Reproducing “Towards Interpretable ReinforcementLearning Using Attention Augmented Agents”

Charles Lovering, Sam Lobel, Denizalp Goktas, Kweku Kwegyir-Aggrey, Albert Webson

Abstract: We attempt to reproduce "Towards Interpretable Reinforcement Learning UsingAttention Augment Agents", a recent work which introduces a novel attention mechanism to understand the reasoning behind the predicted policy for canonical Atari games in a reinforcement learning setting. Like the original paper, our implementation utilizes the Importance Weighted Actor Critic Architecture (IMPALA). Our implementation, despite using considerable resources, was only trained for a fraction of the number of steps done by the original authors. Nonetheless, the results that we obtain from our early-training stage model indicate the validity of the authors’ work.

Track: Replicability

NeurIPS Paper Id: https://openreview.net/forum?id=HJxm5BBeUS