Abstract: Database Management System (DBMS) configuration and optimization are directly tied to overall performance. Default settings rarely achieve optimal performance. Experience-driven manual tuning is constrained by the inability to continuously track system state and data distribution, often yielding suboptimal outcomes. Bayesian optimization (BO), while widely used, tends to fall into local optima on complex objectives, limiting global search. To address this, MTune was proposed, an MCTS-based method for DBMS configuration tuning. A policy tree was used to partition the multidimensional knob (parameter) space into regions, with each node representing a distinct subspace. The MCTS objective was defined by submitting knob configurations to BO and receiving evaluation metrics; nodes were scored with the Upper Confidence Bound (UCB) criterion derived from this objective. K-means clustering was used to iteratively split the initial node to grow the policy tree, progressively shrinking the search space and balancing exploration and exploitation. Continuous region refinement mitigated local-optimum risk and enhanced global search capability. Results show that MTune outperforms strong baselines under YCSB-A and YCSB-B workloads. On PostgreSQL v9.6 with HeSBO-optimal settings, average transaction latency decreases by 97.13%–97.91%, throughput increases by 24.83%–48.56%, and system overhead decreases by 1.62%–16.26%. On PostgreSQL v13.6, latency decreases by about 95% and throughput increases by about 10%–25%; system overhead is on par with or slightly better than DDPG, and MTune is best against SMAC under HeSBO-16. MTune identifies high-quality regions and yields near-optimal knob configurations, and it is robust and effective in practice.