Scalable Intent Recognition in Wargame Using Transformer

The intricacy and dynamism of the wargame present formidable challenges to the stability and reliability of traditional methods for behavioral intent recognition. To tackle this issue head-on, we propose an innovative architecture called Behavior Sequence Compression and Transformer (BSCT) specifically designed to recognize target behavior intentions within wargame. Our research primarily focuses on the development and comparison of various behavioral sequence compression methods. These methods have been meticulously crafted to address the inherent challenge posed by exceptionally long sequences, where the standard transformer architecture may prove inadequate. Through a comprehensive series of experiments, we demonstrate that the integration of the long-sequence compression method with the transformer architecture yields a remarkable improvement of 10 to 30% in the accuracy of behavioral intent recognition within wargame.