An Environmental Model Learning Method Based on Improved Attention Mechanism

This paper addresses the information confrontation and situational intelligent cognition challenges in aerial combat environments by integrating advancements in world modeling research. It proposes an environmental model learning method that, through the combination of Mixture Density Networks with attention mechanisms, adapts the Transformer architecture for state-action input to environmental multivariate prediction output. A new loss function is introduced to balance the accuracy of multivariate predictions. The experimental section first analyzes the principle of action of the loss function and the balancing effect of multivariate predictions through simulation. It then conducts predictive analyses for different levels of noise and varying lengths of missing positional information, ranging from 10, 25, to 50 time steps. Simulation results indicate the feasibility and potential for engineering application of the environmental model learning method based on the enhanced attention mechanism in learning air combat environmental models, implementing multivariate predictions, and balancing prediction accuracy.