Mixed Integer Programming (MIP)

The defining feature of MIP is the ability to model yes/no or countable decisions. Integer variables represent indivisible entities like vehicles, drivers, or stops. A binary variable (0 or 1) can indicate whether a specific vehicle traverses a specific arc in a route. This contrasts with Linear Programming (LP), where fractional solutions are allowed. Without integer constraints, an LP might suggest assigning 0.7 of a truck to a route, which is physically impossible. MIP enforces logical, real-world feasibility.

The primary exact algorithm for solving MIPs. It systematically partitions the problem into smaller subproblems (branching) and solves their LP relaxations. Key mechanics:

• Relaxation: Temporarily ignore integer constraints to get a lower bound (for minimization).
• Pruning: Discard a branch if its bound is worse than the current best integer solution.
• Incumbent: The best feasible integer solution found so far. This creates a search tree, efficiently eliminating vast regions of the solution space without explicit enumeration.

A complementary technique to branch-and-bound that tightens the LP relaxation. Cuts are additional linear constraints added to the formulation that are redundant for integer solutions but slice off fractional regions of the feasible space. For example, in a VRP, a subtour elimination constraint prevents a solution where a vehicle visits a small cluster of customers without connecting to the depot. Modern solvers use a branch-and-cut framework, dynamically generating cuts at each node of the search tree.

A MIP model is defined by three components:

• Objective Function: A linear expression to minimize (e.g., total travel distance, cost) or maximize.
• Constraints: Linear inequalities or equalities defining the feasible region (e.g., vehicle capacity limits, time window adherence).
• Variable Bounds: Domains for each variable (e.g., binary, integer, continuous). The art of MIP formulation lies in translating a real-world logistics problem into this precise mathematical structure. A tighter formulation with a smaller integrality gap solves significantly faster.

For large-scale problems, proving absolute optimality can be computationally prohibitive. Solvers track the optimality gap: the relative difference between the best integer solution found (incumbent) and the best remaining theoretical bound. A solver is typically stopped when this gap falls below a set tolerance (e.g., 1%). This provides a provable guarantee on solution quality—a key advantage over heuristic methods like Genetic Algorithms, which offer no such mathematical warranty on how close a solution is to the true optimum.

The performance of MIP models is heavily dependent on the underlying solver engine.

• Commercial: Gurobi and IBM CPLEX are industry leaders, offering highly optimized presolve routines, heuristics, and parallel branch-and-cut implementations. They often solve problems orders of magnitude faster.
• Open-Source: SCIP and HiGHS provide robust alternatives, with SCIP excelling in constraint integer programming. Solver selection is a critical architectural decision, balancing licensing cost against the required solve time for operational deployment in real-time routing systems.

Real-world logistics rarely optimizes a single metric. Multi-objective MIP models simultaneously balance conflicting goals:

• Minimize total distance driven
• Maximize on-time delivery percentage
• Minimize fleet size utilized Solutions are evaluated on the Pareto frontier, where improving one objective necessarily degrades another. Techniques like weighted-sum scalarization or epsilon-constraint methods convert these into solvable single-objective MIPs.