 Welcome global supply chainers to our third live event of ST0x. Thank you very much for joining today. My name is Alex Rokov. I'm your course lead and I have invited today Chris Kappes who is the executive director of the CTL and also your teacher in the past weeks. You have spent a lot of time with him. Thank you very much for joining Chris. Glad to be here. Glad to be here. I bet your guys are glad to almost be over. Yeah, we have only the final exam and our live event today. I think students are looking forward to that. So what as usual we have set up our Slido account. You can log in at slide.do and put in the event code hashtag s0x-3 and you will have this live feed and the opportunity to interact with us in two ways. On the one hand you can put up questions to us and we will pick these questions as we see them fit to our content that we are going to do today. And on the other hand we are going to put out some polls and I'm running one poll right now about the contents that we have served you with in the last 11 weeks. Before you switch though, so for questions put in a question but also we're relying on you guys to upvote the ones that you want to listen to. So if you see a question in there that you like then upvote it. That way we'll know which ones are the most interest to you all. Absolutely, that's important to us because then we can see that easy. Prioritize. It's like a heuristic. So our agenda for today is the welcome, which we just did. And then Chris will do a recap of what we've learned so far and an outlook of how what we've learned in SC0x ties to all the other courses that are upcoming. SC1x, SC2x, SC2x and SC4x to make the connections clear why we have set up the program the way it is. Afterwards, we have a couple of problems for practice that further should connect what we've done in the course with with all the with different problems that you can come up with in transportation logistics. Let's say that we're going to go through them step by step discuss how we apply our methods and see if you have any other questions you just put in on Slido. So, Chris, I will turn it over to you to discuss with us some of the learnings that we had. Great. So I hope you guys enjoyed this course. It's the first course that many people take when they go into the MicroMasters or the SCx sequence, but it was actually the fourth class that we created. And the reason why we did it is we found that just like when you come to grad school, you usually want to get all your math up front because then you're going to have all the tools in your tool belt to be able to solve things as you go forward. And that's the approach we took here. What we also tried to do was weave in hints at least of how these different techniques would be used in both the rest of the courses as well as in your career and supply chain, because every one of them has a very practical use. I didn't do anything theory for theory's sake. Everything is there to give you the underpinnings. So there are four big things that I hope you learned, and I'm not going to go into equations. I just want to do big picture things. The first thing that we tried to do is make sure you had a level of mathematical dexterity, because it's a model-based course, and 0x, 1x, and 2x are all model-based. Everything you do in supply chain, whether you're explicitly dealing with a model or you're just dealing with the results, is the result of a mathematical model. Network design, procurement, all these things rely on an underlying mathematical formulation of costs, time, some kind of relationships. So I wanted to make sure that you practice in using functions and using approximations and just getting your hands wet and getting comfortable back with the algebra that you think you didn't need to use again when you learned it in seventh or eighth grade, because you need to have that dexterity. So that's the big, big takeaway. Hopefully you're getting more comfortable when you see a summation sign. You don't freak out because it's just notation. The second big thing that we did in the first part of the course was optimization, and I took you all the way from unconstrained, classic optimization using calculus just a little bit so you understand it's not magic, and then went into constrained. So we did minimal, maximum, first-order conditions, and then second-order conditions. How do you know you're at the top of a hill in the dark? And the whole idea is because any step you take takes you down. It's the same concept that you use mathematically in our formulations, and then we also went into constrained. This is where we say, okay, you've got an objective function, you want to minimize or maximize, but you're subject to certain constraints. And so how do you work with that? And you saw the amazing result with linear programming that when you have a feasible region that's convex, that the solutions are going to be at the corner points, which is an amazing result that hopefully you appreciated because we rely on that for a lot of problems. In fact, a lot of the stuff, if you go into future grad work, you'll find that a lot of the work is to try to convert a complex, say, mixed integer linear program or something complex into a linear program because we know how to solve a linear program very quickly. So as you go into practice or into a higher grad courses or into other studies, you'll see that the problems you're solving are bigger and bigger and bigger, and hopefully you realize for integer programming or mixed integer programming, the size will kill you. And so there's all these tricks and things you can learn at a higher level how to convert things into a linear program. But everything we did in optimization, we're trying to come up with an answer. It's all prescriptive. So we did unconstrained optimization, which you will use, and I'll talk about that later in SE1X, constrained optimization, which you'll use a lot, especially mixed integer linear programming, every supply chain uses this. They might not be explicit about it, but when you're deciding where things flow, where things get located, at the bottom of it, it's a mixed integer linear program. And so hopefully you've got a feel for what you can and cannot do with a milk. The other things that we added in, I say probably three runs ago, two runs ago, we added in a section on algorithms and approximations. If any of you are over the age of say 55, and I'm sure there's some of you, you might have remembered a slide rule. But the vast majority of students that certainly I see at MIT or they're on this class don't have a slide. Did you ever have a slide rule? No, not really. I had one in engineering school just to see how it works, but I never worked with it. The thing about a slide rule, it's awesome, but it also helps you think about approximations. And so you think of orders of magnitude very clearly. We've lost that skill. I had to use a slide rule my freshman year in high school, but that was it. The calculators came in. So we've lost the skill of approximation. That's why I put that in there and did some approximations and used the approximation technique