Traditional Machine Learning for Allocation Planning in Supply Chains

This article shares a practical use case of a traditional machine‑learning algorithm in a business scenario, focusing on allocation planning within a supply‑chain context.

After a brief personal reflection on deep‑learning models such as LSTM, the author argues that many classic machine‑learning methods have solid mathematical foundations and are worth revisiting.

The core problem is defined as a "transfer plan" (调拨计划): given an over‑stocked warehouse A and an under‑stocked warehouse B, determine how much of each product should be transferred to balance inventory, reduce cross‑region shipping costs, and respect capacity constraints.

The model includes:

Variables representing the transfer quantity of each product from A to B.

Constraints ensuring that the transferred amount does not exceed available stock at A, does not exceed demand at B, and stays within transportation capacity.

An objective function that minimizes total logistics cost, which consists of shipping cost, cross‑region shipping cost, and transfer cost.

Solving the model can be done by exhaustive enumeration for tiny instances, but the solution space grows exponentially for realistic problems. Therefore, heuristic search methods such as simulated annealing, genetic algorithms, and ant‑colony optimization are introduced as practical alternatives.

The article illustrates the genetic algorithm workflow (selection, crossover, mutation) and mentions that commercial solvers like CPLEX can also be employed.

Finally, the author contrasts the traditional optimization pipeline (problem definition → modeling → solving) with the deep‑learning pipeline (vectorization → loss minimization), emphasizing that both approaches have their place in solving complex business problems.