Online Intention Recognition With Incomplete Information Based on a Weighted Contrastive Predictive Coding Model in...

Area of Science:

• Artificial Intelligence
• Deep Learning
• Military Science

Background:

• Traditional intention recognition methods struggle with incomplete battlefield data, impacting efficiency and reliability.
• The dynamic and uncertain nature of warfare necessitates advanced approaches for real-time situation assessment.

Purpose of the Study:

• To develop a robust deep learning architecture for online intention recognition under incomplete information scenarios.
• To enhance the accuracy, stability, and speed of intention recognition in wargame simulations.

Main Methods:

• Proposed a novel deep learning architecture, W-CPCLSTM, integrating Contrastive Predictive Coding (CPC) and variable-length Long Short-Term Memory (LSTM) networks.
• Utilized CPC to capture global structures from limited intelligence data and LSTM for intention classification.
• Implemented an attention weight allocation mechanism for improved model stability during training.

Main Results:

• The W-CPCLSTM model demonstrated significant improvements in recognition accuracy (7%-11% over LSTM) and speed (6-32x faster than LSTM).
• Evaluated performance across various degrees of information completeness and data lengths in wargame simulations.
• Achieved superior performance compared to traditional LSTM, FCN, OctConv, and OctFCN models.

Conclusions:

• The W-CPCLSTM architecture offers a stable and accurate solution for online intention recognition with incomplete battlefield information.
• The model's enhanced performance makes it a valuable reference tool for command decision-making in complex wargame environments.
• This approach addresses key limitations of traditional methods in dynamic and uncertain operational settings.