to also illustrate algorithms. So some of the algorithms in network optimization that I talked about is shortest path, traveling salesman problem, vehicle routing problem. These are used every day. And so when you have the exact data, you can use exact algorithms or maybe heuristics, those approximations if it's really big. But sometimes you need to make a decision and you have very little information. And that's where these approximations come in. I gave you one example using where it's approximation of local routing distances, which is very powerful. And I'll explain why in a few minutes how it's used. But everything we did for that first part of the class was mainly deterministic, but it was also prescriptive. These are techniques that will give you an answer. How many facilities should I have? Where should I put them? Which things should I do to maximize my revenue or minimize my cost? It gives you an answer. In the second half of the course, we went into more descriptive and predictive things that don't tell you an answer, but they either describe a situation or they predict an outcome based on some certain inputs. So we went into probability and statistics. Probably the hardest week there was when we did hypothesis testing. There was a lot of stuff there, but I wanted to make sure you got it all, and so we applied it. So hopefully you took away how to handle distributions. You'll see the normal distribution throughout your career. Poisson distribution as well, if you get into spare parts and slow moving things. And I introduced triangle distribution. Again, a great approximation if you don't have any detailed data. Hopefully you're comfortable characterizing data. Every project in the 15 years I've been here at MIT, the size of the data sets that students have dealt with to solve research problems has grown by a factor of 100. It used to be everything could be done in a spreadsheet. And now this is why we teach you an SC4X, SQL and large, massive data skills, because you need to be able to deal with and how to characterize large sets of data. Data cleaning is a critical skill, but being able to characterize data through central tendencies, the dispersion, all of that is very important skill that hopefully you picked up. Then we went through all the hypothesis testing. What do you do with the data? The idea of p-values, point estimates, confidence intervals. Hopefully you got a handle on that and the way that you'll get better at that is just doing more problems. And so after a while you see it's very mechanical. That's why we covered a lot of it, but once you understand the concepts, the mechanics are very straightforward. And so when do you use a normal distribution, a Z-statistic versus a T-statistic? And hopefully after through repetition you'll get comfortable with that. Finally we focused in on the last part, regression and simulation. And these are more predictive techniques. They're going to predict what an outcome will be. They will not tell you what the best solution is. There's no such thing as a prescriptive regression model or a prescriptive simulation. They tell you, given these inputs, here's what you expect the output to be. It won't tell you the best. People really get confused over this, but they're two very separate techniques. And they work well together, but they're separate techniques. On regression I went through example after example of building a model, ordinary least squares. You'll use this if you do any forecasting. It's widely used in practice. There's a lot of other techniques that you can go into, but ordinary least squares where you have this dependent variable, it's a function of all these independent variables, thinking back to functions. It's used widely in forecasting. I use it in cost analysis. So you'll see that a lot. Simulation is used mainly a lot of times within a factory four walls. I've seen a lot of simulations or within a network designed to see how a certain solution or a certain policy will go. Simulations make a ton of sense. If the situation is so complicated, you can't create a nice closed form analytical solution or can't solve an optimization model. And it's also really good to do what ifs. What if I change my delivery policy so that it's every third day instead of every other day? And if you set up your simulation, you can see how that effect will be. And then at the end of it you say, is it any better than my previous solution? That's where you use a hypothesis test. So at the end of a simulation or regression, you're using all of those statistical tests to validate the model and to help give you some insights. We also did queuing theory as well in there, and that kind of fits right in with some of that as well. So I asked students which of the concepts they like most was most interesting for them. And it's supposed to be that your drive for constraint optimization has picked up. That was a picture that we saw also before. More than two-thirds of the answers say, well, we like constraint optimization. Actually, simulation seems to be also very well perceived. So personally, I'm showing my bias because I'm more, it seems like you're, it's like religion, right? You're one or the other. I tend to be an optimization guy. I cut my teeth on large-scale, combinatorial optimization. And so that's the biggest tool in my tool belt. I don't know, are you also more optimization based? I'm also more optimization based, although I'm also kind of exploring all that big data stuff where we get more into regression analysis, but it's mostly optimization. But to be fair, I've been using regression on the side for other projects for, gosh, 15 years. Simulation is never in my strong suit, but we have some people here who it's a primary tool. So as you get up to like a PhD level, you find that you generally go to, you find your methodology and you kind of use that a lot, or you find a domain. But we try to make sure you have at least a taste of everything. Yeah. That's good. Okay. So hypothesis testing is still not very much a thing. Hypothesis, yeah, you know what? You use it when you got to use it. Yeah. It's a way of, so if you're working in a warehouse, you're not going to whip out your hypothesis test that often. But if you're doing an analysis to try to justify spending $5 million on a software system that'll improve operations and increase time, yeah, then you might want to do it to say, hey, you know what? This is a solid improvement, or this is not. So you got to pick when to use your tools. And we will cover that in our professional practice a little bit of how we could use hypothesis testing. Yeah. Let me go. Oh, do you have another question? No. Okay. So, so you have these big things that you learned in all things in eight weeks of content. A lot of stuff was put in there. And I know that. I know I put a lot in there. I did that on purpose. Let me explain how it's going to be used in later courses if you continue on through SC1X to 4X and continue with the MicroMasters, which I hope you do. In SC1X, you're going to learn all about forecasting demand, managing inventory and planning transportation. It