AI Now Plans Factory Orders Better Than Humans Can

A new arXiv paper tackles a difficult inventory-planning problem: deciding production quantities for multiple items with limited capacity when demand quantities are known but the timing of demand arrivals is uncertain. The authors model this as a discrete-time Markov decision process and show that stochastic timing significantly increases complexity, including more states, transitions, solution time, and memory use. They then propose a genetic algorithm that searches over inventory policies. In tests on 330 benchmark instances, the algorithm stays close to the exact stochastic solution when that solution is available, with an average optimality gap of about 3.44%. On 90 harder instances, it remains below a 5% gap and achieves an average speedup of about 6.89 times with 95% confidence. For extremely large cases that cannot be solved exactly, the paper uses regression to estimate the speedup.