Video: A Hitchhiker's Guide to Machine Learning and Artificial Intelligence in the Supply Chain

Carter McNabb discusses Machine Learning and Artificial Intelligence (AI), and the significant opportunities for Australian supply chains.

Carter McNabb speaking about Machine Learning and AI in Supply Chains at ASCI2019

In this keynote presentation at ASCI 2019 Carter looks at how Machine Learning and AI are currently being used in the supply chain, as well predictions for their use moving forward.

His presentation explores the emerging technologies using Machine Learning and Artificial Intelligence (AI) that could potentially offer significant opportunities for Australian supply chains.

Furthermore, he explains their potential across all areas of the Supply Chain from Strategy, to Planning and Execution.

Transcript

Ivan Imparato: I've got my good friend Carter McNabb here somewhere. There he is. Carter will talk to us about the Hitchhiker's Guide to Machine Learning and Artificial Intelligence. Carter is a founding member of GRA, Australia's premier supply chain consulting firm, specialising in supply chain strategy planning and execution. Over its 20 year history, GRA has helped 200 plus organisations across multiple industry identify combined savings of more than $10 billion. I said it right, Carter, billion. GRA's focused on practical results delivery and its mission is to turn its clients' supply chains into a competitive advantage. Carter has over 20 years of international supply chain advisor and transformation experience and has taught at Masters level with Monash University Logistics and Supply Chain Management Programme. This series of published articles and white papers, press quotes and frequent speaking engagements, Carter is recognised as an expert in the field. So, ladies and gentlemen please welcome Carter McNabb.

Carter McNabb: Machine learning. The reason why I've called this A Hitchhiker's Guide to Machine Learning and Artificial Intelligence is it's still kind of a nascent topic. There's a lot of hype and a lot of excitement, but the question is, what is it? Where does it fit? How does it apply? And so this is my best attempt to try to create a little bit of a roadmap for what that looks like. I'm noting that it is still early days, it's here, but it's still an emerging technology.

And so we'll explore that a bit. What I'd like to do is start these discussion with just a quick show of hands that says if you've done anything intentionally with AI or machine learning in your supply chain, would you please raise your hand?

It's funny, what I've noticed is it doesn't matter how many people in the room, it's always two people. I was looking for percentages, but if it's 10 people it's two people, and if it's 100 people it's two people. So apparently in any given room. So this is machine learning, right? Artificial intelligence. Oh dear, okay. So before I jump to the, kind of the applications in supply chain, let me just make sure this works. Yay. I just wanted to point out a couple of things that sort of underscores the importance of this from a business perspective.

I also want to make a comment that for me, technology is neither good nor bad. It's the intention of how it's used that determines actually what the implications are in the world.

So, some of the things in AI machine learning, I look at it, I get a shudder. It's like, oh, I don't know about that. So it doesn't all sit totally comfortably with me, but the reason why I'll point it out is that, look it is here, and the question is, you know, how do we embrace it and deploy it in a way that actually serves humanity as opposed to, yeah, treating people like people and not like things?

I think is very important. Speaking of people, there's a lot of us.

I remember flipping through a National Geographic several years ago and there was a picture in the National Geographic, and it had a graph that kind of went like that. And what it started it with it said, it was one AD, so height of the Roman Empire. There were 200 million people on the planet. And the graph went like this went up and I think by 1880 then we had a billion people. So let's call it two thousand years later we had a billion. And so, now it's, what's the year again? 2019. And we're sort of seven and a bit billion. So, a seven time, population's increased by factor of seven times since 1880. There are some people alive who have seen the population triple or quadruple in their lifetime, so that's, I used to say that's unprecedented, but I'm pretty sure when, I think it was, was it?

I'm going to get this probably wrong. I think it was like 125,000 years ago. Human beings got down to apparently 30 pairs, so it's about 60 Homo Sapiens left. Southern Africa, all of us are descendants from those people and, in fact, if you trace our mitochondrial DNA back, we all come from one of seven women.