is the core course. It's the first course I created. It mirrors STM 260, which I've been teaching here for more than 15 years at MIT. And it's the classic trade-offs between demand, planning, inventory and transportation. So having too much and having too little, having a fixed cost and a variable cost. And as soon as you think of that word trade-off, you should be thinking optimization because you're finding the best under certain situations. So in forecasting demand, we'll do a lot of work with functions. You'll use regression. You'll use statistics to understand the accuracy of your forecasts. And it's all one big function, right? So your demand is going to be a function of maybe some different patterns, a level, a trend or some kind of seasonality. You'll use all of those regression and statistical tools to come up with a better forecast. For managing inventory, you'll use unconstrained optimization. The classic inventory model is called EOQ, Economic Order Quantity. You'll recognize it is that you can solve it with a first-order condition. And you'll end up with the EOQ formula. But hopefully by taking SC0X, you know where it came from. It's not a magic formula that someone handed down to you. But you'll see how it came about. You also have a lot of minimum costs under uncertainty. So we'll use a lot of statistics there and distributions. We'll use a lot of the normal distribution and the Poisson, because you want to understand if you have a certain amount of inventory and your demand distribution, which is a stochastic variable, what's the probability I'll stock out? So you use probabilities a lot. And so the idea is you're going to be figuring out how much inventory to stock, such that my probability of stocking out, which is the probability the demand is to the right of your level, is at some level, less than 10%, let's say. So we'll spend a lot of time understanding how you set your inventory replenishment levels based on probabilistic demand. And then in transportation, you'll really look at optimization of time versus cost. And you'll see how the uncertainty and variability in transit time affects the amount of inventory you need to hit a certain level of service. And we'll do a lot of mode selection, understanding how variability is. So we'll really focus in. But the big thing for SC1X is you're going to see at the end of the course, it really boils down to one big equation, the total cost equation. And you'll see that forecasting affects the equation in a certain way. Inventory is affected a certain way and transportation impacts it a certain way. So what you see is how to make the trade-offs of all these different levers that you have in your total cost. So SC1X will use all these models. Can I add a question that fits into all that inventory discussion? Yeah, sure. Because Prashant asks, are we going to cover different inventory models such as a smooth, lumpy, and erratic intermittent demand too? So these are more special circumstances that we have in inventory management. Do we cover that in a similar way? Well, that's what Prashant's describing are not really inventory models or domain patterns. So yes, we will. We'll cover Krosten's method, which really looks at how, if you have lumpy demand, and when do you have lumpy demand? Let's say if I have a parts replenishment. I give spare parts out to cars or manufacturers might order from me. I might get an order occasionally and the order might be of very different sizes. So if I look at just the average, it doesn't make sense because there's a lot of time periods where there's no demand. But when it comes in, it might be a big burst. So when you have demand that's very sparse or very consolidated into bursts, then there's different methods you can do. There's also, what you'll see is the inventory model, and maybe this is what Prashant was getting at, is you use the inventory model that matches your demand distribution. So if demand is deterministic, there's no variability, then we have one set of models. If it's variable and it's variable, but it's deterministic, I know where it's going to change, I might use a different model. And so if it's very a short period of time, I have a different set of models. So what you'll see is different inventory models are used based on the assumptions of the demand. Hopefully that makes sense. It'll make more sense if you take SC1X. We spend probably three weeks covering these different models. And perhaps the follow up to that is, are we also using different distributions then? So Prashant also asks if we're going to use, I think he had Poisson and negative binomials. Yeah, so the lion's share, to be honest, most systems assume normal. Small, then we'll obviously cover normal, we'll cover Poisson. I won't go into log normal. I'll hint at it, and there's different ways you can do it. I won't go into any binomial, negative binomial. You can go truncated normals. The inventory modeling pool is very deep. And so there's a lot of depth you can go. We won't go into multi-eshelon inventory. We touch upon that in later courses. I think we touch it a little bit in SC2X. That's a more advanced topic. But the concepts are all the same. And hopefully the concepts of it, if you change the distribution, all that matters is how you figure out that probability of stocking up. That's the only thing that changes. Whether it's a normal, whether it's triangle, whether it's Poisson, all you're doing is it's a different method you plug in to figure out your probability of stocking out or your expected number of units short. That's all that it is, and then it just gets mathematical. But we'll cover most of that. All right, then in SC2X, if you like mixed integer linear programming, you will love SC2X. It's all about design. SC2X is supply chain design, and there are three elements of the flows. You have the physical flow, the information flow, and the financial flow. And so the first half of the course is faced on the physical flow. And that's network design, network flow analysis, channel design. How does product get from origin to final consumption? And so that includes where should I set up facilities? How many should it be? Which facilities should serve which endpoints? How do you flow things through? And it's all a huge mixed integer linear program. And you'll get more skilled on the different techniques because we go a little deeper than we did in SC0X. In fact, we actually explained some things in Excel doing that because SC2X was the second course I created. And so you'll probably be able to do it even easier if you've gotten comfortable with SAS or Ample, to be honest. But I show things in Excel because I blow up the constraint matrix because to me that's how I learned. I learned doing it by hand. And so hopefully when we go into the details, looking at it in the constraint matrix, that'll make more sense to you. But we go into a lot of depth there in mixed integer programming for aggregate scheduling, production planning, and a lot of different things. We also will go into finance and other areas, but they're not as model-based. And SC2X is kind of like the