揭示生物集群系统内部信息耦合机制的深度网络模型

Analysis model for revealing mechanism of internal information coupling in biological collective systems based on deep network

生物集群在自然界中广泛存在,如鱼群、鸟群、兽群等,这些集群通过内部的信息耦合能涌现出有序的协同行为.但是,集群内部交互复杂、情况多变,导致微观层面的行为分析还缺乏行之有效的标准工具.对此,以鱼群运动数据为研究对象,借助图注意力网络对鱼群中的单体行为进行自动化建模,目的是提炼出适于复杂系统分析的通用网络工具.首先将单体的感知信息映射到高维状态空间,然后生成软注意力数值以表征单体之间的交互强度,再对所生成的软注意力数值规范化,所得规范结果既可作为多邻居信息耦合的关键指标,又可通过解码器将所抽取的注意力信息稀疏解耦为运动决策.实验结果表明:所生成的注意力数值不但能揭示群体内部的信息耦合关系,还能进一步对隐藏交互强度可视化,从而为鱼群的视觉交互理论提供新的科学证据.所提出分析工具拥有信息耦合可解释、交互强度可显现、系统规模可缩放、状态偏移可泛化等优良特性,有望发展成为复杂系统解耦分析的标准范式,在社会行为分析、机器人集群控制以及智能交通系统安全性评价等方面具有潜在的应用价值.

The biological collective motion exists wildly in the natural world,such as fish schooling,birds flocking,herds migrating,etc.These speciaes can emerge cooperative behavior orderly through internal information coupling.However,due to the complexity of the internal interaction and ever-changing environment,there is still a lack of effective tools for behavioral analysis at the micro level.In this work,an embedded graphical attention deep network is employed for automatically building the model of the individual information interactions from the data of fish schooling,aiming to extract a general network tool suiting for the complex system analysis.This research maps the low-dimensional individual observations to the high-dimensional states space followed by the generation of soft attention values to represent the interaction strength between the individuals.These soft attention values are numerically normalized,which can be used as a key indicator for the information coupling of multi-neighbors.A decoder network is designed for transforming the extracted attention information into the motion decision of individuals.The experimental results show that the obtained attention value can not only reveal the hidden coupling relationship of the information interactions in collective systems,but also visualize the information interactions of the individuals,which can be used as scientific evidence for proving the visual communication theory on fish schooling.The presented analysis tool has the following excellent characteristics:First,the coupling of internal information can be explained;Second,the interaction strength of individuals can be visualized;Third,the quantity of individuals in the system can be scaled;Forth,the model can be generalized to the different distribution of collective states.In conclusion,the proposed tool is promising to become a standard artificial intelligence paradigm for decoupling analysis of complex systems,which has potential application values in the behavior analysis of social systems,distributed control of swarm robotics,and safety evaluation for intelligent transportation systems.