主题: Graph Neural Network–Guided Monte Carlo Tree Search for Solving Mixed-Integer Linear Programs

中文题目：用于求解混合整数线性规划的图神经网络引导的蒙特卡洛树搜索

时间: 2026年1月27日 上午 10:30-11:30

地点: 管理科研楼一楼 第四教室

Bio:张公伯，北京大学光华管理学院博士后研究员。主要研究方向包括仿真优化理论与应用、机器学习、人工智能（AI运筹/科学智能）等。在IJOC、IEEE TAC、IEEE TASE等高质量期刊上发表学术论文。主持国家自然科学基金青年科学基金（C类）、中国博士后科学基金特别资助和面上资助项目。入选北京大学博雅博士后项目，曾获IJOC Meritorious Paper Award、智能制造系统工程学术会议优秀论文奖、北京运筹学会青年优秀论文奖等。

照片：

Abstract: To address the challenge of myopic decision-making in Mixed-Integer Linear Programming (MILP), an inherent limitation of traditional heuristics and learning-based branching approaches, we propose a neural-guided Monte Carlo tree search framework that treats MILP solving as a sequential lookahead planning process. We leverage a graph convolutional policy-value network to extract structural insights from the bipartite representation of each LP relaxation, refined through a pretraining-then-finetuning paradigm that transitions from a supervised warm start to asynchronous distributed training to effectively guide the tree search. To enable large-scale parallel exploration, we design robust interaction mechanisms that bridge neural-guided search with existing non-thread-safe solvers, ensuring numerical stability and consistent pruning. On the theory side, we establish the asymptotic consistency of our search policy. This framework not only demonstrates strong robustness on large-scale MILP instances where commercial solvers often struggle, but also highlights how integrating graph representation learning with lookahead search can mitigate the greedy behavior and local-optimum traps of existing solution methods.