So, we are actually literally all related. It is a big family, believe it or not. It's just, it's the truth. So, why am I pointing this out? I guess the other point is that at 2045, it's supposed to level off about around nine billion. But, keep in mind that there's no more water. There's the same amount of water, there's the same amount of resources and we are consuming much more than we used to.

So not only is there an exponential increase in population, there's an exponential increase in consumption. Business might say that's wonderful, the sustainability view might say that's concerning. But, I think our challenge is we have to find a way to bring economical and ecological considerations together. They can't be in opposition, they can't be separate because they're quite intertwined. So look there are a lot of us running around.

The next point I'll make is, and we leave a giant wake of data with us as we run around these days. So, this was another National Geographic magazine article. I've just taken a scan of it. And this was several years old, but this was an article describing what MIT labs could predict, with a very high degree of accuracy, just by looking at search engine behaviour. And, you know one of the bubbles on there says, "Pregnant and doesn't know it yet." So, they were able to work out if someone was pregnant by just observing changes in search engine behaviour, and again they were able to detect this before the person who was pregnant recognised it. Now, that's interesting. Again is that good or bad? You know, sort of depends. Depends on how that's used. But this is the level of insight that can be generated from data these days. And again, this was a few years old. And again, this is proliferating quite exponentially.

So we've got a lot of folks running around. A lot of data as well. And then on top of that, in the world some of us live in, is the supply chain world, I'm not even going to come close to going through those individual dots bit by bit. But the other point is, supply chains have got a lot more complex. In some ways supply chains have gotten more fragmented, more granular, more particulate. We've got many, many more sourcing options than we used to. And if you look at the number of fulfilment options and possibilities we have now than what we had say 10 or 15 years ago, it's been an explosion.

So, product proliferation, shorter life cycles, more trends, more competition. You've heard all that before. But I guess what I'm saying is that the task to try to manage all this is getting a little bit more challenging. And, I think that's why something like machine learning and its possibilities may become a fundamental way that we do things moving forward, because with that amount of data, with the amount of options and complexity that there is in supply chains, the ability to efficiently extract an insight from all that data becomes a really key point. It is going to change the way we work. I've been asked the question a few times, well you know, because we're actually seeing roles. 

I was at another talk a couple weeks ago and there was a guy from a recruiting company, he said, "We're seeing requests for these sorts of roles." And I said, "Well what is it?" He said, "It's a supply chain that's on architect." I said, "What do they do?" He pulled it out. He said, "They have to have a..." What was it? It was a mathematical understanding of how to solve complex supply chain problems, strategic insights, and you know business concepts, and macroeconomic considerations. They have the ability to understand warehouse and DC layout design based on fulfilment options. They need to understand the different automation technologies, build a business case, and see it through to implementation." I said, "Have you found anybody like that?" He said, "No, we haven't." I was like, "Neither have we." I mean 'cause, I've never met one person who has all those skills. I just haven't seen that yet. Not at a suitable level of depth.

So, I know there's a lot of interest in like recruiting of data scientists and these sorts of things. Again, a good skillset, but remember that you know, maths is a philosophy, it's a language. Statistics is a science built on the language of maths. So just because I'm good at working with numbers doesn't mean I understand the art of its application in a particular industry.

So, it's good to have a data science, it's better to have someone who understands how to use data to solve a problem and has industry experience and knowledge. But I think all that stuff is important. And look it is a disruptive technology, and I'll show you some of the examples here with a couple little videos, which give me a break and take the attention off me for a a little while which, believe it or not, I do like.

So, what is AI and machine learning? At a really high level, it's a mathematical construct that looks for relationships in data. It creates a model and uses that for future prediction. Another way of saying that, it's something that looks at the relationship between correlation and causation.

So, I'll do some more artificial intelligence and machine learning right now. I can deduce clearly that breakfast presentations create darkly coloured jackets. I mean, it's factually true, there's correlation here right? We're all here at a presentation, and there are a lot of people wearing dark clothing. So therefore, breakfast presentations. So if I were to set up another breakfast, I could expect to see lots of darkly coloured clothes, right? I mean, so there's a correlation, but did it cause that?