Keystone course. Everything prior to that, 0 and 1, is all very mathematical modeling halfway through. And then we start getting more qualitative. So we go into the financial aspects. We go into procurement and some other different things. In SC3X, we get in, and the big focus there is we now introduce exogenous factors. Say, OK, here's what the model says, but you've got these state regulations or you've got a border or you've got uncertainty that's just coming in a disruption. So how does the real world influence your models? So we'll introduce system dynamics, which is a mathematical simulation technique that I thought about putting into zero, but I think it belonged closer to the application. And so system dynamics is an interesting technique. It helps you understand the dynamics of a system, how different entities interact with each other, because that's what a supply chain is. We'll get into risk contracts. So if I have a supplier and say a manufacturer and a retailer, and how do they share the risk between how much they order? And for that, we'll leverage some of the stuff we did in SC1X called the news vendor problem. And that will leverage things that you just learned in SC0X, which is probability distributions and unconstrained optimization. SC4X is now we're going to large scale things. And so the big learning from SC4X is how to do this analysis in practice. So instead of doing things just in a spreadsheet, you've got to use SQL. So you have a relational database. So we spend a lot of time understanding why you use relational databases, how to use relational database, because if you get in real life and in real research projects, you'll have millions of records that even though Excel can handle a lot, it doesn't make sense. There's other tools. And so MySQL, we use that to help you understand how you can handle large scale data. And then also machine learning, which is kind of again something we thought about teaching in 0x, but I think it really hits home when you see the problem of having massive data sets. So kind of in the very last course, we introduced some new math again. But prior to that, you're really using only the stuff that you've learned so far. So SC0X is really setting you up and providing you a foundation for all the SCX courses. Think of it as a toolbox. So now you've got all these tools and you'll use all of them throughout your course, not all at the same time, but it's very, I think it's very useful for you to know what's in your toolbox. So when you see a problem, you say, you know, that looks like an optimization problem or that might be best solved by simulation because what you'll see in practice and in research, the same problem can be approached by many different techniques. And so you've got to pick the one that you're comfortable with and you think gives you the best answer. That make sense? Absolutely. Thank you, Chris. We have a couple of questions from our learners that are getting more thumbs up and they all circle around how do we use these methods in practice? So Mo, for example, asks, many concepts have been covered during SC0X. I can't imagine how to use these models and methods in real-life supply chains. The other Sahil is asking that this was a great course with lots of problems, lots of practical problems, whether the other courses have similar practical components are they practically related? And the third one that I just saw is we've gained a good knowledge on SC0X. My question here is, how to get the most benefit possible of implementing some of these concepts? So for me, there are two components. One is how does that relate to further courses because we are setting the stage for further courses? I think you've covered that already. The other question is, how does that help me in practice? Yeah. And so I'm going to do some examples here, but it depends where you are. I mean, to be honest, if you're in a planning group, which a lot of our graduates go into planning groups, whether you're doing demand planning, inventory planning, procurement, where you're doing more strategic kind of analysis, you will use a lot of these techniques. And if you're in any demand planning, you will use regression. You will use some kind of stochastic modeling without a doubt. If you are in procurement, you're using optimization. It might be hidden because you might be using a system like Ariba or other systems that are out there, but underlying, you'll know that there's an optimization model. And that's the big thing. What a lot of students complain about sometimes is, like, why do I need to learn this math if I've got software that can do it for me? There's just so many cases where people think software is magic. And it has limitations of what you can and can't do. And so I think it's exceptionally important, and that's my hallmark here, is to make sure you understand the fundamentals. We're not going to teach you necessarily how to use each software tool. We do some in recitations, but all of them sit on the same fundamentals. And as long as you know those fundamentals, in my opinion, it's easier to use one or the other when you understand the strengths, the weaknesses, the limitations of each of these methodologies. So that's why we always go to fundamentals. If you want to learn just how to use a certain software, go to YouTube. And just you can, my gosh, I use that if I want to brush up on something, find a video of someone entering something into SAS or into SAP or into Cplex. And you'll learn that, but it's dangerous because you need to understand what things you're standing on, what assumptions you're making when you use that software. Okay. So to add some of the practicability of our methods, we have prepared a couple of problems from practice. And I suggest we go just step by step through them. So I picked these. Right. We don't, okay. So what we want to do is we want to present the problem to you and then ask you with a quick poll of how you want to solve them. So yeah, dip into the tool belt and think about, well, how would I approach that? Which method do you think is most appropriate to solve the problem? And I will open that first poll for the first problem and you can, while Chris is presenting it, you can think about, well, how I would tackle that. Perfect. All right. So I picked these problems because these are actually mainly based on work that I've done with companies here while at MIT. Some resulted in a capstone or thesis projects, but every one of them has a company sponsor. So the first one, you want me to just do it? There you go. You got it? Okay. Is with a company called CH Robinson. And it's level of service, LOS versus price. And so you can read it, but let me just paraphrase while you're reading along. The situation is that CH Robinson is actually the country's, the world's largest freight brokerage company. A brokerage company is simply a middleman or a third party between a shipper, someone who has goods to move, and a carrier, someone with a truck, right? And so they broker the relationship. Sometimes they're called a 3PL or a third party logistics provider. I'll just use brokerage. And