Joe Bellini, Chief Operating Officer of One Network Enterprises, discusses machine learning and prescriptive analytics.
Q: Joe, in our previous conversation on control towers, you were starting to touch on machine learning capabilities. You said that machine learning might only improve a forecast 5% to 10% yet still have a major impact in the supply chain. One Network uses prescriptive analytics as opposed to merely predictive analytics. Is that improvement related to machine learning and prescriptive analytics?
Joe: What I was referring to was just on the demand front end — 5% to10%, improvements in forecast accuracy. But the machine learning (ML) actually goes much deeper if you look at the prescriptions. There’s a lot of variables across a network related to demand and supply and production and capacity. There’s lots of different ways to solve a problem. Let’s say there’s a shortage, you could: Add a second shift at the plant. You could expedite a shipment with a carrier. You could change the carrier mode. You could reallocate from other customers. There are so many different ways to solve these problems, but only if you have access to the data.
Now, because of our Master Data Management (MDM) infrastructure, all nodes in the network, if they have the right permission, can get access to all data and work together, collaborate, to solve a problem. So, now you’ve got the workbench as part of your control tower, and might see three or four different prescriptions proposed, the things I just mentioned, to solve a problem.
And you can program the machine learning AI capability in NEO intelligent agents to be autonomous. If there’s a certain way when those conditions occur that you say, “Well, the best way is always to do X in that situation.” In those cases, you can make it autonomous and run it all day long as conditions change.
Or, if it’s a bigger issue where you’d like to make a decision, you can push the recommendations from NEO to the workbench, and let the planner, or scheduler, or a group of trading partners who all have access to that screen, collaborate and decide what the best course of action is.
Remember that the machine learning is going to get better and better over time, due to the way we’ve designed it. It will learn what decisions for what types of disruptions lead to the best outcomes. And then the next time a similar condition occurs, it will make a better prescription based on those learnings."NEO intelligent agents can be programmed operate autonomously. You set the parameters within which they operate, monitor their performance and fine tune and expand their range as your comfort level increases." Click To Tweet
And the other thing about the machine learning the way we’ve designed it, and because we have the multiparty MDM, is that it’s looking at the entire landscape. I used to do a lot of linear programming, and you basically have limits. It’s almost like a box. You’ve got the four corners of the box, but in those limits, you can try to fine tune the linear algorithms. You don’t have that limitation anymore with machine learning. It’s what they call “vectors.” And you can have these vectors learning, whether it’s a weather pattern coming in, or a political disruption within a country where you’ve got key sourcing, or whatever.
It can start to learn about, “Well, how important are those trends right now to a potential outcome or a disruption within the supply network, and what should I do now to adjust for it?” So, it’s looking at decision “landscapes,” and then understanding what the high point is on it, what that optimal decision is.
Let’s say, it’s like where I live in Denver, the Rocky Mountain range, there’s a lot of trails out there. Which one is the best one to climb that day to resolve the problem? And it will learn about those landscapes and the right paths to take and which one to pick in any given situation.
Q: Talking about which “path to take.” That’s similar to route optimization, where it’s rules-based and expert systems. Am I right?
Joe: I used to do a lot of expert system work, too, and it’s a little different. A good analogy is putting on a golf range. So, you’ve got the little book, and it tells you the slope, and you can look at the grain, and you can determine humidity and temperature. And then a lot of people think mountains nearby influence how the ball turns and there’s a giant lake somewhere close by. And, so, you can try to calculate all that stuff and make the putt, but the probability of it going in the hole if you’re far enough away is very low.
Now, if someone’s ball is right behind yours and you get to watch them make the putt that you’re going to need to make. That’s data, right? That’s a direct data feedback as to what’s going on. Well that’s what machine learning does.
Machine learning actually looks at the data and then relates it back to the variables to say, “Well, what were the conditions that day? What about the mountain? What about the lake? What about the grain?” And based on the data it can understand how they all relate. It’s almost the opposite of the way we used to do it.
If you analyze a million putts on a green from every spot on the green, then you can tell somebody, depending on where they land, what the right way is to putt it, you’re going to get the ball in the hole a lot more than someone else just trying to calculate it using slopes and the grain. That’s how we used to do it. So, that’s why machine learning is different. The more data you give it, the more insight it’s going to have about how the variables on that particular day affect the outcome of the data. And it will learn and learn and learn, and get better and better over time.
So, that 5% or 10% improvement in the demand forecast that we’re seeing initially, for example, that could go as high as 25% or 30%. That will evolve over time.
Q: I see. Now, which specific machine learning technique does NEO leverage?
Joe: Yes, we’d have to talk to the development group about the specifics. There are a lot of methods out there and they’re all good for different things as to how the algorithm is going to learn.
Typically, what you’ll do is, you’ll run data sets through different ML techniques to see which ones better predict what the actual outcome was, and they learn that over time. So, there’s still a little bit of tuning that goes in. It’s not a cut and dried thing where the machine learning can automatically just do all this stuff on its own. It needs interaction with data scientists. That’s what that team is doing. So, if you want an update on exactly what we’re doing there, we can schedule a chat with the development team who are deep into that.
Q: Going back to the notion of prescriptive analytics versus predictive, what is it about machine learning and those various vectors that are responsible for increasing accuracy of the predictions with machine learning? What is it about it that enables it to prescribe your best bet or next best course of action?
