 Hello, global supply trainers. So happy to have you all here today for our second live event in SC0x Supply Chain Analytics. I am Imma Borrella. I'm a first doctoral associate here at MIT. I'm originally from Spain, and I'm the course lead of SC0x of this run. And most of you already know me, so welcome. The course is going great. We have more than 28,000 enrolled students and 1,779 of them are very fine. So great learning community. Today I'm really happy to have with me Dr. Matias Winkenbach. Welcome, Matias. Thanks so much for being here with us. Thanks for having me, hi. Dr. Winkenbach is the director of the MIT Megacity Logistics Lab. And he's a research scientist here at the MIT Center for Transportation and Logistics. And his current research focuses on multi-tier distribution network design in the context of urban logistics and last mile delivery, urban freight policy and infrastructure design, as well as data analytics and visualization in urban logistics context. And he has a wide and really interesting bio that you can consult in the link I shared with you earlier for the live event. So thanks, Matias. Today we'll be discussing with him a very interesting topic, urban distribution network design for B2B last mile delivery. So the agenda briefly will be as follows. First, Matias will introduce the topic, discussing about data-driven urban distribution network design. After that, we will let you go to the breakout rooms. We will propose a couple of questions for you to discuss with your peers. And you will have 15 minutes there and then come back to the main room again with your comments and questions. And Matias will continue his presentation and introduce the idea of how they will be accounting for a certainty in these network design models. And of course he will present a case study of a project they develop in Bogota, right? Yes, exactly. Okay. Then we'll have 10 minutes or so for questions and answers and then we'll wrap up the live event. Okay, sounds good. So let's begin, Matias. All right. Whenever you want. Welcome once again for my side. So I'm gonna talk a little bit about urban last mile distribution network design. Some of you might have already heard something about this. And as we heard before, the second part of this presentation will then focus a little bit on how do we actually do this if certain elements of the network or of the underlying market that we're trying to serve are uncertain. But let's start with the basics. And the basics are relatively well-trivial you might think but actually one big challenge in urban distribution network design is how do you actually capture let's say the complexity of an urban market? And complexity means for instance the geographic complexity. Cities don't usually come in a standard shape. They are not perfectly round or perfectly square. So you need to have a way of basically capturing the geographical shape of a city. And then you also have things like demand being not just uniformly distributed across the entire city but there are very non-trivial patterns in the way that your demand that you're trying to serve might be distributed. And the way that in our research we typically tackle this and here you see on the slide the example of Bogota, basically the capital of Columbia. You see that we, well, we call it pixelize the city. So we cut the city into a large number of city segments or we usually call it pixels. And in this example here you for instance see a number of, I think about 2000 pixels that we cut Bogota into an each pixel is basically one square kilometer of the city area. And you see with this pixelization we already capture kind of the geographic complexity of the city because we only consider parts of the landscape that are actually carrying demand. So where people are actually living or in this case study where actually some retail activity is going on so where there are customers to be served. And then what we can do for every single pixel is we can characterize the most important pieces of the input data that are relevant for the network design problem. So for instance, one or the most important part of the data is demand like how many customers do we have per pixel? So what's the density of stops that need to be served? Or what's the drop size distribution of this stuff? Meaning like when we visit one of these customers how many packages or cases or what kind of volume do we actually have to deliver to a single stop? And then the other side is the infrastructure that we're operating on like the road network infrastructure for instance. And again we can use those pixels to characterize that infrastructure in a high resolution kind of way. So we can for every single pixel for instance use open street maps data or Google API data to inform the model about the network capacity. So how many lanes they are, how many roads there are, how dense the road network is, whether there's any direct generality in there. So whether there are one way streets basically and also how direct vehicles can travel between one customer and the other whether they can almost travel on a straight line distance or whether maybe because of one way streets they have to make a lot of detours to get from A to B. All of this can be characterized mostly based on the open the available data sources like Google on the pixel level. So on a very high resolution. And the result might look like something like this. So the case study that we're gonna discuss later on actually produce this kind of map of Bogota. You see again the different pixels and every single pixel here represents one square kilometer of Bogota. And we just color coded it a little bit. So the color of the pixel tells you a little bit about the density of customers in that area. So we for instance, see that in the right part of the map you see that downtown that reddish area which is characterized by a lot of traditional retail for instance, so a lot of small stops that need to be served. And the green areas are characterized by fewer customers. So for instance, bigger retailers. However, the size of the pixel tells you something about the volume that needs to go there. So how many cases or packages do go to that pixel in total. And you see that there are some pixels that are pretty big and pretty red. That means a lot of volume, but a lot of customers or a lot of fragmentation. And other pixels that are similarly big, but much more green. So there we have a stronger concentration of demand. And this kind of complexity you can only depict if you work with this kind of high resolution discretization of the city because you cannot just assume that demand is evenly distributed and it's the same everywhere within the city that would not basically pay respect to the true complexity of an urban market. Now the question is how do we design distribution networks that serve this kind of demand? And distribution network design can actually be explained very simply. You have probably three major decisions. First decision is what kind of facilities are you using to serve demand? So how many distribution centers do you need and where do you locate