one of the lines of business for CHR is that they will completely outsource the management of the transportation department. And so in this role, they will select which truckload carriers to use. And they will actually operate that on a daily basis. When they make the selection of which carrier to use, they typically make it some kind of trade-off. Whenever you hear trade-off, you think optimization, right? Between price that they have to pay the carrier and the level of service they're going to provide. And so level of service is one of those squishy things that means different things to different people. But in this case, it meant a combination of on-time delivery, on-time pickup, the transit time, the acceptance ratio, the damage ratio, all these kind of things. Acceptance ratio in truckload trucking is this. If I offer a load to a carrier and they reject it, then that dings against them. And so I look at that percentage that they accept. Contracts in the U.S. are a little weird. So acceptance ratios can be like 90%. But if it starts dipping low, then that's a problem because then the shipper or the broker in this case has to go and find another carrier. So you try to keep the acceptance ratio very high. You want a carrier to accept all the loads offered to them. And remember, I don't think I talk about truckload carriers. I might a little bit when I do the transportation approximations. Truckload carriers, as opposed to less than truckload or parcel, go from point to point. So a truckload carrier will pick up a load from this manufacturing plant and go deliver it straight to a customer's DC or distribution center. So they go point to point like a taxi cab, as opposed to something that picks up a lot of smaller loads, consolidates, and then delivers that way. That's more like a bus. So we're focused on truckload trucking. So CHR does this for a bunch of companies and they came with a question and they said, basically, does level of service have any relation to the cost? So the amount that I pay, because it seems an intuitive thing. We pay more and so that's the question and say, should we pay more for better service? Do we get better service if we pay more? What drives that level of service? Is there something that we can do to improve the level of service that carriers provide us or reduces the cost? So that's the question. So how are we doing? We have about 20 answers. People are still digesting. Oh, come on. You guys got to make decisions. But I think some of them have an initial intuition that maybe it is a constrained optimization problem. So you set up an LP or a mixed energy LP or that it may be something for regression and think about how different factors that we were talking about influence the price. So unfortunately you guys missed the one. Only 11% is saying the biggest thing, large-scale data analysis. Because we were doing a project. Now this is ongoing. We've had a series of projects with CHR. We're doing one right now looking specifically at the impact of lead time and level of service. And it's massive amounts of records. And the interesting thing, this will be in practice, you'll get data from multiple sources merging that data, cleaning the data, explaining what the heck they meant with different fields. That's a big time job. I have two students working on this now for a couple of weeks trying to understand and get the clean data set. So before I can do any of the analysis that we've done in this course, you've got to make sure the data set is clean, pristine, and you understand everything that's there. The way we're handling this is mainly going to be regression. I threw a red herring if you know what that means when I talk about optimization. But really what you're doing is you're saying, okay, my dependent variable here was cost per load. And my independent variables were the region, the destination region, the distance, and then possibly level of service. We also did a series of models using regression, but it's a different type of regression that looked at the probability that a carrier accepted a load or not, 0, 1. It's called the binary loget model. It's a little more advanced, but it's a regression. It's the same thing. I have a dependent variable. It just happens to be a 0 or a 1. We talked about that, right? So I can't use just ordinary least squares, a slightly different technique, and what factors influenced that a carrier would accept it or not. So we tried to quantify that, and then we used hypothesis testing to see if I have a low cost carrier or is a high cost carrier is there a substantial difference in their level of service. The punchline is, and this was a surprise to everyone, but we've repeated this result in a lot of different examples, is that level of service and price are not correlated. So paying more doesn't necessarily mean you're going to get better service in truckload trucking. What we found, the interesting insight, and this is why research is so much fun, we found that some of the best performing carriers were the cheapest ones. And to understand why that is, you have to take SC1X, but I'll give you the hint. They tended to be smaller regional faced carriers that operated within a certain region instead of what's known as nationals that are operating everywhere. And so what we found is some of these smaller regional carriers offered better service because they were CH Robinson or the shipper was a bigger piece of their business and they tended to operate on a known set smaller area of the network and they were working with a very large customer. The nationals, they have to operate everywhere, so they tend to be more expensive because they have to cover a lot of holes where there isn't a lot of traffic and their level of service is not as great. So it was kind of an interesting analysis and we've done this in a bunch of different ways to understand different factors that influence the cost. You also have to build more of a geographic or a spatial model to understand that if I leave from origin A to B, that's not going to be the same cost as if I leave from B and go to A. And that's something we'll talk about in SC2X when we talk about network design because what we found is, and this is pretty well known in the industry, it's not symmetric. The cost of going from, I always use this example in the United States, Chicago to Miami is not the same as going from Miami to Chicago and it will talk much more about that in SC1X. Very interesting result there. So you've related that problem to SC1X and SC2X for transportation network design. Is there also a course that talks about how to manage big data? Do we touch that? We do a little bit SC4X. SC4X we cover it, but we do enough to make sure you understand how the big systems and supply chains work. There are some great data analysis courses on edX. A new MicroMasters is coming out from our IDSS that is going to be specifically on a large scale data analysis. So there are other very specialized courses where you can go much deeper into the math for these, but we cover some of this in SC4X. I think that was one question that came up. That's a