So, that's one of the tests with these things is, is the data actually meaningful in terms of actually a predictive piece. And machine learning is trying to establish those sorts of things. The relationship between correlation and causation. One of the things that's exciting about it is that, having done a lot of work in supply chain design and planning, you know a lot of times what we're told with you know, forecasting and planning is, you know, it's great to run a promotion on this item, but I have a sense that a few weeks later the whole category takes a hit. Like, in other words, it's not just, so yes the promotion worked. I call that vertical. This is how I think. And I look at that like as a vertical issue, but when I drop that vertical issue in the water, it creates horizontal ripples. And the rest of the product range and historically, we haven't had a really great way of analytically dealing with that.

Machine learning now, and AI, gives us the ability to understand a broader context. Because really what it's doing is it's looking at multiple data sets that don't have to be formatted the same way, they can be disparate. But as long as there's some consistency between the data sets, it's looking for relationships and saying, "Aha, these are the things that actually "make a difference. "And now that I know that, I can focus on it."

The other thing about machine learning is that it can be very data hungry, so you don't necessarily want to try to boil the ocean every time you want to make a piece of toast. You know, the trick is, what are the things that are actually relevant? So a lot of the idea with proof of concepting with machine learning, is to figure out what really matters.

We did some work with a convenience retailer and, you know, it was something around what happens when you change the price of the 1.5 litre? When you bring it down, what we started to see is people would switch out of the 600 mil and start buying the 1.5. But there were 24 different permutations of price points and so on in relationships, but we found only two of the 24 mattered. Only two of the 24 actually, so then now we know where to spend our time. We're not worried about all 24 of those, just those couple. So what are we seeing in the market? Very much still in the innovation cycle in the sense that you know, it's a new, shiny thing. It is exciting, but what I'd offer is that machine learning and AI isn't an opportunity to skip steps. So, in other words it's like, we don't have this capability so we'll just do AI machine learning and that'll solve the problem.

I describe AI and machine learning as like the top of the wedding cake. You have to have all of the other layers before it really starts to add value. And I'll explain a little bit more about what I mean in a minute.

But it's still early on, but supply chains are data rich, we can get data from lots of different points. And this is the, one of the challenges I find with AI machine learning is that there are so many possibilities and few are eventualities right now. I think that'll change, but what I want to do today is  try to frame that up just to give you a sense of that landscape.

So in terms of, this is just one of our views of how we see supply chains. There's a lot of information on it. I like to think of this in three chunks. There's a strategic element, which basically says, if I know my business strategy, what my customer offer is, I can then come up with my network strategy and build my supply chain based on the business strategy and the customer value proposition.

And in fact they should be linked 'cause my customer value proposition should inform my supply chain structure. So, at a strategic level, that's kind of that piece.

Generally speaking, decisions at that level tend to be longer term. They tend to be higher level, and the data that we use tends to be less granular. It's less detailed, right? Because if I'm trying to make a decision for 15 years from now, I'm not necessarily concerned how long it takes to drive from this street to this street at Tuesday at 9:20 in the morning, for this particular problem. So it's a design question.

The integration, I'll just talk to it, 'cause it's there. But the integration piece is then round the organisational constructs that are required to bring this stuff to life. So your S&OP processes, your roles and responsibilities and policies and so on. But I'm not going to spend much time on that.

The next level's the planning level. So I've got my supply chain, but now I need to plan it. So forecasting, inventory planning, replenishment planning, S&OP, transportation planning, and so on. And again, there tends to be again, more data that I use. It tends to be more sort of short to medium term.

Whereas then I drop down to the execution section, which is really around, you know, looking at warehouse fulfilment, warehousing processes, manufacturing processes, and so on. Things are starting to get a lot more granular. Much shorter term horizon, but a lot more granular.

So I just framed that up in terms of those three kind of horizons, or levels, and then how it fits in. I guess the interesting thing is when pulling the presentation together, because literally it was just like, what would you like to present? I was like, "I don't know, why don't we do this?" And we spelled it out and me and the team started looking through it, and the funny thing that we found is there's not a lot of great examples of machine learning and AI in the strategic realm.

