arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2203.11547cs.AIcs.CY
收藏

强化学习中的可解释性:观点与立场

Explainability in reinforcement learning: perspective and position

University of Zagreb Faculty of Electrical Engineering and Computing(萨格勒布大学电气工程与计算学院);Institute Ruder Boskovic(鲁杰尔·博什科维奇研究所)

Agneza Krajna, Mario Brcic, Tomislav Lipic, Juraj Doncevic

摘要:

人工智能(AI)已经嵌入到人们日常生活的许多方面,让AI为他们做决定已经成为常态。相对于其他机器学习范例,强化学习(RL)模型增加了可解决问题的空间。一些最有趣的应用是在具有不可微的预期奖励函数的情况下,在未知或未定义的环境中操作,以及用于超越任何教师的性能的算法发现,从而代理通过简单的反馈从实验经验中学习。应用范围及其社会影响是巨大的,仅举几例:基因组学,游戏(国际象棋,围棋等),一般优化,金融投资,政府政策,自动驾驶汽车,推荐系统等,因此,通过解释来提高基于RL的系统的信任和透明度至关重要。大多数关于人工智能可解释性的文章都提供了涉及监督学习的方法,而在RL领域很少有文章涉及这一点。其原因是信用分配问题,延迟奖励,以及无法假设数据是独立同分布的。本文试图对可解释强化学习领域的现有方法进行系统概述,并提出一种新的统一分类法,建立和扩展现有方法。立场部分描述了如何观察可解释性的语用方面。特别强调了接受解释和提出解释的各方之间的差距。为了缩小差距,实现解释的诚实和真实,我们建立了三个支柱:主动性,风险态度和认识论约束。为此,我们说明了我们的建议的最短路径问题的简单变种。

英文摘要:

Artificial intelligence (AI) has been embedded into many aspects of people's daily lives and it has become normal for people to have AI make decisions for them. Reinforcement learning (RL) models increase the space of solvable problems with respect to other machine learning paradigms. Some of the most interesting applications are in situations with non-differentiable expected reward function, operating in unknown or underdefined environment, as well as for algorithmic discovery that surpasses performance of any teacher, whereby agent learns from experimental experience through simple feedback. The range of applications and their social impact is vast, just to name a few: genomics, game-playing (chess, Go, etc.), general optimization, financial investment, governmental policies, self-driving cars, recommendation systems, etc. It is therefore essential to improve the trust and transparency of RL-based systems through explanations. Most articles dealing with explainability in artificial intelligence provide methods that concern supervised learning and there are very few articles dealing with this in the area of RL. The reasons for this are the credit assignment problem, delayed rewards, and the inability to assume that data is independently and identically distributed (i.i.d.). This position paper attempts to give a systematic overview of existing methods in the explainable RL area and propose a novel unified taxonomy, building and expanding on the existing ones. The position section describes pragmatic aspects of how explainability can be observed. The gap between the parties receiving and generating the explanation is especially emphasized. To reduce the gap and achieve honesty and truthfulness of explanations, we set up three pillars: proactivity, risk attitudes, and epistemological constraints. To this end, we illustrate our proposal on simple variants of the shortest path problem.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