Abstract: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.

Deep learning reveals hidden interactions in complex systems

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

Beyond Tracking: Using Deep Learning to Discover Novel Interactions in Biological Swarms

Evolving Neural Networks Reveal Emergent Collective Behavior from Minimal Agent Interactions

Navigating the swarm: Deep neural networks command emergent behaviours

Deep learning systems as complex networks

Deep learning based parameter search for an agent based social network model

VectorNet: Encoding HD Maps and Agent Dynamics From Vectorized Representation

Graph Attention Inference of Network Topology in Multi-Agent Systems

Learning Individual Interactions from Population Dynamics with Discrete-Event Simulation Model

Deep attributed network representation learning of complex coupling and interaction

Uncovering interaction patterns of multi-agent collective motion via complex network analysis

Modeling agent decision and behavior in the light of data science and artificial intelligence

Model-free inference of direct network interactions from nonlinear collective dynamics

Predicting the long-term collective behaviour of fish pairs with deep learning

Scaling Large-Language-Model-based Multi-Agent Collaboration

Multi-Agent Game Abstraction Via Graph Attention Neural Network.

Behavior-Inspired Neural Networks for Relational Inference

Graph MADDPG with RNN for multiagent cooperative environment

Looming Detection in Complex Dynamic Visual Scenes by Interneuronal Coordination of Motion and Feature Pathways