Strategic interactions among major powers are characterized by complex interdependencies, conflicting national interests, and dynamically evolving international environments. Traditional quantitative approaches, including multiple regression and time-series analysis, primarily focus on statistical relationships among political and economic variables, limiting their ability to characterize adaptive decision-making and reciprocal strategic responses. To address these limitations, this study proposes a data-informed multi-agent deep reinforcement learning framework for modeling and simulating strategic interactions between major powers, using Sino–U.S. relations as a representative case. First, a multidimensional strategic state representation is constructed by integrating bilateral trade, economic capabilities, exchange rates, diplomatic interactions, and geopolitical security indicators. Historical diplomatic records are processed through event extraction, semantic categorization, and quantitative relationship scoring to characterize both accumulated strategic influences and short-term interaction dynamics. Second, a multi-agent stochastic game model is developed to represent major powers as autonomous decision-making agents with interdependent strategic objectives. A multidimensional reward mechanism incorporating economic benefits, security interests, diplomatic gains, and confrontation costs is formulated to support adaptive strategic behavior modeling. Third, a deep reinforcement learning mechanism is introduced to enable agents to optimize their policies through repeated strategic interactions and environmental feedback. Furthermore, a scenario-based simulation framework is designed to investigate strategic behavior evolution under alternative conditions, including economic cooperation, trade competition, diplomatic negotiation, and security confrontation. Historical consistency assessment, comparative baseline analysis, and parameter sensitivity evaluation are incorporated to examine model credibility and behavioral robustness.