There's some in the planning, a lot more in the execution, but not much in the strategy piece yet. That doesn't mean it isn't coming, it's just not as well developed. So jumping into that. If we look at those three horizons, and this is what are some applications, what might they be?

The other thing about this is get creative and start maybe thinking about this in your own organisation. The way I look at this is, do you have a hunch that something may be related to something else? And you've always known, I bet these things are related, or you know when we do this, this sort of thing happens. Because it's really interesting to then try to get that data and do a proof of concept with AI and machine learning to see if that actually, if the hypothesis is true. And you can proof of concept it and do it in sort of a controlled way. But one might be population and demographic data to preposition inventory and strategic locations.

There was an article in the AFR. I'm sorry. I'm just going to have to say this. I mean no offence. I acknowledge that it may cause it. But the fact that the AFR has a title that says, "Woolworths learns not to mess with Plumpton's love of Doritos," and this is what we're using AI and machine learning for, sometimes I just have to chuckle. It's like, really? That's okay. But the point about this was is that if you really look at demographics and you really look at preferences, and I guess supermarkets have been doing this for a while, which is kind of regional or kind of locale based ranging, but really starting to get some of those customer preference attributes and using that to actually come up with the stocking policies.

So not just demand planning, not just an allocation process, but what we're actually seeing in the kind of dynamically range that fits in a particular demographic market, that's one of the applications that we're seeing there.

So that's happening. Another one might be, you know, if you're doing network planning or modelling, the way that we've done this historically, is what's the business strategy? What's your CVP? And then what are the scenarios? What happens if revenue goes up by 40%? What if it stays the same? What if it goes down by 40%? And so we usually put a bunch of what ifs in there to see how robust the recommendation is since that gives us a sense for its sensitivity.

But generally those inputs are guesses. Those are best guesses, right? So what if we now can start using machine learning to inform some of those hypotheses about the sorts of scenarios that we're going to see. What about population growth? What about the income in those populations? What about the preferences in those populations?

Now let's build that into our offer moving forward, okay. What about things like as well the projected congestion in the network as the population increases, and what does that say about delivery times and the CVP?

Might we then change the model from a DC fulfilment to a dark store idea because of the density, the location, and the expected real estate prices? So, these are the sorts of things we might start bringing into the strategic realm. 

This is an example, and I mean it's interesting, a lot of this stuff, what I've seen is a lot of this stuff is already being used in agriculture for yields. 

Why is that? I think one of the reasons for that is because I think the relationship between correlation and causation is much clearer. Nutrients in the soil, sunlight, water levels. I mean it's pretty clear that these things are related and relatable.

So what you're seeing on the slide, we've got live sensor data, for example, that talks about optimal harvest times. And this is happening. So we can look at, you know, recent rain fall, are the conditions right for harvest? Off we go.

Six months ahead, projected weather changes will impact the harvesting timeline, so we can start making adjustments six months out. Maybe 12 months out we're saying lower than expected yield rates, we might change our labour schedule, the workforce sort of distribution.

And then commodity prices over 24 months might be changing, so that might impact the decisions we're making now. And the little graphy thing that's moving around is actually showing the yield over a period of time based on a machine learning approach.

So interesting, a lot of more strategic ones I've seen so far have been more in agriculture and it kind of makes sense to me why.

Planning. We've done some work ourselves in this. Planning is really now more for if we talk about our set, say our forecasting, demand planning, inventory planning, and so on. What I said before is a great example is this notion, if I run promotion, what's the net effect on all products? Not just on the one I'm running the promotion on.

So what's the net impact of a particular piece, and imagine the power of that now in say a sort process. You know, as opposed to just looking at it as a vertical issue, it's a horizontal issue.