them? And then if you have a multi-tier distribution network where you have smaller satellite facilities somewhere within the city where some sort of transshipment from larger vehicles to smaller vehicles, for instance, takes place. You also need to decide how many of these do you need and where do you locate them? That's kind of the first level decision, the facility decision. And where we usually model this is with a mixed linear programming model. So with a mathematical model it helps us optimize the network. And basically the decision on the number and location of facilities is mostly taken care of through binary decision variables. So you have a set of candidate locations and those binary variables tell you which of those candidate locations to use to locate, for instance, a distribution center or a satellite facility. The second level of decisions relates to the assignment of service areas to the different facilities. So let's say your first level decisions told you you have one distribution center and three satellite facilities. The next thing you wanna know is which part of the city is actually being served from which of these facilities. And again, this can be modeled within the mixed integer linear programming model as a binary variable, maybe also as a, let's say a floating point variable that is however constrained to be somewhere between zero and one. But for simplicity's sake, let's assume it's also a binary variable. And those binary variables, and there's a lot of them will then tell you which of the pixels that you saw earlier will be allocated to which of the facilities. And that basically then defines the service area being served from each of the facilities. And then the last decision level is actually the model choice. So let's say you have different vehicle types to choose from, a big 10-pellet truck, a smaller truck, and let's say an electric bus or something. Then this is gonna be modeled again as a binary choice, right? For every allocation between a facility and a city pixel, you then make the choice like which of those three vehicle options you actually want to activate to serve demand in that pixel from that facility. So it's basically a whole lot of binary or integer variables that we're working here. That's why we use mixed integer linear programming. Whereas the objective function is in the most cases, the total cost of operation. So most companies that we work with want to minimize the cost of serving demand. And that's usually then basically a pretty long formula that takes into account different cost components like routing cost, equipment cost, facility cost, and so on and so forth. The most tricky part is the routing cost. The routing cost need to be approximated because you can imagine if you're dealing with a city like Bogota, you're dealing with thousands of customers that might be served on a single day. That means you're dealing with hundreds of vehicle routes. And it's very hard to explicitly optimize vehicle routes for the scale of an entire city. So that's why we use approximation techniques and I'm going to touch upon that in the next slide. But approximation techniques that basically don't explicitly optimize the vehicle routing part, but basically just tell us what is a good estimate for the optimal length or the optimal duration of routes into a certain area of the city and to what cost to do these lengths and durations actually translate to. Obviously there are also different optimization objectives sometimes some companies don't want to optimize for cost but maybe for market reach. Maybe they want to optimize their emissions footprint in a city or other things like responsiveness to the customer service level. All of these could be potential optimization objectives. Most of the models that we built and that we see in the literature focus on cost though. Well, I mentioned that we use approximation techniques and in the course of this little session we can't explain this in very much detail. But the thing that I want you to take away is that doing a location routing problem, so deciding at the same time where to locate facilities and how to route vehicles around them is an extremely complex thing to do. It's a combination of a location allocation problem and a vehicle routing problem. Both of them are known to be was known NP hard so computationally very complex. And obviously if you combine the two and try to solve them jointly you get an extremely complex problem and that could not be solved in reasonable amounts of time. So we use continuum approximation which basically can look like something like this if you're really interested or reach out to me and I can give you some literature to dig into. But basically what it does it, it just uses geometric probability theory to basically estimate how long a vehicle would travel within a certain area of demand with a certain number of customers assuming that it would follow an optimal route. So we're not explicitly figuring out that optimal route we're just assuming that there is some optimization going on and then we can basically use probability theory to estimate the distances between two stops and therefore also the total distance of a route, the duration of that route and the cost of that route. And that's what a whole bunch of the research projects that we're doing here is actually leveraging every single day. And I think that brings us, so this was a very brief introduction to just network design as such without any stochasticity involved yet. And I think now you're gonna break out to discuss what could be done on top of that. Thanks Matias, it was a great presentation. I enjoyed it very much. It was brief, but we will have more after the breakout rooms. So please join the breakout rooms now and meet your fears. We will have some CTAs moderating or facilitating the discussion in a few breakout rooms. You can see their pictures and their names in the screen right now. And there's a couple questions we would like you to discuss. So it's just some ideas to guide your discussion. We would like you to talk about what you just heard and how this connects with what you have learned during the past four weeks in SC0X. So just discuss how the techniques that you've learned can be applied to resolve a real and complex network design problem like the ones they are approaching in the Megacity Logistics Lab. They are dealing with that every day. And the second question, we like to think that in all of our problems in SC0X, we have always assumed a deterministic context in which everything is stable, doesn't change. But we know reality changes continuously and the city is particularly a very complex environments, contents where there are very, a lot of factors that can affect our network design. So how would you take uncertainty into account when designing a network design model? So we know maybe you're not an expert in this area, but you know a little bit about network design by now. We've been learning about it in SC0X. So we would like you to take a thought. How would you introduce uncertainty to these models so they are more realistic? See you in 15 minutes.