AI 무기체계의 자율 판단 신뢰도 향상을 위한 베이지안 신경망 적용 방법론

Bayesian Neural Network Approach for Reliable Autonomous Decision-Making in AI Weapon Systems

Recent advances in artificial intelligence (AI) have rapidly transformed the landscape of modern warfare, with AI-enabled weapon systems offering unprecedented improvements in target recognition, operational speed, and autonomous mission execution. However, despite these benefits, the reliability of autonomous decision-making remains a major concern, particularly in certain environments where incorrect predictions can result in unintended engagements or civilian casualties. Addressing this issue requires a means to not only improve prediction accuracy but quantify the uncertainty behind each decision. This study proposes a novel methodology for evaluating and enhancing the reliability of autonomous decisions in AI weapon systems by leveraging the framework of Bayesian Neural Networks(BNNs). Recognizing the computational limitations of applying BNNs to an entire deep learning model, a Partially Bayesian Neural Network (PBNN) architecture is introduced. This design enables the system to estimate the uncertainty of its own outputs while maintaining real-time performance—an essential requirement in military applications. Furthermore, a semi-autonomous control structure is implemented, in which human intervention is selectively triggered based on the estimated uncertainty. The results show that the proposed method produces significantly lower output variance and higher confidence scores when predictions are correct, while consistently identifying high-uncertainty cases associated with incorrect predictions. This allows the system to distinguish between reliable and unreliable decisions in real time. The proposed approach offers a technically viable pathway toward balancing autonomy and human oversight in AI weapon systems. It also serves as a foundational framework for the future implementation of Manned-Unmanned Teaming (MUM-T).