hip topic and everybody thinks about it. Students probably want to know more, so we're doing a little bit about that. So here's the thing, all the techniques you've learned when we do data characterization, that's big data analytics. You just have to be able to do it outside of Excel. So the concepts are the same. Now machine learning is a different animal, and so that's different because that's where when we do regression, we hopefully made the point that correlation does not mean causation, but that you want a story. I kind of stress this. So we did something where we did, I think I did it on pricing. Cost is a function of distance, all these other things, and so they had to make sense. I couldn't make it my independent variable be the color of the trailer or the number, the last digit on the license plate. Because let's say there's a correlation there. What am I going to do with that? Machine learning is just the opposite. They don't care. You just find correlations. You're hunting for correlations and there's a lot of great techniques and things, whether you're supervised or unsupervised. So SC4X is a good course that will cover some of this and it's a nice launch pad for you if you want to go deep. Because again, just like inventory theory, machine learning is a deep pool and the math gets really tricky, really deep, really fast. Lots of different ways to do it too. So many different methods to apply machine learning is such a big pool. So supervised machine learning, regression is supervised machine learning. A lot of it's just the names, the labels put out, but it is a great field and it's really expanding what we can do. So we'll move to the second problem. I will open up a new poll. You're presenting with new house students thinking about that. So this is a problem we did with Starbucks about two years back and this is based on a thesis by and Cecil, efficient supply chain design for highly perishable food. So this is based with Starbucks and hopefully everyone's heard of Starbucks. I think they're everywhere. They're in Italy now. Quick service restaurant. They mainly service coffee. I mean, they're a coffee shop, but they do other things, but they replenish each store daily from a regional DC and RDC using a third party distribution company, usually between like 10 o'clock at night and 4 in the morning when it's closed. But in addition to selling coffee, they started selling fresh food. Not just stuff that's frozen, but actually freshly produced each day because they're finding they want to get more into the afternoon. They want to have lunch meals and not just be a morning breakfast place. And so the fresh foods are made by a third party. They don't do it themselves. They have different companies located around their network and they make these sandwiches. These stores will send an order in the afternoon. It gets assembled by, I think, like 8 o'clock at night. By midnight it's ready to be distributed and it has to get to the stores on a daily basis. And so right now the fresh food component is very small compared to the other products, right? But the question is, the Starbucks had is how do I flow this fresh food products to its stores? Should I have the supplier that's making these sandwiches and assembling them? Should they send it straight from its plant to the Starbucks DC? Then it gets combined with the other things getting distributed. Should the supplier send it directly to each of the stores? Should they consider, when do they consider one distribution channel versus another? So this is the question the Starbucks had. How do I get my fresh food to the stores daily as efficiently as possible? So that's the question. That's the question and people are starting to think about it. I think we give them perhaps one or two seconds 16 answers and counting hopefully counting. And so one of the skills hopefully you'll learn by doing this is, yeah you have the tools and I am glad you guys ask the questions, when do you use which tool? Because that's part of the thing is first identifying what the problem is. The second is adding, how should I approach the problem? The third is, how do I actually solve it? What data do I need? What techniques do I need? So it's kind of steps identifying what is the problem, what class of problem is it, and then what do I need to actually solve it? Because a lot of times you'll identify something and say, oh, that's an optimization problem, but the data doesn't support it. You don't have the data to be able to do that, so you've got to find a different approach. And that's why I covered some of those approximation techniques early on in the first half of the course. So about 33 answers right now, constraint optimization is in the lead so they want to set up a network design problem I would say and then there is a second group that is perhaps more concerned about the size of the problem. So they are thinking about algorithms and approximations here and simulations perhaps. Yeah, so what we developed was a total cost model of how much time it takes for practice flow and if it went on one vehicle versus another, and what they found is it's really a function of how much is flowing through the system. So currently it makes sense if they were going to deliver straight from that food production plant to each of the stores themselves, they'd have to do a multi-stop, they'd have to go to multiple stores, like 15 different stores to have enough to fill a truck to justify that route. So initially it probably made sense to go deliver all the fresh food to the Starbucks DC which adds time, right? And then it gets added on to the other standard deliveries. What that does is that almost adds a day right? And so what we found out what we could do, you could try to see can we shift the time when they start? They start at 8 and finish by midnight, can they start at say 4 and finish by 8 and then we can combine them with those that same day. So the challenge they found once they understood the tradeoffs was that if I piggyback off the existing deliveries coming in for all the other product, I lose a day. If I go directly, I gain a day but it's more expensive. And so they did do a constrained optimization and they actually used some approximations. And so because what are we doing here? It's the same approximations that you did. So they're going to a number of stores in a set area of a density, of a set density in Boston is where we did the analysis. So you don't know exactly which routing they're going to use, so you don't use an exact algorithm use an approximation. So if I know I have to deliver to 20 stores within this 10 mile area, I know the density and I can figure out the expected local distance, I can figure that cost. So what we did is we used an approximation of the local delivery cost to make that tradeoff of do I go through existing channel or do I do a direct fresh channel. And so they found that the current situation for the sample, the pilot that we looked at, it made sense to piggyback. However as the volume increased there's a tipping point when it makes sense to go direct to stores with the fresh food. And so