Supply Chain Optimization Modules

Traditional demand forecasting systems have struggled to overcome the trade-off between speed and accuracy. For example, trying to improve the accuracy of daily forecasts for a large number of products or stock keeping units (SKUs) requires enormous computation time, making it impossible to complete the calculations in a timely manner. As a result, the current situation is that companies either settle for rough weekly or monthly forecasts, or use simple calculations such as exponential smoothing or ARIMA for forecasting.

With our system that integrates automated machine learning (AutoML) and automated optimization (AutoOpt), high-precision forecasts can be achieved in a short time by combining AutoML, which tries various machine (deep) learning methods to select the most accurate one, and AutoOpt, which selects the appropriate method based on data on past demand forecast calculation time and accuracy. Furthermore, the probability of method selection can be automatically adjusted by monitoring measures of forecast variability and instability.

Our inventory optimization solution targets multi-echelon inventory across the entire supply chain. While single-echelon inventory models can be easily solved using inventory formulas, solving multi-echelon, network-type inventory models requires state-of-the-art technology.

However, significant cost savings and improved service rates can be expected compared to applying classical inventory formulas to each echelon separately. For safety stock placement, which is a mid-term decision-making model, we have developed a fast solution method that can solve larger-scale problems than conventional methods. For short-term inventory policy optimization, we have developed a new method that combines simulation and deep learning to calculate the optimal dynamic inventory policy.

Our Logistics Network Design (LND) solution utilizes MOAI (integration of machine learning and mathematical optimization) technology, which enables faster and better solutions compared to approaches using only traditional mathematical optimization solvers. Typically, for large-scale LND problems, even using state-of-the-art mathematical optimization solvers can require several hours of computation to obtain a solution with an error of around 10%. As a result, it is often impossible to perform sufficient "what-if analysis" or scenario analysis needed for decision-making.

In recent years, there has been an increasing number of cases where stochastic optimization and robust optimization, which take into account data uncertainty, are applied to LND problems. These modern uncertainty-aware optimization methods are known to be more difficult than normal deterministic optimization, and without MOAI technology, it becomes impossible to solve large-scale problem instances. By incorporating MOAI technology into stochastic (distributionally robust) optimization, highly accurate solutions can be calculated in a short amount of time for large-scale problem instances.

Production scheduling optimization, which considers the trade-off between product delivery times and lead times, the trade-off between setup costs and inventory costs at multiple stages, machine uptime, work burden constraints such as employee breaks and travel, and production sequence constraints, cannot be solved with simple rule-based heuristics that program the rules of the shop floor.

Our production scheduling optimization solution is not just about scheduling; it is a fusion of several optimization models that can optimize the entire production process within the factory.

ur delivery optimization solution combines state-of-the-art optimization methods to calculate optimal solutions for medium-sized problem instances with hundreds of customers, and approximate solutions with a few percent error for large-scale problem instances with tens of thousands of customers. Furthermore, as a result of over 30 years of joint research with many real companies, it is designed to allow various constraints necessary for practical use to be added without compromising solution performance or speed.

As an interesting case study, we have a problem of replenishing vending machines, taking into account product inventory over multiple periods (thousands of customers for 30 days). We have successfully achieved cost savings of over 30% for this large-scale problem instance. In addition, we have developed MOAI (integration of machine learning and optimization) technology that uses machine learning to quickly calculate highly accurate solutions by storing pairs of past data and solutions.

This is an analysis tool for proactively addressing disruptions in the supply chain due to disasters and other events, and supports decision-making for building a more resilient supply chain.

MOAI Supply Chain Risk Analytics integrates bill of materials, production information (which factory manufactures which parts), and transportation information (which factory transports to which factory), and performs optimization-based analysis to identify parts and their production factories that have a significant impact from disruptions.

We also offer dynamic pricing solutions that consider the supply-demand balance and corporate profits for goods that decrease in value after a certain period, such as airplane seats, train and ship seats, hotel and inn rooms, rental cars, baseball tickets, television commercial slots, coin parking lots, food, home appliances, and fashion.

Using a state-of-the-art algorithm that combines AI prediction and optimization, we automatically perform not only prediction and rule-based supply-demand balance adjustments, but also revenue maximization simultaneously. In addition, we can handle various practical problems such as considering customer psychology in response to price fluctuations (prospect theory from experimental economics), consecutive stays at hotels, transfers of airline tickets, and consecutive numbered seats.