Title: Towards a self-driving data center with a focus on AI-driven supply chain management


Technical Area: Machine Learning



The use of the cloud services has grown tremendously, and investment in related infrastructure of IT equipment and software has exploded. IDC is the infrastructure cornerstone of Alibaba’s ecosystem, is aiming towards an automatic delivering, proactive optimizing and self-evolvement data center, with the ultimate goal of making digital economy constructed easier. We take the lead in building the “digital, automated, and intelligent” operations and management systems for the global data centers, continuously exploring the optimal solution for the data center’s supply chain management, while guaranteeing the stable, reliable and effective operations for the rapid development of Alibaba’s business. That sort of rapid growth and unpredictability presents challenges across today’s data center supply chain, including:

- Quick delivery of orders with complex product requirements;

- Orders with small quantities and several different product types

- More emergent and personalized requests


The benefits AI technologies bring to supply chain planning are the ability to provide speed and accuracy beyond human capability. Supply chain optimization isn’t a collection of disconnected functional modules. It looks at the supply chain as a whole, not just individual links in a broken chain. Adding AI on top of this foundational backbone will help IDC supply chain soar to a completely new realm by realizing benefits and value on a much bigger scale. It all starts by building out extended AI functionality in the areas of end-to-end visibility, demand forecast, inventory planning, scenario simulation and resource allocation optimization. You’ll know sooner if something’s gone amiss and be able to act faster to correct it, reducing decision latency and improving operational and financial performance. That’s a revolution in supply chain planning on its own.



This project is to optimize IDC supply chain management, support dynamic scenario demand planning, inventory optimization and resource allocation.


The measurable outputs of the research may include but not limited to:


Related Research Topics

1. Demand Forecast Model:

Over the last few decades, traditional time series forecasting methods such as moving average, Holt-Winters methods, exponential smoothing, Box-Jenkins ARIMA model, Bayesian dynamic estimation, and multivariate regressions have been proposed and widely used for long-term supply chain demand planning. These methods are chosen because of their ability to model trend and seasonal fluctuations present in demand data.


Some machine learning techniques, including artificial neural networks and support vector machines, have been used to forecast demand in supply chain. However, how to improve the quality of long-term forecasts is still an outstanding question in both theoretical and industrial fields.


Companies have tried for years to optimize production and inventory deployment decisions through more effective forecasting of customer (retailer) and consumer (end-user) demand. Too often, though, the forecasting models they use have contributed to, rather than reduced, uncertainty. With the increasing complexity of demand drivers such as the timing and type of promotional activities employed, and competitors’ pricing moves, a robust and dynamic forecasting model is needed to incorporate a customized set of drivers and compensate for the additional volatility that these factors inject into the forecast. A better demand forecast model could be used to inform better production and inventory deployment decisions. It also enables the company to highlight the inefficiencies that behaviors bring to supply chain performance.


2. Inventory Optimization:

Since forecasting customer demand can be challenging, companies often add inventory to protect against inaccurate forecasts. Inventory planning is a critical aspect of enterprise- wide optimization. Inventories are used in production and logistic networks to coordinate supply cycles and to mitigate the risks associated with uncertainty. The optimal inventory policies for the most recurrent inventory management problems have been studied for decades. The importance of inventory management in industrial applications derives from the effect of stockouts in the levels of customer satisfaction and the impact of stock in the economic balance of companies. The potential savings from stockout prevention and inventory related cost offer a huge opportunity for optimization.


Many optimization models have been developed for supply chain planning because they offer the possibility of finding strategies that lead to greater economic benefits. The traditional models have focused on finding the optimal decisions of the supply chain planner in a deterministic context. Many optimization strategies have been proposed to manage inventories since Harris introduced the Economic Order Quantity (EOQ) model in 1913. The EOQ model was developed to balance ordering and holding cost for problems with a deterministic demand rate. Other classical models for inventory management with uncertain demand include continuous-review (r, Q) policies and periodic-review basestock policies; the main purpose of these models is to minimize the expected cost of replenishment and stockouts, since complete satisfaction of uncertain demand might be too expensive or impossible.


However, they all deal with the question of determining the replenishment strategies for a single inventory system subject to demand. These models consider linear cost functions for ordering, holding, and stockouts with or without fixed charges. Some models include lead times, but none of them has capacity constraints and uncertainty from other sources. The only source of uncertainty considered in the classical models is given by stochastic demand.


It is widely recognized that uncertainty and external decision-makers play a fundamental role in the economic success of industrial supply chains. Therefore, a supply chain optimization model in realistic industrial environments are needed to mitigate the risks and disruptions faced by supply chain planners with respect to the future performance of the system.