Collab-Solver: Collaborative Solving Policy Learning for Mixed-Integer Linear Programming

Mixed-integer linear programming (MILP) has been a fundamental problem in combinatorial optimization. Conventional MILP solving mainly relies on carefully designed heuristics embedded in the branch-and-bound framework. Driven by the strong capabilities of neural networks, recent research is exploring the value of machine learning alongside conventional MILP solving. Although learning-based MILP methods have shown great promise, existing works typically learn policies for individual modules in MILP solvers in isolation, without considering their interdependence, which limits both solving efficiency and solution quality. To address this limitation, we propose Collab-Solver, a novel multi-agent-based policy learning framework for MILP that enables collaborative policy optimization for multiple modules. Specifically, we formulate the collaboration between cut selection and branching in MILP solving as a Stackelberg game. Under this formulation, we develop a two-phase learning paradigm to stabilize collaborative policy learning: the first phase performs data-communicated policy pretraining, and the second phase further orchestrates the policy learning for various modules. Extensive experiments on both synthetic and large-scale real-world MILP datasets demonstrate that the jointly learned policies significantly improve solving performance. Moreover, the policies learned by Collab-Solver have also demonstrated excellent generalization abilities across different instance sets.

💡 Research Summary

**
The paper introduces Collab‑Solver, a novel multi‑agent framework that jointly learns policies for two tightly coupled components of mixed‑integer linear programming (MILP) solvers: cut selection and branching. Traditional MILP solvers such as SCIP, CPLEX, and Gurobi rely on a branch‑and‑cut (B&C) paradigm where cutting planes tighten LP relaxations before the search tree is branched. Existing learning‑based approaches improve individual modules (e.g., branching via graph neural networks, cut selection via reinforcement or imitation learning) but treat them in isolation, ignoring the strong upstream‑downstream interaction: the cuts added at a node directly affect the quality of the LP relaxation and consequently the choice of branching variable.

To capture this interdependence, the authors model the interaction as a Stackelberg game. The cut‑selection agent acts as the leader, choosing a subset of generated cuts based on cut‑specific features and a global MILP representation. The branching agent, as the follower, observes the leader’s action and selects a branching variable using the same MILP representation. The joint objective is to minimize solving time and primal‑dual gap, formalized as a bi‑level optimization problem where the follower’s policy is optimal given the leader’s policy.

Collab‑Solver’s architecture consists of three key innovations:

1. Data‑communicated encoders – Separate neural encoders (MLP for cut features, GNN for the bipartite variable‑constraint graph) map both agents’ observations into a shared latent space. This enables the leader’s cut decisions to be directly communicated to the follower and vice‑versa, creating a feedback loop that aligns the two policies.

2. Two‑phase learning –

   • Pretraining: Using trajectories collected from a conventional solver, both agents are pretrained via imitation learning while simultaneously training the shared encoders. This phase provides a strong initialization and ensures that each agent can interpret the other’s data.
   • Fine‑tuning: A two‑timescale update rule is applied. The leader’s parameters are updated with a small learning rate, providing stability, whereas the follower’s parameters use a larger learning rate to quickly adapt to the leader’s latest cuts. This mitigates the non‑stationarity that typically arises when multiple policies are trained concurrently.

3. Scalable game‑theoretic formulation – Although the paper focuses on cut selection and branching, the Stackelberg framework can be extended to additional modules (e.g., node selection, primal heuristics) by introducing more agents and appropriate communication bonds.

The experimental evaluation covers eight NP‑hard benchmark sets, ranging from synthetic instances to large real‑world problems (e.g., power systems, portfolio optimization). Metrics include total solving time, number of explored B&B nodes, and final primal‑dual gap. Collab‑Solver consistently outperforms state‑of‑the‑art learning‑based baselines: it reduces solving time by roughly 15‑20 % and cuts the number of nodes by about 10 % on average. Importantly, the method exhibits strong out‑of‑distribution generalization; when tested on instances with different variable counts or constraint densities, performance degradation is minimal, indicating that the shared encoders capture problem structure effectively.

Ablation studies isolate the contributions of each component. Removing data communication (training agents independently) leads to a 12 % increase in solving time. Replacing the two‑timescale scheme with a single learning rate causes unstable training and a 8 % performance drop. Limiting cuts to the root node (as in a prior collaborative work) also harms performance, confirming the benefit of allowing cut‑branch interactions throughout the search tree.

Contributions:

• Formalization of cut‑branch collaboration as a Stackelberg game.
• Introduction of a data‑communication pretraining phase and a two‑timescale fine‑tuning regime that together stabilize joint policy learning.
• Empirical demonstration of superior solving efficiency and robust generalization across diverse MILP datasets.

Limitations and future work: The current implementation only addresses two agents; extending to a full suite of solver modules will require a more complex multi‑leader/multi‑follower game model. The reward function is primarily time‑centric, so incorporating additional objectives such as memory usage or solution quality could further improve practicality. The authors suggest exploring meta‑learning for rapid adaptation to new problem families and investigating multi‑objective reinforcement learning to balance competing performance metrics.

In summary, Collab‑Solver represents a significant step toward truly collaborative, learning‑driven MILP solvers, showing that joint optimization of interdependent heuristics can yield measurable gains over isolated learning approaches while maintaining strong transferability to unseen problem instances.