Learn to Interpret Atari Agents

By Zhao Yang, Song Bai, Li Zhang, Philip H. S. Torr. 2018

30 citations recorded by the source corpus.

Deep Reinforcement Learning (DeepRL) agents surpass human-level performances in a multitude of tasks. However, the direct mapping from states to actions makes it hard to interpret the rationale behind the decision making of agents. In contrast to previous a-posteriori methods of visualizing DeepRL policies, we propose an end-to-end trainable framework based on Rainbow, a representative Deep Q-Network (DQN) agent. Our method automatically learns important regions in the input domain, which enables characterizations of the decision making and interpretations for non-intuitive behaviors. Hence we name it Region Sensitive Rainbow (RS-Rainbow). RS-Rainbow utilizes a simple yet effective mechanism to incorporate visualization ability into the learning model, not only improving model interpretability, but leading to improved performance. Extensive experiments on the challenging platform of Atari 2600 demonstrate the superiority of RS-Rainbow. In particular, our agent achieves state of the art at just 25% of the training frames. Demonstrations and code are available at this https URL

Subfields: Artificial Intelligence

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

Citations in the corpus

Exploratory Not Explanatory: Counterfactual Analysis of Saliency Maps for Deep Reinforcement Learning (2019, background): “…Smilkov et al., 2017; Zhang et al., 2018) and to explain how agents choose actions in reinforcement learning (RL) applications (Bog-danovic et al., 2015; Wang et al., 2016; Zahavy et al., 2016; Greydanus et al., 2017; Iyer et al., 2018; Sundar, 2018; Yang et al., 2018; Annasamy & Sycara, 2019).”
Counterfactual States for Atari Agents via Generative Deep Learning (2019, background): “…of the past work on explaining machine learning has focused on explaining what features were important for a prediction (e.g. [Ribeiro et al. , 2016]) or on identifying regions of the visual input that cause the agent to perform a certain action (e.g. [Greydanus et al. , 2018; Yang et al. , 2018]).”
Free-Lunch Saliency via Attention in Atari Agents (2019, methodology): “At time t, input frames st are turned into embeddings vt, ht is the recurrent state: (a) Sparse convolutional block; (b) Dense convolutional block; (c) Region-Sensitive Module [42]; (d) DARQN [36] (N = 256) and [9] (N = 512); (e) architecture from [26]; (f) our architecture (ht ≡ 0).”