【论文阅读】Robustness and performance of Deep Reinforcement Learning.

一.论文信息

题目： Robustness and performance of Deep Reinforcement Learning.【深度强化学习的鲁棒性与性能】

发表年份： 2021

期刊/会议： Applied Soft Computing(中科院SCI 2区, 影响因子：8.263)
APPLIED SOFT COMPUTING

论文链接： https://doi.org/10.1016/j.asoc.2021.107295

作者信息： Raid Rafi OmarAl-Nima, Tingting Han, Saadoon Awad MohammedAl-Sumaidaee, Taolue Chen, Wai LokWood

二.论文结构

Abstract
1 Introduction
2  Literature review
	2.1. Literature without neuron coverage(没有神经元覆盖的文献)
	2.2. Literature with neuron coverage(有神经元覆盖的文献)
	2.3. Literature with genetic algorithm
	2.4. Contribution of this paper
3 Theoretical background
	3.1. Deep reinforcement learning
	3.2. Neuron coverage
	3.3. NC as a fitness function for the GA
	3.4. GA to generate more training samples
	3.5. How GA works in our setting
	3.6. The GANC algorithm
	3.7. Exploited DRL-RT model
4. Results
	4.1. General parameters
	4.2. Practical experiments
	4.3. Prior results
	4.4. GANC parameters and NC results
	4.5. Enhanced performances
	4.6. Comparisons
5. Conclusion

三.论文内容

Abstract

Deep Reinforcement Learning (DRL) has recently obtained considerable attentions. It empowers Reinforcement Learning (RL) with Deep Learning (DL) techniques to address various difficult tasks.

In this paper, a novel approach called the Genetic Algorithm of Neuron Coverage (GANC) is proposed. It is motivated for improving the robustness and performance of a DRL network. The GANC uses Genetic Algorithm (GA) to maximise the Neuron Coverage (NC) of a DRL network by producing augmented inputs. We apply this method in the self-driving car applications, where it is crucial to accurately provide a correct decision for different road tracking views. We evaluate our method on the SYNTHIA-SEQS-05 databases in four different driving environments. Our outcomes are very promising – the best driving accuracy reached 97.75% – and are superior to the state-of-the-art results.

摘要

深度强化学习(Deep Reinforcement Learning, DRL)近年来受到了广泛关注。它赋予强化学习(RL)深度学习(DL)技术以解决各种困难的任务（address various difficult taks）。

本文提出了一种新的方法，称为神经元覆盖的遗传算法(GANC)。它的动机是提高DRL网络的鲁棒性和性能。GANC使用遗传算法(GA)通过产生增强输入（by producing augmented inputs）来最大化DRL网络的神经元覆盖率(NC)。我们将这种方法应用于自动驾驶汽车应用中，在这些应用中，为不同的道路跟踪视图准确地提供正确的决策至关重要。我们在四个不同的驾驶环境下的SYNTHIA-SEQS-05数据库上评估了我们的方法。我们的结果非常有掐图-最佳驾驶精度达到97.75% -优于最先进的结果。