In recent years, Unmanned Combat Aerial Vehicles (UCAVs) have achieved rapid development, and have played a great role in local combat due to the advantages of strong survivability, low cost-effectiveness ratio, and flexible configuration, etc. Autonomous air combat decision-making of UCAVs is one of the key technologies contributing to the winning of air combat, and reasonable evasion of incoming targets is the primary condition for air combat decision-making. Along with the development of information technology, modern air combat missions tend to be multi-level and multi-stage, with intricate and complex situations, diverse and varied target types, and scarce and intermittent intelligence information, which together constitute a highly challenging combat environment. In such a context, the improvement of the survival rate of high-value unmanned combat vehicles under the premise of accomplishing all stages of the combat mission has far-reaching implications and significant value for the realization of combat objectives, the maintenance of the overall level of combat capabilities, and the ultimate achievement of combat victory. Facing the multi-stage air combat mission, the critical factor to improve the survival rate of UCAV is the situational awareness and decision-making capability of the model. Reinforcement learning has strong decision-making ability, but the algorithm performance is constrained by the perception ability, and the traditional reinforcement learning method has slow strategy convergence and low data utilization.

For these issues, this paper proposed a data-knowledge driven intelligent avoidance decision-making method by utilizing the flight characteristics of sustainable large maneuvers of UCAVs. Facing the multi-stage air combat mission, the method employed MDP to formally model the avoidance decision-making process of UCAV in the joint air-sea-land combat, introduced the self-attention mechanism of blocked interaction with self-enemy situation based on reinforcement learning, and constructed a data-knowledge driven strategy updating method. Through the sample mixing of offline database-online interaction chain and the strategy mixing of knowledge decision base-avoidance agent, the rule knowledge was extracted from the existing historical data and decision-making experience, which guided the unmanned combat aircraft to generate accurate and efficient avoidance solutions in the process of continuous interaction and feedback with the multi-stage air combat mission environment.

Based on a simulation platform, this paper analyzes the avoidance performance and evaluates the air combat effectiveness of the proposed model, and the results show that this model not only improves the avoidance success rate against incoming targets, but also achieves a more competitive combat effectiveness compared with the traditional reinforcement learning DQN and rule-based model. Specifically, compared to the DQN and rule-based models, the proposed model not only accelerates the convergence speed while increasing the overall avoidance success rate by about 10% and 18%, but also dramatically improves the survival time of the fighters in the air combat process, which provides the living force support for the completion of the multi-stage mission. Besides, the experimental results also reveal the fact that tactical maneuver strategy is the most effective defense against incoming missiles during air-air combat, while it is difficult to evade anti-aircraft missiles from sea vessels and ground vehicles because of their own fast speed and strong end-to-end guidance capability, which is a pain point and difficulty for long-range pre-planning of intelligent decision.

These results show that the data-knowledge driven intelligent avoidance decision-making method can improve the survival rate and mission completion of the UCAV, realizing efficient avoidance decision-making under the multi-stage air combat mission, which demonstrates it has a certain theoretical significance and reference value for the intelligent modeling of the decision-making behavior of the UCAV.

CGF (computer generated force) acts as a core component in modern military simulation systems, and the authenticity and intelligence level of its decision-making behavior modeling directly determine the effectiveness of war simulation, tactical training, and operational deduction. Traditional modeling approaches suffer from rigid knowledge representation, scarcity of high-quality training samples, insufficient modeling of complex decision-making processes, and limited behavioral evolution capabilities, which severely restrict the development of intelligent and highly realistic CGF systems. LLMs (large language models), with their superior natural language understanding and generation, knowledge utilization, and few-shot reasoning abilities, provide a new technical paradigm to address these bottlenecks. This paper systematically reviews the research progress of LLM empowered CGF decision-making behavior modeling, which is of great theoretical significance and practical value for promoting the intellectual upgrading of military simulation systems and supporting the research and development of next-generation intelligent CGF.

This paper first identified four core challenges faced by current CGF decision-making behavior modeling: decision knowledge representation, decision complexity modeling, decision behavior evolution, and scarcity of high-quality decision samples. Then, it clarified three enabling paths of LLMs for CGF, including data and knowledge enhancement, decision intelligence generation, and capability iterative evolution. On this basis, a complete LLM based CGF decision-making behavior modeling framework was constructed, which consisted of five key modules: perception, decision-making, action, role, and memory. The technical implementation routes and representative research works of each module were elaborated in detail. The perception module transforms heterogeneous battlefield situation data into standardized semantic information; the decision module generated reasonable and interpretable strategies through military knowledge fusion, structured reasoning, and hierarchical planning; the action module converts high-level strategies into executable instructions constrained by equipment performance and battlefield rules; the role module endows CGF with personalized decision-making characteristics to solve the problem of decision homogeneity; the memory module realized experience accumulation and behavioral evolution through memory modeling, retrieval, and dynamic evolution mechanisms. Finally, this paper summarized potential research directions from five aspects: decision real-time performance, decision quality, decision fidelity, evaluation system, and decision risk control.

LLMs have demonstrated remarkable application potential in CGF decision-making behavior modeling and have become a key technology for constructing evolvable, interactive, and highly realistic intelligent CGF. At present, relevant research is still in the initial exploration stage, and there are still key problems to be solved urgently, such as the mismatch between LLM reasoning delay and tactical real-time requirements, LLM hallucinations affecting decision reliability, the lack of a complete and standardized evaluation system, and decision security risks in military scenarios. Future research should carry out in-depth exploration in lightweight model customization, large and small model collaborative decision-making, human factors embedded modeling, standardized evaluation system construction, and full-process security risk control. Meanwhile, strengthening interdisciplinary integration of artificial intelligence, military psychology, operations research, and other fields will continuously promote the maturity and practical application of LLM empowered CGF technology, and provide strong support for the development of intelligent military simulation.