Large-Model Reinforcement Fine-Tuning for Intelligent Decision-Making in Combat Simulation Deduction

Objective: To address multisource heterogeneous situation information, resource- and time-constrained task-level actions, insufficient semantic representation in conventional reinforcement learning (RL), and the limited ability of large language models (LLM) to use environmental feedback, an intelligent decision-making method based on large-model reinforcement fine-tuning is proposed. Methods: Task-level actions such as patrol, strike and interception are formulated as a constrained discrete decision process, and time, force status, regional deployment, resource margins, target situation and executable-action sets are organized as structured semantic observations. LLM is used as the policy generator and fine-tuned efficiently with low-rank adaptation (LoRA). An independent value network is introduced to estimate state values. Action parsing, fallback execution and invalid-action penalties are used to construct a closed simulation-interaction loop, and the policy and value networks are jointly optimized using a proximal policy optimization (PPO) objective. Results: In 30 test episodes of the deduction scenario, the LLM-RL method achieved a success rate of 90.0%, outperforming the behavior tree (BT) strategy, distributed proximal policy optimization (DPPO), and prompt-only large language model (Prompt-only LLM) method. Compared with Prompt-only LLM, action validity increased from 71.7% to 92.2%. Conclusion: The proposed method transforms constrained planning decision-making in combat simulation deduction into a semantically enhanced reinforcement fine-tuning problem, improving task completion quality while enhancing action executability.