One of the tool sets we work with in the U.S. is integrated with FRED, which is the Federal Reserve Economic Database. And what it's doing it's looking for relationships and changes for things like interest rates, fiscal policy and monetary policy, housing starts and so on. And then looking for, as these things change, is there anything, does it look like it's having a correlation. In other words, and extrinsic impact on the demand. And as it starts learning from that, as we start to see things changes, the systems start to make automatic adjustments.

Now we can start actually getting some real data around it. When we make pricing changes, how does that actually impact things? Again, that was pretty hard to do in the past 'cause they're multi-variant, they're non-linear, You know and we just really haven't had great maths to solve those problems, but that's now here.

Things around network inventory optimisation, which would say, I've got five locations. I want to have an effective 99.9% in-stock position for the network. One way of doing that is holding inventory at 99.9 on all five locations. You guys know the inventory versus service and cost curve. That it gets exponential at that you know, those last few percentage points. So the point is, well hold on a second. If I've, do I need to do that, or could I hold 95% in all five locations, which means that my probability being out of stock is .05 to the fifth minus one, which means I get basically a 99.9% service level. And it means 5% of the time I'm distributing from a non-optimal location, but on a total cost basis, that's the lower option.

So we're getting into these sorts of capabilities, that next level of refinement and cost out.

Another one is planning process automation. We're seeing that some of these demand planning tools are actually learning, they'll actually learn from planned behaviour. So how would filling orders, so we've got a recommended order, but are we rounding up to an inner or to an outer, or are we trying to build full containers or pallets? How are we doing that? And again, over time it'll learn, but the other thing about this is it doesn't mean it necessarily has to do the work for, but it can actually make a more informed recommendation as, this is what we think you would likely do because this is what you have been doing and this is what we've prepared for you.

Or you can automate it to say that for, you know, for orders less than a certain dollar value that have good forecast accuracy, whatever, get automatically approved so we don't have to look at those. So there are some options there.

Another area we're seeing some development is, again, another place that's kind of been a black art, which is New Item Introduction and NPD. So what are all the conditions around a new item launch, because again, we typically don't have great history.

The historical approaches have been copy history from a like set of items, bring it over, learn from that. But again, now the machine learning will start to take all of that into account and bring that in. Now, this is the bit that I think is kind of fun. The research is, and I've got several university white papers on this that were done by mathematics departments, is the bottom line is if you run machine learning by itself, it gets to historical data set versus kind of the traditional forecasting techniques. The traditional forecasting techniques do better. 

So, the thing about math, remember it's always horses for courses, and the best analogy that I've come up with so far, I'm sure there's a better one, is that if you get a block of land and it's covered in trees, but you want to build a garden, do you get the gardener or a landscaper to get rid of the trees?

So machine learning is a subtle process of refinement. It's actually very sensitive. It's very subtle. So, I don't want to use machine learning and AI to do my base forecasting. In fact, I've talked to several companies that have tried this and changed it because they said it made it worse.

I talked to one of the big grocery retailers and they said, "We're still using our existing forecasting approach for our forecasting, but we're using AI and machine learning at a store level to look at foot traffic during the day, weather patterns, and a few other things to then figure out what we need to do in the store that day." So again very short-term horizon.

So it's not a replacement for, it's an in addition to. I just want to make clear, so you don't get to skip steps. We don't really do good forecast testing, so we'll just do all machine learning and we're right. This is the first point.

The second point is like AWS, Amazon Web Services has an AI and machine learning engine. So they say, "Throw us your data, "we'll give you a forecast back." Have a think about that? What does Amazon do? They sell products. What are they asking for? Data on your sales history. I don't know .

Have a think. But what I'd be saying is, also be aware that you need to be very careful about what you're using it for. I would not recommend using it for your base forecasting. There's a lot more on the execution space.

So a lot more, because again, we're getting to this lower level. What I might do is just show you a couple videos, but we've got things like IoT sensors that are giving us real time information to help us with predictive maintenance, as you know as a good example, autonomous vehicles. And this is the area we're starting to see the AI and automation really come, you know, marry up.

