On Neural Consolidation for Transfer in Reinforcement Learning
This work addresses the challenge of understanding transfer mechanisms in reinforcement learning, which is incremental as it applies an existing method (distillation) to a known bottleneck (predicting transfer).
The paper tackled the problem of predicting knowledge transfer between tasks in deep reinforcement learning by using network distillation as a feature extraction method, showing that distillation does not prevent transfer, including from multiple tasks to a new one, with comparisons to transfer without distillation.
Although transfer learning is considered to be a milestone in deep reinforcement learning, the mechanisms behind it are still poorly understood. In particular, predicting if knowledge can be transferred between two given tasks is still an unresolved problem. In this work, we explore the use of network distillation as a feature extraction method to better understand the context in which transfer can occur. Notably, we show that distillation does not prevent knowledge transfer, including when transferring from multiple tasks to a new one, and we compare these results with transfer without prior distillation. We focus our work on the Atari benchmark due to the variability between different games, but also to their similarities in terms of visual features.