Interpreting Decision-Making Behavior in AI-Piloted Aircraft in Aerial Combat Scenarios: An Approach to Enhance Human-AI Trust

With the continuous advancement of artificial intelligence (AI) technology, AI algorithms have demonstrated exceptional aircraft control capabilities in highly dynamic and complex scenarios such as aerial combat. However, the inherent lack of explainability in AI algorithms poses a significant challenge to gaining sufficient trust, presenting potential safety risks that could lead to aircraft loss of control. This limitation hinders the widespread adoption of AI in practical applications. To enhance human–AI trust, improve system stability and safety, and advance the deployment of AI algorithms in practical settings, this study proposes an approach to describe and explain AI decision-making behaviors using natural language. Natural language is a straightforward medium for expressing information, which avoids the need for additional decoding or interpretation, particularly in rapidly changing battlefield environments, enabling pilots to quickly comprehend the intentions of AI algorithms and thereby fostering trust in AI systems. This study constructs a dataset of AI decision behavior description and interpretation based on adversarial temporal data in an aerial combat scenario and introduces an encoder–decoder framework that integrates an attentional mechanism. Findings from the experiments suggest that this approach effectively delineates and elucidates the AI decision-making behaviors, thereby facilitating mutual trust between humans and AI.