Hidden pains of globalized production: how can machine learning help

Sam Gilliam, Along, 1969

A discussion about growing but underrated supply chain pains, and the relevance of machine learning in treating those.

A. An outlook on contemporary supply chain issues

With globalization, complexity in the supply chain increased steadily in the last 20 years.

Here are some key factors that noticeably contributed to this evolution

a) Many industrial sectors have become more consolidated. This in turn fostered centralized set-up in cross-borders corporations, notably in transversal functions such as procurement

b) More risks arose from the lengthening of supply and value chains, specifically many unpredictable influences on supply times, such as port strikes, natural disasters, etc.

c) Employees are becoming increasingly specialized and departments tend to operate in silos

In a striking contrast, supply chain teams at large haven’t been equipped with tools able to cater to this new environment.

a) ERP configurations are still lengthy and expensive.

b) More generally, there was a crying lack of innovation in this space between 2000 and 2015, especially when benchmarked again other engineering or marketing activities

c) Supply chain teams, precisely the ones now facing more challenging decisions, sometime suffer from a shortfall of recognition and/or management attention.

This gap between heightened complexity and lack of resources creates tensions and inefficiencies on a wide scale — and ultimately massive hidden costs:

a) Stressed employees lacking time

b) Decisions involving many variables being taken suboptimally. This includes supplier selection, purchasing optimization, supply chain integration, etc.

c) Teams putting up more stocks all over the value chain, as they feel that overall risk is more difficult to manage and aim to avoid out-of-stock events

B. Description and application of machine-learning based solutions in this context

Without delving into many details, we would like to outline the three types of technologies that are most commonly referred to as Machine Learning and their differences.

a) In supervised learning, the computer is presented with example inputs and their desired outputs, given by a “teacher”, and the goal is to learn a general rule that maps inputs to outputs Ex: image recognition, spam detection. This is what is most commonly understood when machine learning is brought up.

b) In unsupervised learning, no labels are given to the learning algorithm, leaving it on its own to find structure in its input. Unsupervised learning can be a goal in itself (discovering hidden patterns in data) or a means towards an end. A typical illustration of that would be clustering, where one wants to discover the inherent groupings in the data (i.e. part of a set of molecules are drugs and part are not but you do not know which are which and you want algorithm to discover the drugs)

Both these techniques require large sets of historical data to begin with and are associated to big data.

c) Reinforcement learning. The algorithm gets rewarded or punished at a certain moment in time and infers from his parametrization at the moment in which direction it should improve itself. Solving games like go or chess, as well as video games, are classic example of reinforcement learning. Unlike other techniques it is not associated to large sets of historical data, the machine learns from itself.

Specific qualities derived from these machine-learning-affiliated technologies are particularly relevant in the resolution of the above-described dire supply chain issues, and hence actively sought-for in this context.

Here are some of these qualities

a. Rapidity. The “enhanced heuristics” derived from reinforcement learning make for very rapid computation which allows for example the production of many simulations, and then for continuous optimization. Examples of companies building on this quality in the context of industrial operations: GenLots, Trufa

b. Flexibility. The possibility to add parameters “on the go” to a model without altering its fundamental functioning means is extremely valuable. Conversely it also means that solving problems is possible even with restricted of partial input data. Ex: GenLots, Alteryx

c. Complexity. The capacity of advanced models to deal with a number of “separate” problems simultaneously proves reinforcement learning-inspired models fit for multi-dimensional environments described above Ex: GenLots, Celonis, Elementum

d. Discovery. Patterns in the data identified through unsupervised learning can be exploited — for example to fix supply chain interruptions issues Ex: Llamasoft, Trufa, Risk Method, Concentra

e. Accuracy. The ability to account for more inputs and operate in layers of networks allows predictive models built with ML — notably supervised learning — to perform better than traditional models Ex: Prognosix

C. A concrete example with GenLots

GenLots identified one specific problem which is suboptimally treated today as a result of the above considerations: order planning, which sends back to the question — when to order how much materials for the production?

This critical decision, especially relevant when the planning is required over a rolling horizon of 30 weeks or more, is most often conducted today with a combination of static tools, which over-simplify the reality and fail to consider multiple parameters, and manual adjustments.

The process is lengthy and vast quantity of SKUs are not optimized.

The challenges associated to this problem are:

a) The need to avoid out-of-stocks

b) The absence of one « neutral » entreprise-level criteria when taking this decision, because departments tend to operate in silos

c) The difficulty to standardize the approach as different parameters impact different categories of products

GenLots will typically evaluate the problem against a neutral, company-wide criteria: Total Cost of Ownership.

It will build on qualities of machine learning to consider vast amounts of parameters simultaneously, as to eliminate the need for additional manual adjustments.

Because it works fast, optimization is possible on all SKU’s and not only high runners, and because it handles complexity well the result is trustworthy.

When applied to real-world data from fine chemicals or pharmaceuticals companies, GenLots demonstrated the ability to mix standard ERP parameters, such as requirement for production, safety stocks, and inventory levels with advanced parameters such as perishability, quantity discounts, production in campaigns and many more, through a proprietary platform involving reinforcement learning parameters, whose configuration is fast and largely automated.

It resulted in outstanding results both on human and commercial levels, respectively a decrease of 5 to 8% of total purchasing costs and a gain of several hours by employees such as material planner. It also contributes to mitigate the risk, with the certitude that key parameters are accounted for, and helps manager to project strategic decisions by running simulations.


All in all it needs to be recognized that the increased complexity in supply chain structures, has been progressively generating hard-to-detect costs and pains in internal systems. Those can take the form of excessive safety stocks, suboptimal decisions on multiplex questions, inability to account for new parameters, inaccurate forecasts, excessive stress within the teams and more. While machine learning, and adjacent technologies deep learning and artificial intelligence, are sometimes abusively brandished as some-form of magical one-size fits all formula, they entail extremely useful characteristics when used with acute understanding of the underlying domain.

Mitigating supply chain pains, harmonizing operational activities, gaining granularity in global production systems form some of the most promising applications of machine learning technology, and much remains to be done.