Research on automatic pilot repetition generation method based on deep reinforcement learning.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Air Traffic Management

Background:

• Air traffic control (ATC) simulation training traditionally relies on pilot seats, which are costly and inefficient.
• Developing automated methods for generating realistic pilot communications is crucial for advancing ATC simulation.

Purpose of the Study:

• To propose and evaluate a deep reinforcement learning model, RoBERTa-RL, for generating pilot repetitions in ATC simulations.
• To enhance the efficiency and reduce the cost of ATC controller training through automated communication generation.

Main Methods:

• Utilized RoBERTa, a pre-trained language model, enhanced with transfer learning to address data scarcity in the ATC domain.
• Employed reinforcement learning algorithms to optimize the RoBERTa model, improving generalization capabilities.
• Trained and tested the model on real-world area control and simulated tower control datasets.

Main Results:

• RoBERTa-RL achieved high ROUGE scores (e.g., 0.996 for ROUGE-L on area control data).
• Keyword-based evaluation demonstrated high accuracy (98.8% on area control, 81.8% on tower control).
• Significant improvements in generalization were observed, with a 56% increase over the baseline model.

Conclusions:

• Deep reinforcement learning effectively enhances deep learning models for text generation, mitigating generalization issues.
• The RoBERTa-RL model shows significant promise for improving ATC simulation and training.
• The proposed approach has potential applications in other related text generation domains.