Agent Behavior Prediction Using Observable Driving Parameters

The patent extracts and removes the problematic acceleration measurement from the behavior modeling process. Instead of using directly measured acceleration (which is noisy and unreliable), the system extracts only the necessary observable state parameters (position, velocity, heading) and uses these to predict acceleration through machine learning models, thereby eliminating the measurement accuracy problem while maintaining behavior modeling capability

The patent replaces the mechanical/physical measurement approach (direct acceleration sensing through double differentiation of position) with a computational approach (machine learning-based acceleration prediction). This substitution uses observable state parameters as inputs to predict acceleration, avoiding the inherent problems of direct mechanical measurement while achieving more reliable behavior modeling

The patent applies preliminary action by pre-training machine learning models offline using historical trajectory data. The complex computational work of learning the relationships between observable state parameters and agent-specific parameters is performed in advance during the training phase. During actual autonomous operation, the pre-trained models quickly infer parameters from current observations, significantly reducing real-time computational complexity while maintaining high prediction accuracy

The patent implements dynamics by using different computational approaches for different operational phases: offline training phase uses intensive computational methods to learn from large datasets, while online operation phase uses lightweight inference from pre-trained models. This dynamic allocation of computational resources optimizes the balance between prediction accuracy and real-time computational complexity