摘要：郑重声明：原文参见标题，如有侵权，请联系作者，将会撤销发布！

 

 

 arXiv:1902.08102v1 [stat.ML] 21 Feb 2019

 

Abstract

 

1. Introduction

 

2. Background

 

2.1. Bellman equations

 

2.2. Categorical and quantile distributional RL

 

CDRL.

 

QDRL.

 

3. The role of statistics in distributional RL

 

3.1. Expectiles

 

3.2. Imputation strategies

 

3.3. Expectile distributional reinforcement learning

 

3.4. Stochastic approximation

 

4. Analysing distributional RL

 

4.1. Bellman closedness

 

4.2. Approximate Bellman closedness

 

4.3. Mean consistency

 

5. Experimental results

 

5.1. Tabular policy evaluation

 

EDRL.

 

QDRL.

 

5.2. Tabular control

 

5.3. Expectile regression DQN

 

6. Conclusion