we're able to identify how to find that tipping point and when that would happen because that helps planners. Right? I can start saying, okay my forecasters say, you know what, forecast again a technique for regression I'm seeing demand increase like this so therefore I can play in 18 months out. I'm probably going to need to have a different distribution model maybe go direct. So what this pilot, this thesis was able to do was help them one quantify very quickly how to what parameters are driving the decision of whether to go direct or to piggyback off of existing and then two what signs to look forward to know when you want to make a switch. I'm trying to think what else we did there. I think that was mainly it was a fun project. It was really nice and we developed it was all in Excel. This was totally in Excel and we handed the project the spreadsheet off to them where they can make this decision very quickly just changing some parameters. The other thing we can do is we can play with this and say okay what if situations change? What if the density changes? What if the value of the products change because we also looked at the holding cost of having too much inventory because one thing you could do is deliver every third day but then you've got inventory hanging around as opposed to delivering every day but the tradeoff is if I deliver more frequently I have less perishable items getting spoiled I don't have to care as much inventory but I'm paying more that frequent delivery as the opposite I deliver less frequent so my transportation costs go down but my inventory holding cost goes up and my waste cost might go up because I might have some things that you don't want to buy the sandwich that's four days old so you have waste. So there's a tradeoff there and we're able to capture those costs and then use an optimization model to figure out the best path. So a lot of techniques that we taught in SC1 X used and we will be doing that in SC1 X and SC2 X more detailed and dig more into network design problems and how we can actually tackle these problems, right? Right. Want to go to the third problem? Yeah, let's go to the third. And while you're talking I will again open up the polls to see where students are going. So this seems to be the same problem but it's a little different. This is Walmart. Again, the world's largest retailer. We've done projects with them for years. They're a great company. They've done, they've innovated in the transportation logistics base for decades. So they're a large discount retailer and originally they were dry goods. They got into grocery in the late 90s and they are now the certainly the United States largest grocer. They might be the worlds but they're certainly the United States by a large margin and so as the difference between just being a discounter where you can put things in drive-in trucks you don't have to worry about a shirt spoiling. For grocery, you have to deliver different product temperatures so they're generally three. Ambient, which is just room temperature, drive-in but then it's refrigerated and then there's frozen and there's even something in between so you don't want to have ice cream delivered in a refrigerated truck and you don't want to have say grapes or other fruit delivered in a frozen truck. It'll damage the product. So you have these three different types and traditionally they would replenish a full truckload in a specific temperature zone to a store so they're ordering their drive-in stuff, their regular dry goods stuff. They have an ambient truck and they have all stuff that can be kept at room temperature. They deliver that. They have a full refrigerated truck and so all the milk, all the dairy, all that stuff gets delivered in one and so it might be that that truck would have to go to two stores to unload. Well what's happened is they've started a new format of smaller stores in urban areas and it was turning out that they'd have to do many multiple stops. It wasn't just one or two to take care of a full truck. It'd be three, four, five and as you start doing that you start running out of time in the day and so it starts becoming uneconomical and so what they've started looking at is multi-temperature trucks. So think of a truck with zones and they can actually move the sliders. One can be frozen, then refrigerated, then ambient. And so the question is should I use these trucks? They cost a little more. And then the question is if I have these how should I use them? Should I make it so that I use the classic trade office? Should I have single zone trucks delivered to multiple stores or should I try to have a multi-zoned truck delivered to single stores or a multi-zone truck delivering to multiple stores? And so trying to understand which stores get serviced by which type of vehicle and then how much space do you allocate for that? That's the question that Walmart wanted answered. That's interesting. I have never thought about it. I'm not a transportation person. I have never thought about trucks with multiple compartments. Are they more expensive? They're a little bit more and they take up some space because you're taking up space by having these dividers and anything and you learn SC1X risk pooling. If I take one big zone I can store more stuff in one big zone than in three zones that sum up to the same area because I'll have gaps in each. But it is a little more expensive because the other thing is you're paying for the generator unit even for that ambient. Because it takes a generator refrigerator unit. Our students have about 30 replies here thinking that we can constrain optimization, algorithms and approximations. There's coming in some regression now and simulation again. We thought about simulation. That was our first go-to because we could trade different policies like every store delivered every day. What would that do? And then we'd have to figure out a rule or an algorithm for when to use multi-stop. But we could test different policies. We chose not to because we went with the constrained optimization approach for this actually. And so as a mixed integer linear program here where the objective function was the total cost. And the total cost you can think of it as for each pallet being delivered. So you look at every store had a fixed demand. So think of your constraints store A had certain demand and we did weekly base buckets, number of pallets frozen, refrigerated, number of pallets ambient for each store. You had to meet that. And then the question is my decision variable was the number of pallets assigned to each type of truck. And so we minimize the total cost which is the transport cost, the stoppage cost and the holding cost. And you'll learn in SC1x what holding cost is. But essentially the idea is if I buy a week's worth of stuff and I keep it and buy it all at once and it's there versus I buy a week's worth, a day's worth of stuff every day. You can see there are different inventory levels. So the holding cost is the cost of holding that inventory for a period of time. And what you'll learn is that it's not the total purchase cost. You have to look at what the