Graying the black box: Understanding DQNs
This work addresses the interpretability of deep reinforcement learning models, which is an incremental improvement for researchers and practitioners seeking to debug and optimize neural networks in RL.
The paper tackled the problem of understanding Deep Q-networks (DQNs) by developing a methodology and tools to analyze them non-blindly, revealing that DQNs aggregate the state space hierarchically and enabling interpretation of policies for Atari2600 games.
In recent years there is a growing interest in using deep representations for reinforcement learning. In this paper, we present a methodology and tools to analyze Deep Q-networks (DQNs) in a non-blind matter. Moreover, we propose a new model, the Semi Aggregated Markov Decision Process (SAMDP), and an algorithm that learns it automatically. The SAMDP model allows us to identify spatio-temporal abstractions directly from features and may be used as a sub-goal detector in future work. Using our tools we reveal that the features learned by DQNs aggregate the state space in a hierarchical fashion, explaining its success. Moreover, we are able to understand and describe the policies learned by DQNs for three different Atari2600 games and suggest ways to interpret, debug and optimize deep neural networks in reinforcement learning.