So this is one on you know, shipping using AI, which is saying that if I start to understand that you know, chance of freight being delayed in Shanghai, what might I do as a rerouting option? There's a tropical thunderstorm that's going to be happening off the coast of Thailand, maybe we should switch to air freight for this particular leg. So again it's starting to look at weather patterns and saying, what will that actually mean for delivery and what are our options? I'll show you a little video real quick.

Video: It's no longer a choice, supply chains must digitally transform to gain a competitive advantage. The key to that advantage is visibility. Despite decades of trying, nobody has gotten the visibility they need. Why is that? Because until ClearMetal, no one has been able to make sense of the underlying data. ClearMetal has solved this problem with modern egestion, machine learning, and AI. We're the only provider that delivers the visibility you need and can trust. Forget all the marketing hype, buzz words, and misleading promises. Shippers around the world are using ClearMetal's visibility to increase revenue and drive down costs. We help optimise inventory, proactively manage exceptions, reduce transportation expenditures, and establish a competitive advantage. The future of supply chain is data intelligence. The future of visibility is ClearMetal.

Carter McNabb: That's not an endorsement of that particular offering. I'm just giving you an example of what some of the shippers are now putting forward as the point of difference. I haven't used that, so I have no financial links to that organisation whatsoever. Where we're also seeing it is in some of the smart storage systems. So some of you may have seen AutoStore.

Video: So why waste time on shelf-based solutions when AutoStore is here? AutoStore is a cube-based system, making use of all space for proper warehousing. Turn that wasteful air into storage, and double, triple, or even quadruple the inventory capacity without moving to a new building. Bins are stacked right next to each other. On top of each other. Radio controlled robots drive on tracks above the cube, lifts down to grab bins and deliver them to work stations for order fulfilment or replenishment. All operations get done efficiently and accurately in high-speed work stations.

Carter McNabb: It's a cube-based storage mechanism that works basically in three dimensions. It works horizontal and vertically, but there's an AI algorithms that are actually driving the automatic, because it's automatically relaying itself, you know, as it's receding and dispatching. So there's AI driving that.

You've probably seen the Boston Dynamics robot. Now this is still a prototype, but this is a picking robot. Kind of cuddly aren't they. Basically what I'll say is this, is that think a way into this there's a lot of question about how do we put our toe in the water and that is, I think that's the answer, is it is a toe in the water approach. Is really start with a hypothesis.

You probably already have an intuitive sense for there's something in the business that if I had, if I could relate these and make meaning out of it, and could use it to predict that it might add some value.

So let's create the hypothesis, develop of proof of concept. And in the proof of concept then you can start to see what things actually do make sense, what is relatable, what does add some value, you can then make the business case for it, refine the model, and then look to implement and integrate.

So you do that in a controlled way. And make sure it's safe and you're looking at the right things with the right amount of data. And again, the final point I'd make on this is that, you know again, technology is neither good nor bad. It just is. It's the intentions of how we try to use it, or how we intend to use it. I wish you all wonderful and humane intentions in how you use technology, and good luck. And thank you for your time.

Reproduction of GRA whitepapers and articles

GRA permit the reproduction of GRA authored whitepapers and articles so long as all the following conditions are understood and met:

  • Entire credit details must be included:
    • Author's name(s)
    • GRA name and contact details
    • GRA URL link to the original article
  • All hyperlinks within the article must also be retained
  • Articles must not be resold
  • GRA retain full copyright.

If you have any queries about reproducing a GRA article or whitepaper, please contact GRA Marketing

Testimonials

"GRA worked closely with us to develop a comprehensive supply chain strategy to strengthen our business in a rapidly changing retail landscape. The level of interaction and engagement with our team was exceptional, ensuring a high degree of buy-in for the strategy."

– Brett Kelly, National Supply Chain Manager, Officeworks

Typical results

  • 20-40% inventory investment reduction
  • increased service levels ranging up to 99.9%
  • 10%-15% reduction in supply chain operating costs
  • 5%-20% spend management savings
  • the ability to fund business initiatives from operating cash flow (OCF) improvements
  • improved return on capital employed (ROCE)
  • a minimum 3:1 ROI (10:1 to 30:1 typical)