How Are Supply Chains Optimizing 45K Variables and 100K Constraints

The way a solver approaches that search depends on the structure of the problem. Some methods aim to prove mathematical optimality. Others prioritize speed or feasibility. Increasingly, modern optimization platforms combine multiple techniques to balance solution quality and computation time.

Think of this as the workhorse of supply chain math. It finds the best possible answer to any problem where costs and volumes scale in straight lines: double the shipment, double the cost. The solver searches for a defined region of possibilities and always finds the true optimum.

The formula says: find the values of your decision variables x that minimize total cost, given a set of linear rules:

The constraints draw a fence around what is physically possible. The solver finds the lowest cost point inside that fence. Because everything is linear, the fence is a clean geometric shape (a polytope), and the optimal answer always sits at one of its corners, which algorithms like Simplex find very efficiently.

The real world rarely scales in straight lines. Energy costs curve upward as machines approach capacity. Supplier discounts kick in at volume thresholds. When those curves matter, you need nonlinear optimization, which handles any smooth mathematical relationship, at the cost of more computation.

The formula is the same in spirit, but f and g can now include powers, curves, or products of variables:

The solver still searches for the best point, but now the “fence” can be curved, making the search harder. For convex curves (bowl-shaped costs like the quadratic above), it can still guarantee a global optimum. For more irregular shapes, it may find a very good local answer without being able to guarantee it is the absolute best.

Sometimes the problem is simply too large to solve exactly in reasonable time, for instance, routing hundreds of vehicles across a city. A heuristic does not try to prove it has found the best answer. Instead, it follows a smart rule to build a very good answer quickly.

The nearest-neighbor rule works like this: start at the depot, always go to the closest unvisited stop next, return when the vehicle is full:

The constraints are practical, respect vehicle capacity C and serve every stop, but they are enforced incrementally as the route is built, not solved globally. The result is a feasible, serviceable plan in seconds. It will not be perfect, but it will be good enough to operate from, and it scales to problems that exact solvers cannot touch.

Metaheuristics take the heuristic idea further: instead of building one solution with a greedy rule, they maintain and evolve a whole population of candidate solutions, gradually improving quality over many iterations. Genetic Algorithms (GA) are the most intuitive examples.

Imagine you have P candidate supplier networks, each represented as a list of on/off decisions across supplier–country pairs. Each candidate is scored by a fitness function that combines what matters:

Over hundreds of generations, the population drifts toward high-scoring regions of a solution space that would be impossible to search exhaustively (2ⁿ combinations for n suppliers). No optimality guarantee, but the progressive improvement is systematic rather than random.

In some supply chain problems, particularly in manufacturing scheduling, the primary challenge is not “find the cheapest plan” but “find any plan that works at all.” Regulatory cleaning requirements, equipment certifications, and batch sequencing rules can interact in ways that make most schedules infeasible before cost even enters the picture. Constraint programming is designed precisely for this.

The model defines variables (when does each batch start), their domains (which time slots are possible), and hard rules that must all hold simultaneously:

The solver propagates these rules aggressively. Every time it fixes one variable, it immediately eliminates impossible values for all related variables, shrinking the search space before trying the next decision. When no valid assignment exists for some variable, it backtracks and tries a different branch. Finding a single feasible schedule is treated as the win. Cost reduction comes only after feasibility is confirmed.