Joe: Because we’ve got the federated MDM in the network and we’ve got this permissibility layer, we know all variables in the network related to everything. All planning and execution variables are known. They’re not trapped in silos and hard to access, and they’re not stale. They’re real time and they’re known. So, the machine learning has a huge advantage because it can look across an expanded variable cell which has never been available before. And even the simplest thing, like trying to combine variables like on-time-in-full. Yet, everybody wants that because that’s a key shipment indicator, “Am I meeting customer service levels at the least landed cost?”"The secret sauce behind the success of machine learning on One Network, is that it has all the data and federated master data from the network, all planning and execution variables are real-time and known." -Joe Bellini Click To Tweet
But to do that, you need to combine data from order management, from the carriers, from receiving, you know, the bar code that you’re going to input into inventory so you can authorize payment. There’s a lot of different interactions there where in the old-style ERP you’re dealing with three or four different modules that don’t talk to each other. So, it’s impossible to actually submit that data to a machine learning algorithm in an efficient and real time way, so it can predict when there are disruptions, try to handle them, and make decisions to improve things.
Whereas, in our network, all that’s available. Now you’ve got an expanded data set for the machine learning, the vectors. It’s similar to what Netflix did with AI in recommending shows. As they got better at it, they started doing a better job of writing the one or two paragraphs describing a movie so they could better predict who would like a particular movie. They added more data about the vectors, and as it got richer, they were able to make much better recommendations to people who would really like that movie. It’s the same thing across the network.
There is different data related to what’s really happening around transportation and logistics, and what’s really happening around forecast accuracy in the demand profile. Similarly on the supply side, there’s data around what’s really happening on supply availability in terms of what are you replenishing that’s really being consumed versus what’s going into safety stock. This gives you different degrees of freedom in how to make decisions, solve problems or take advantage of opportunities. Let’s say, a promotion is double what you thought the volume was going to be. People love it. Well, what do you do? How are you going to react to that through the network?
That’s really where the machine learning comes in is, because it is all about data, and the more data you give it, the better it gets. A network like ours that has access to all that data and can feed those vectors into machine learning, the machine learning is just going to be all that better. Soon, you’re making better predictions, and based on those predictions, I can offer prescriptions up as to how to take advantage of opportunities or solve problems.
Also, the system understands what the outcomes have been. So, when they’re looking at KPI’s, (key performance indicators), it can relate those to decisions that led to those outcomes. So, when they’re looking at the prescription, it will include an indication of what we believe the outcome will be if you choose a certain prescription to solve a problem.
Q: Right. Now what sort of innate advantages are there to deploying machine learning through a dynamic, a virtual or software agent such as NEO?
Joe: So, the agent can not only recommend prescriptions based on the data that’s available in the network, it can also execute them. The agent can actually execute the decisions that are made, make it actionable. For example, it can change a plan, change a schedule, change a load on a carrier, reallocate inventory even if it’s on a ship in a container. We know which items are on there, and by the time it lands at the port, if we’ve made a different allocation decision, the carriers know about it and can go ahead and distribute on that basis. That’s really the value of having a dynamic agent infrastructure, it’s being able to not only analyze and predict but actually prescribe and then execute, and you can make it autonomous."Yes, there is some value in robotic process automation automating administrative tasks. But the real key is if you can generate better outcomes in your KPIs autonomously. That's where we're heading." Click To Tweet
So, all this RPA, robotic process automation, if you look at some of the firms that are doing that, a lot of it is just trivial. They’re automating administrative tasks that tend to be somewhat repetitive. Sure you’re going to get a little bit of labor savings. But the real key is if you can make it autonomous in terms of generating better outcomes in terms of your KPIs. So in some quarters, you may want to go for revenue, others you may want to go for market share. Other quarters you may want to go for margin. There’s different pressures on the business to generate different outcomes. Those are all different strategies that are going to drive different decisions. And the NEO autonomous agent knows that, what you’re trying to do at any particular period, it can optimize it for you.
Q: Would you go so far as to say that’s a trend impacting machine learning, to actually deploy it in conjunction with an agent that can act?
Joe: Oh, yes, absolutely. Otherwise, you’re just analyzing stuff, analysis paralysis. What good does making better predictions do you if you don’t have access to the network, to go ahead and execute that better outcome; and then improve on it over time?
Q: Exactly. Analysis paralysis…
Joe: I’m a mathematician, so I heard that through my career. I don’t know if you remember, there used to be this commercial of these consultants that were coming out of a meeting with a bunch of executives and presented them with a big analysis. They’re going down the escalator laughing to each other and saying, “Can you believe those guys really wanted us to implement this?” And the company guys were going up the escalator, looking at these consultants and saying, “Well, yes, of course you have to implement it. Otherwise, you can’t get any value out of it.”
Q: Of course, else what’s the point? Well, Joe, like I said, it’s always terrific to chat with you. I had a lot more questions, but we’ll have to save that for another day. Thanks Joe!
You might also like…
- Prescriptive Analytics: A Conversation with Joe Bellini
- Control Towers: A Conversation with Joe Bellini
- Supply Chain Resilience and the Power of Business Networks
- Supply Chain Resilience: Surviving Hurricanes, Typhoons, Flooding and Heat Waves
- 3 Strategies Every Executive Must Deploy
- Prescriptive Analytics: A Conversation with Joe Bellini - October 16, 2020
- Control Towers: A Conversation with Joe Bellini - October 9, 2020
- Supply Chain Resilience and the Power of Business Networks - September 30, 2020