精准智能理论:面向复杂动态对象的人工智能

Refined intelligence theory: artificial intelligence regarding complex dynamic objects

新一轮科技革命和产业变革正在萌发,以深度学习和大数据为基础,以Alpha Go等为典型应用场景掀起了人工智能的第3次高潮.传统的基于统计线性化动态建模的人工智能,在处理复杂对象时遇到了可解释性、泛化性和可复现性等发展瓶颈,迫切需要建立基于复杂性与多尺度分析的新一代人工智能理论,我们称之为精准智能.针对复杂系统的非线性特征,精准智能构建内嵌领域知识和数学物理机理的系统学习理论,包括复杂数据科学感知、复杂系统精准构建、复杂行为智能分析3个层次.具体而言,通过复杂数据科学感知建立内嵌时空特征与数理规律等具有可解释性的科学数据系统;通过复杂系统精准构建反演具有非线性复杂逻辑关系的多层次、多尺度、可解释的人工智能动态学习模型;通过对系统复杂行为智能分析建立面向系统行为演进和全局动态分析的可解释可调控人工智能新理论和新方法.将上述精准智能理论应用于群体智能,提出了群体熵方法,实现了群体激发和汇聚行为复杂性度量与有效引导调控.

A new round of scientific and technological change and industrial transformation is emerging. Based on deep learning and big data, a third achievement of artificial intelligence is represented by Alpha Go and other typical application scenarios. When dealing with complex dynamic objects, the traditional artificial intelligence based on statistical linear dynamic modeling experiences bottlenecks related to interpretability, generalization,and reproducibility. It is urgent that a new generation of artificial intelligence theory be established that is based on complexity and multi-scale analysis, which we refer to as refined intelligence. To deal with the non-linearity of complex systems, we constructed a systems learning theory of embedded domain knowledge and a mathematical physical mechanism that is accurate and intelligent at three performance levels: complex data perception, complex system refined construction, and complex intelligent behavior analysis. Specifically, we built a scientific data system with interpretability, including embedded spatiotemporal characteristics and mathematical laws, through complex data perception. We also built a multi-level, multi-scale, and interpretable artificial intelligence dynamic learning model that can deal with the nonlinear relationship of complex logic based on complex system-defined construction. We developed a new theory and method for interpreting and controlling the evolution of artificialintelligence-oriented behavior and global dynamic analysis based on complex intelligent-behavior analysis. We applied the above refined intelligence theory to the proposed method for crowd intelligence with crowd entropy, and found that it can measure complexity and provide effective guidance regarding system stimulation and aggregation behavior.