets has resulted in improvements in real-world applications such as object detection, text-to-speech synthesis and machine translation. Deep reinforcement learning (DRL) has similarly shown impressive results in board and video games, but less so in real-world applications such as robotic control. To address this, I have investigated three factors prohibiting further deployment of DRL: sample efficiency, transfer learning, and interpretability. To decrease the amount of data needed to train DRL systems, I have explored various storage strategies and exploration policies for episodic control (EC) algorithms, resulting in the application of online clustering to improve the memory efficiency of EC algorithms, and the maximum entropy mellowmax policy for improving the sample efficiency and final performance of the same EC algorithms. To improve performance during transfer learning, I have shown that a multi-headed neural network architecture trained using hierarchical reinforcement learning can retain the benefits of positive transfer between tasks while mitigating the interference effects of negative transfer. I additionally investigated the use of multi-headed architectures to reduce catastrophic forgetting under the continual learning setting. While the use of multiple heads worked well within a simple environment, it was of limited use within a more complex domain, indicating that this strategy does not scale well. Finally, I applied a wide range of quantitative and qualitative techniques to better interpret trained DRL agents. In particular, I compared the effects of training DRL agents both with and without visual domain randomisation (DR), a popular technique to achieve simulation-to-real transfer, providing a series of tests that can be applied before real-world deployment. One of the major findings is that DR produces more entangled representations within trained DRL agents, indicating quantitatively that they are invariant to nuisance factors associated with the DR process. Additionally, while my environment allowed agents trained without DR to succeed without requiring complex recurrent processing, all agents trained with DR appear to integrate information over time, as evidenced through ablations on the recurrent state.Open Acces