Online Relational Inference for Evolving Multi-agent Interacting Systems

TL;DR

Experimental results demonstrate that ORI significantly improves the accuracy and adaptability of relational inference in dynamic settings compared to existing methods, and offers a robust solution for real-time applications in complex, evolving systems.

Abstract

We introduce a novel framework, Online Relational Inference (ORI), designed to efficiently identify hidden interaction graphs in evolving multi-agent interacting systems using streaming data. Unlike traditional offline methods that rely on a fixed training set, ORI employs online backpropagation, updating the model with each new data point, thereby allowing it to adapt to changing environments in real-time. A key innovation is the use of an adjacency matrix as a trainable parameter, optimized through a new adaptive learning rate technique called AdaRelation, which adjusts based on the historical sensitivity of the decoder to changes in the interaction graph. Additionally, a data augmentation method named Trajectory Mirror (TM) is introduced to improve generalization by exposing the model to varied trajectory patterns. Experimental results on both synthetic datasets and real-world data (CMU MoCap for human motion) demonstrate that ORI significantly improves the accuracy and adaptability of relational inference in dynamic settings compared to existing methods. This approach is model-agnostic, enabling seamless integration with various neural relational inference (NRI) architectures, and offers a robust solution for real-time applications in complex, evolving systems.

Figures (7)

• Figure 1: A brief illustration of the proposed Online Relational Inference (ORI) framework.
• Figure 2: Prediction results of ORI with MPMr decoder and the baseline MPM in the springs system. (a) the relation accuracy in the two models throughout the training (top) and visualization of the target and predicted adjacency matrix in our model (bottom). (b) target and predicted trajectories in our model.
• Figure 3: Prediction results of ORI with NRIr decoder in the charged system with evolving interaction and parameters (a) and ORI with MPMr decoder in the springs and charged systems with evolving interaction and dynamics (b). 1-st row compares the relation accuracy between constant learning rates and AdaRelation. 2-nd row shows changes in the relation learning rate throughout the training.
• Figure 4: Comparison between ORI and existing methods with respect to the relation accuracy (a), variance in the adjacency matrix (b), and variance in the predicted trajectory (c) depending on encoder complexity. The number in the MPM ($\cdot$) represents the dimension of hidden states in the encoder.
• Figure 5: Prediction results of ORI with MPMr decoder and MPM in CMU MoCap dataset. 1-st row represents the last frame in the predicted and target trajectory from ORI. 2-nd and 3-rd rows visualize the top-30 stongest interaction edges in the corresponding frame from ORI and MPM. Note that MPM allocate higher relation strengths in the front foot while ORI focuses on the foot behind.