Abstraction for Deep Reinforcement Learning
This work addresses the abstraction challenge in deep reinforcement learning, which is incremental as it reviews and synthesizes existing approaches without introducing new methods or results.
The paper characterizes the abstraction problem in deep reinforcement learning and reviews existing AI/ML developments that could enable the adoption of analogical reasoning and associative memory methods, which face challenges due to the requirement for end-to-end differentiability.
We characterise the problem of abstraction in the context of deep reinforcement learning. Various well established approaches to analogical reasoning and associative memory might be brought to bear on this issue, but they present difficulties because of the need for end-to-end differentiability. We review developments in AI and machine learning that could facilitate their adoption.