| 引用本文: | 苏剑波,陈叶飞,马哲,黄瑶,向国菲,陈若冰.从AlphaGo到BetaGo——基于任务可完成性分析的定性人工智能的定量实现[J].控制理论与应用,2016,33(12):1572~1583.[点击复制] |
| SU Jian-bo,CHEN Ye-fei,MA Zhe,HUANG Yao,XIANG Guo-fei,CHEN Ruo-bing.From AlphaGo to BetaGo —— Quantitative realization of qualitative[J].Control Theory and Technology,2016,33(12):1572~1583.[点击复制] |
|
| 从AlphaGo到BetaGo——基于任务可完成性分析的定性人工智能的定量实现 |
| From AlphaGo to BetaGo —— Quantitative realization of qualitative |
| 摘要点击 3384 全文点击 2611 投稿时间：2016-06-24 修订日期：2017-01-13 |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| DOI编号 10.7641/CTA.2016.60440 |
| 2016,33(12):1572-1583 |
| 中文关键词 AlphaGo 特征完备性 表征空间 任务可完成性 抗干扰 美感评价 人工智能 |
| 英文关键词 AlphaGo feature completeness representation space task realizability disturbance rejection beauty evaluation arti?cial intelligence |
| 基金项目 国家自然科学基金项目(61533012, 61521063)资助. |
|
| 中文摘要 |
| AlphaGo程序利用深度学习算法和蒙特卡洛树搜索算法在围棋领域取得了突破性进展,用定量分析技术实
现了围棋这一传统定性式的任务.此次突破, 实现了对定性人工智能的定量研究, 这对人工智能体的研究具有重要
借鉴意义.对于人工智能体而言, 执行任务前需要考虑任务的可完成性, 对下棋任务而言, 任务目标是取得胜利,因
此, 本文先从任务可完成性角度出发,分别从特征完备性、表征空间构建及基于表征空间的搜索角度分析AlphaGo
程序. 其次, 人工智能体在任务完成过程中, 不可避免地受到各种扰动的影响, 对AlphaGo而言, 本质是对人下棋过
程的建模, 因此, 本文从抗干扰的角度出发, 分析了AlphaGo的缺陷. 再者,人工智能的研究是人类用科学技术的方
式模拟大脑活动的过程. AlphaGo所体现出的围棋思想与人类棋手的围棋美学之间的差异,也是定量分析与定性描
述之间的差异. 因此,本文从美感评价角度对AlphaGo进行了分析和展望. 通过上述三个角度,本文诠释了AlphaGo
程序所包含的原理以及对定量化分析定性人工智能体研究的借鉴意义.本文认为, AlphaGo虽然取得了里程碑式的
进展,但在定性描述(如: 美学,艺术)以及系统未知扰动方面仍存在大量问题值得研究.人工智能的跨越式发展, 即
从Alpha级别提升至Beta级别,应该包含对事物定性分析的能力. 最后, 希望人工智能算法的研究工作者通过本文能
更关注于挖掘定性描述与定量分析之间的关联, 并进一步将人工智能算法提升至BetaGo乃至更高. |
| 英文摘要 |
| AlphaGo, using deep learning and Monte Carlo Tree searching algorithms, has achieved great progress in
the game of Go, which realized the qualitative task by quantitative analysis. Such progress has realized the quantitative
research of qualitative AI, which has a signi?cant reference value for researchers. For AI, task realizability should
be taken into consideration before executing the task. The task goal of the game of Go is winning of the game.
Therefore, AlphaGo is ?rstly analyzed in the aspect of task realizability, including feature completeness, establishment
of representation space and searching method based on representation space. Secondly, during the process of task
executing, AI will confront various disturbance inevitably. The essence of AlphaGo is modeling the process of playing
chess of human. Hence, the drawbacks of AlphaGo is analyzed in the aspect of disturbance rejection. Thirdly, the
research of AI is a simulation of human brain activity by scienti?c technology. The beauty evaluation difference
between AlphaGo and human player re?ects the difference between quantitative analysis and qualitative description.
Therefore, AlphaGo is analyzed and prospected in the aspect of beauty evaluation. In this paper, the principle of
AlphaGo and the signi?cant reference value of quantitative realization of qualitative AI are annotated through the three
aspects mentioned above. Though AlphGo has achieved remarkable progress, we consider that there are still plenty of
problems remain to be studied in aspects of qualitative description (e.g. beauty evaluation, art) and unknown disturbance
of system. The signi?cant progress of AI should contains the ability of qualitative analysis, which is an leap from
Alpha level to Beta level. At last, the analysis in this paper is supposed to make AI researchers pay more attention to
dig the relationship between qualitative description and quantitative analysis and enhance the AI to BetaGo and higher level. |
|
|
|
|
|