Query-based learning for aerospace applications.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Aerospace Engineering

Background:

• Training neural networks for real-world applications is challenging due to numerous parameters and wide dynamic ranges, making data acquisition costly.
• Existing methods struggle with large-scale datasets, necessitating efficient learning strategies.

Purpose of the Study:

• To investigate and enhance query-based learning (QBL) for efficient neural network training in complex applications.
• To improve model generalization and reduce data acquisition costs using active learning techniques.

Main Methods:

• Employed network inversion (discrete and continuous) to identify performance-critical data for QBL.
• Introduced a novel heuristic for selecting target values in continuous network inversion.
• Utilized node decoupled extended Kalman filter (NDEKF) training and a causality index (CI) for dimensionality reduction and efficiency.

Main Results:

• Demonstrated the effectiveness of QBL in improving neural network efficiency and generalization.
• Successfully applied the QBL approach to aerospace classification and control distribution problems.
• Showcased the benefits of NDEKF training and CI in reducing input search dimensionality.

Conclusions:

• QBL, augmented with network inversion and advanced filtering techniques, offers a viable solution for training neural networks on complex, data-intensive problems.
• The proposed methods significantly enhance learning efficiency and generalization, particularly in aerospace applications.