Supply Chain Science: applying AI and ML in 2021
- 5 January 2021
- By Manhattan Staff
Today, people and organisation often use the terms AI and Machine Learning (ML) so interchangeably that it is difficult to discern where one begins and the other ends. And while AI and ML are synonymous with each other, it’s important to recognise that these technological cousins are not the same thing.
AI is a branch of computer science tasked with building machines capable of intelligent behaviour, while the team at Stanford University in California, define ML as; “the science of getting computers to act without being explicitly programmed”. While there will always be a need for AI researchers to build the smart machines, it is the ML experts that make these machines truly intelligent – that being said, it’s easy to see the ambiguity between the two!
For us at Manhattan, ML sits at the intersection of data science, computer science and statistical mathematics: it is the ability for systems to automatically learn and improve analytical tasks from experience, without needing to be explicitly programmed.
So, how does ML fit in with AI? And more importantly maybe, how does it actually work in the context of supply chain science? Think of it like this. If AI is essentially the intelligence, ML is the implementation of the compute methods that support it. ML is the workhorse and enabler of AI through its algorithms which provide systems with the ability to automatically learn and improve from experience without being explicitly programmed.
David A. Kolb said in 1984, that: “learning is the process whereby knowledge is created through the transformation of experience.” Regardless of whether learning is human or artificial, learning is driven by a requirement to improve upon a previous performance or experience.
Once we can recognise the difference between the two terms and pick-apart Kolb’s key point regarding the process of learning, you can apply both AI and ML to a number of highly complex processes (including TMS, WMS and even omnichannel scenarios) present in modern supply chains.
When we’re considering the application of ML to supply chains and the primary building blocks of this applied supply chain science, there are three parameters that underpin work in this field.
First, supervised learning is an important aspect of ML for predicting things when data and features are clear. The ‘supervision’ aspect of it comes from knowing the data and the aim.
For example, you could use this to predict the price of a house. Since we know things like square footage, number of bedrooms/bathrooms, and because we know the details of a number of other houses that sold in the same area, we can therefore use that data as our ‘training’ set for supervised learning to make new predictions.
In the context of supply chains, a retailer could use supervised learning to more accurately predict peak season shopping trends through the use of historical customer and website traffic data, and even potentially recent social listening and weather forecast data – this in turn can help retailers to more effectively fine tune their Omni-inventory strategy around peak times.
Secondly, unsupervised learning is used to look for groupings, patterns, or relationships within data, especially when we have little to no real idea of what we are looking for.
Consider this, there is no denying that the number of promotional events has increased tenfold, and it is more difficult than ever before to predict the impact each promotional campaign will have on demand. Having access to this type of performance insight is crucial to growing profitably, but it is not easy to get it.
With a multitude of promotions often running simultaneously, it is even more challenging to determine which promotional activity will drive demand, so how can you accurately predict how demand will change?
Manhattan’s Demand Forecasting, with promotional planning, uses the most advanced data science and machine learning models so that you can extrapolate the value of your hidden data and create the most accurate forecasting models possible.
Finally, reinforcement learning is where you have little or no data but have an environment to interact with.
For example, think of some type of agent navigating and interacting with the environment to try and achieve an objective. The environment provides either a reward (like getting closer to the objective) or a penalty (like getting further from the objective) for each agent decision – which it records and ‘learns’ from.
It’s the same technique used in some chess simulations. Or you could think of a robot arm or cobot (such as those you would find in the Manhattan Automation Network), that is picking or moving items in a warehouse, learning about the size, speed and weight of different items to become more and more efficient in its picking patterns.
Through the process of analysing large amounts of historical observations (data), of a given analytical task and environment, ML can help to solve or improve solutions to key challenges within supply chains that are often simply too large or complex to mathematically model accurately, or to code precisely.
Just like you wouldn’t use a hammer to drive a screw, or a rake to dig a hole, the same principle applies to data science tools as well. ML is a great development, but it’s important to be aware that there are many types of analytical problems to solve across such complicated and fluid environments as the supply chain, each one requiring a specialized scientific and approach to reach the best solution.
Applied machine learning is already a powerful tool and will become increasingly important over the course of the decade as we continue to need technology that can listen, learn and adapt on its own – right now, ML provides one of the best opportunities in the supply chain to achieve that.
Now consider AI. According to a recent global study from McKinsey, adding AI to supply chains is already delivering tangible benefits for companies putting it in place.
The study found research 61% of executives reported decreased costs and 53% reported increased revenues as a direct result of introducing AI into their supply chains. And, more than one in three reported a revenue bounce exceeding five percent.
Areas generating revenue in supply chain management include sales and demand, forecasting, spend analytics, and logistics network optimization such as the warehouse and transportation spaces.
Consider for a moment the warehouse environment. A warehouse is often a chaotic place and an aisle that is clear one minute, can be obstructed the next, causing significant challenges for human workers and their cobots – AI can offer a practical solution here.
AI can train a cobot’s vision system to recognize and react to its environment. In supervised training, a robot can be repeatedly shown pictures of various obstructions, from torn cardboard of different sizes, shapes, and colours in different lighting conditions, right through to other obstructions such as people or other cobots.
After a period of time cobots learn to recognise obstructions of different types, in different scenarios and ‘learn’ independently how to navigate them safely. It can also deliver deep analysis down to the SKU level, not practical with a manual, human-led approach.
There are other applications of AI in supply chains, beyond robotics however, with end-to-end visibility and actionable insights being two of the key areas.
In today's complex networks of supply chains, where data is created and collected from numerous sources, an AI platform can create a single, ‘live’ virtualized data layer; in the process revealing cause and effect, bottlenecks and opportunities for improvement in real-time.
In terms of actionable analytic insights, AI can sift through large amounts of information to discern patterns and quantify trade-offs at a scale, far beyond what's possible with conventional human-based systems.
Finally, when it comes to informed decision-making, AI can provide far more accurate prediction modelling to improve supply chain performance. It can also provide implication-based forecasts across various scenarios in terms of time, cost and revenue, and by acting autonomously over time, it can continuously improve recommendations as conditions and variables change.
In the 2019 Hype Cycle, Gartner highlighted that applied ML and AI is still very much on an upward trajectory in terms of its market and application maturity, however, many businesses are still yet to plug into these opportunities.
However, today at Manhattan, this is not the case, and ML and AI are both very much science fact, rather than science fiction, with many of our ML/AI solutions already delivering demonstrable and positive impacts for customers all over the globe.
It is important to recognise that AI and ML are not a panacea for every analytical or technological challenge facing businesses today. However, it is worth noting the considerable benefits both disciplines are already bringing to modern-day supply chain environments.
We have seen the positive effects that AI and ML can have first-hand when used in conjunction with automation, robotics and even human processes over the last eight months of the global pandemic in particular.
Whether it’s protecting your workforce, streamlining your supply chain networks, keeping your brand promise to end-customers, helping your business to be more environmentally aware and sustainable, or simply helping to improve bottom-line profits, AI and ML can help and they’re here to stay, so surely now is the time to ask yourself: ‘are we machine ready?’