hurdle rate is, the H or the R that we call it, which is what your cost of capital is. So it might be like 20%. But as you order less frequently your holding cost or your inventory holding cost goes up. As I order more frequently, my holding cost goes down, but my transportation cost goes up. So you have that trade-off of transportation versus inventory. And so what we did is we really made a trade-off between the different modes because some stores if they ordered they had a demand that was high enough then it justified a full truck load. And so this is the model that we used, the Mixed Intermediary Linear Program that determined how to replenish each of the stores. You could have easily done this also in the simulation it'd be a cool way to do that as well where you're testing different policies but it would not tell you which trucks to use. It will tell you, it'll say given this policy here's what you will get. That would actually have been my question how do you determine because there are a lot of students who don't know about simulation why not use simulation for this project or also for the second problem with Starbucks why don't you use the simulation why do you opt for an optimization? I think we didn't one thing is that the company wanted answers and so it was less of a policy situation and in fact for what you see for both of those situations in practice what a lot of companies will do is you use both of the tools so you use a simulation and you do an optimization to find an optimal solution then you simulate how things will look because in optimization as you saw the demand is deterministic everything is fixed and so what you can do is two things and we'll do this in SC2x I can run the optimization multiple times where I take a random sample of my input variables and so we do this for network design and you can see so you run the model multiple times based on essentially a Monte Carlo pull of the data the other is you have a solution then you change the demand and see how well does it perform and you can play with both of those so we thought about that and if I had a follow on maybe I would consider that but for the simulation you would have to have an optimization embedded within it because what the simulation is what are you simulating and you'd be simulating demand so if demand changed what you're doing is you're saying demand for store A is now 10 instead of 20 so you need to have an algorithm or heuristic that says ok based on these demands I would use this truck, this truck, this truck so the simulation is really just simulating the demand you'd still need a piece in there that is making a decision of what loads to assign to what trucks to what stores also sort of an axle this problem no it fed into a gosh what did he use, we had a PhD student help with this and he coded it, I want to say Cplex you use Cplex for the optimization in Excel we use that as the front and it called the other what we're finding now is more the in fact we're doing this in here in SCM the courses here we're teaching Python now so if you want to go to grad school before you come any edX courses, great Python edX courses we're finding that for the large data Python is the language to go to it's easy to use and you can, a lot of libraries where you can do a lot of these different things so if you have free time, yeah it's open sourced learn Python we're finding it, I wish I stopped coding using C in the early 90s so learning new languages but I recommend it if you haven't done it already so we have four minutes left so I would suggest we tackle some of the questions that are more general, if that's okay with you yep so perhaps we talk shortly about the final exam, do you want to do that no, you can so final exam is what you have seen in the midterm the midterm is essentially the same as and reflects what we are going to do in the final, we're going to have problem sets plus checkbox questions and the entire content that we served you with from week one to four and was that seven to ten will be covered so we are going to ask questions about all of that content to see your proficiency so it can also be just to be very specific, it can also be the content of one week one through week four it's not just the second half of the course but other than that you can expect exactly the same it's going to be released today and no, tomorrow sorry, tomorrow it's going to be released tomorrow and we leave it open for five days for seven days until December 5th it will be a timed exam again so you will have four hours once you open it so if you decide when you want to take it but once you open it the clock is running and you have four hours remember that if you're taking it on Wednesday in our morning so right before the deadline you need to have four hours before the deadline because when the deadline hits December 5th at 1500 UTC we will close the exam so you need to account for that with the four hours it's an open book exam so you can use all the tools that you developed you can use the videos you can use the practice problems if that helps and you have time to revisit them you can take a look at your notes so it's open but it's all your work so you're not allowed to coordinate you're not allowed to work in groups that is absolutely forbidden I think that should be covering it what other questions do we have? so we have a couple of questions about the course setup so is it mandatory to take the courses in sequence sc1x and sc2? no it's recommended because if you take say sc2x without taking 0x you might be wondering what a mixed integer linear program is if you're not brushed up on your mathematical dexterity with algebra you might get tripped up on sc1x hypothesis testing is used a little later so you can take them out of sequence we recommend taking them in sequence but you guys have busy lives sometimes people want to take two courses at once you want to start so we try to set expectations and you can always go to the key concept document for the whole course and kind of go through and say ok what's going to be covered in this course am I going to be out of my depth so we recommend it but we don't enforce it anything else? so we have a couple of questions left but time is running out I think we've covered a lot today thank you very much for taking the time Chris that was really interesting to see all these programs and how we can actually use the tools I encourage you to take sc1x to me it was the first course and sc1x and 2x are my favorite courses because they really tie together things and you'll use them a lot and sc1x is coming up right? yes it is it's starting in first week of January so you can sign up and see how we think about inventory management forecasting all that kind of stuff give it a try even if you don't want to do verify just go for the first couple of weeks we'll start talking about demand forecasting you'll start seeing how these models and everything you did the pain you went through to learn these models how they're actually used so thank you very much guys the video will be up on our channel if you want to revisit it thank you guys see you soon