复杂系统中的因果涌现研究综述  

Causal emergence in complex systems

作　　者：王志鹏 张江 WANG Zhipeng;ZHANG Jiang(School of Systems Science,Beijing Normal University,100875,Beijing,China;Swarma Research,100085,Beijing,China)

机构地区：[1]北京师范大学系统科学学院,北京100875 [2]北京集智研究院,北京100085

出　　处：《北京师范大学学报（自然科学版）》2023年第5期725-733,共9页Journal of Beijing Normal University(Natural Science)

基　　金：国家自然科学基金资助项目(61673070)。

摘　　要：聚焦复杂系统:就涌现、因果以及因果涌现进行了定量描述;阐述了因果与涌现的联系;系统梳理了基于粗粒化和信息分解2种定量刻画因果涌现的方法,以及基于信息分解和神经信息压缩2种因果涌现辨识方法;详细介绍了各方法的基本原理、优缺点,以及相关应用等.基于粗粒化方法定义了因果涌现框架,其可应用于离散动力系统;利用信息分解方法求解时结果会依赖冗余信息;基于信息分解所提出的指标可识别数据中的因果涌现,且找到一个充分条件;基于数据驱动的神经信息压缩方法可扩展并应用于连续动力系统,并可自动提取不同层级的粗粒化函数,以及构建不同层级的动力学,还可识别不同类型动力学系统的因果涌现;改进现有方法并用于识别更为复杂的系统与学习自动化分组,解决通用动力学大模型等问题.Focusing on complex systems:a quantitative description is provided for emergence,causality,and causal emergence.The relationship between causality and emergence is elucidated.Two quantitative methods,coarsegraining and information decomposition,are reviewed for characterizing causal emergence.Two causal emergence identification methods based on information decomposition and neural information compression are presented.Basic principles,advantages,disadvantages,and relevant applications of each method are detailed.The framework of causal emergence is defined based on coarse-graining method,which can be applied to discrete dynamical systems.The use of information decomposition method in solving problems relies on redundant information.The proposed indicators based on information decomposition can identify causal emergence in data and find a sufficient condition.The data-driven neural information compression method is scalable and applicable to continuous dynamical systems.It can automatically extract coarse-grained functions at different levels and construct hierarchical dynamics,and identify causal emergence in different types of dynamical systems.Existing methods are improved and applied to identify more complex systems,automate learning groups,and address challenges in general dynamics of large-scale models and other related problems.

关 键 词：复杂系统 因果涌现 有效信息 粗粒化 信息分解 神经信息压缩 

分 类 号：N949[自然科学总论—系统科学]