Issues and Opportunities in Last Mile Logistics

all right thanks Andy thanks everyone for coming and yeah so we appreciate this opportunity to present some ideas to you the way we're gonna organize the presentation of course Martin and I do quite a bit of research together I think we're both going to talk a little bit more detail about some of the things that we've done separately and then Martin will cover sort of the second half of the talk so advances in last model Geist six is the title and here's a rough outline we're gonna talk about what's motivating this research it's primarily the first bullet point their growth in e-commerce I'll introduce you some basics in last mile logistics organization and optimization and then we'll talk about a number of innovations some in more depth than others okay so that's the agenda so let's start with e-commerce growth so something is changing has changed the world has changed significantly let's say and that change has largely been driven by the growth in e-commerce so there's plenty of statistics that you can look at that support this type of growth here we look at global measures of total ecommerce sales and billions of dollars over the starting in 2013 and then projected forward into 2018 the rate of change is slowing but you can still see high levels of growth projected 16 15 % growth year over year if you look at one of the biggest purveyors of e-commerce Amazon look at their growth in merchandise sales in spanning from 2010 to 2015 here you can see again this high level of growth almost approaching exponential growth and it's harder to read on this slide but what it's important to look here this red line is showing you the well sorry the blue here is showing you the net shipping costs in order to support these levels of sales and shipping costs are an extremely important part of allowing e-commerce to happen effectively so what we're going to look at primarily today are ways to manage systems that are driven by e-commerce and a lot and that's sort of what we're call modern last-mile logistic systems and trying to manage the costs of delivering certain levels of service one of one particular innovation that is people want goods faster and faster so there's this notion of same-day delivery it means different things to different people I'll try to define it a little bit in a second same-day delivery is also forecast it to grow sort of at least exponentially for a while maybe eventually with this s-curve type shape okay so people are demanding products in a different way and they're demanding them faster than ever okay so here's sort of the way I look at the fundamental change that's occurred in commerce right and you can see this picture here pictures like this are fairly common what is it it's a shopping mall that has the roof caved in and snow on the escalators right so the idea here is that there's a diminishing role in retail for retail stores so stores are basically local inventory points where shoppers go in look at goods make a purchase out of a local inventory stock this type of Commerce is diminishing and it's being substituted by commerce that provides product delivered directly to consumers okay so it started with catalogue phone sales a long time ago this isn't something new but it's really become more explosive with mobile mobile internet connected devices one thing to think about is that the definition of what a consumer is purchasing the product that they're purchasing is also something you should think about it's not just the physical item anymore but it's also the level of service that comes with that item right so you're purchasing a widget but you're also purchasing that you will get it in two hours or that you will get it in one week so that service element is embedded in some sense in products and the same physical product might be provided by different providers with different service levels and as they're sort of different products from that perspective when can I get it matters right and all of this growth here has led to enormous changes in last-mile logistics and the increased complexity of this systems so let's talk a little bit about the basics the focus in this seminar because of this growth in direct-to-consumer commerce is going to be on this direct-to-consumer type of logistics so making deliveries to individuals not necessarily making last mile deliveries to other types of facilities retail stores distribution centers etc so why is the modern form of last model logistics so challenging well consumers typically design they demand product sizes that are much smaller so shipment sizes are much smaller the order sizes are smaller and you think about if you compare how much an individual is going to be receiving in a shipment versus how much you might make delivering to a retail store hopefully you can obviously see a difference unless you like it my family where we get a large shipment almost every day of the V commerce case many constantly changing geographically dispersed locations of course so individuals are in their homes their workplaces you'll see later we might be delivering to their cars or we might be somehow trying to deliver to them wherever they might be at any individual point in time obviously individuals can be in a lot more geographic locations than the limited number of stores or distribution centers etc that were in traditional last-mile distribution and we're going to look in this basic section at understanding how these systems perform primarily looking at cost per delivery so I'm not going to talk a lot about service levels here we're going to assume that sort of a service level isn't and I mentioned this earlier the service level is sort of part of the product so the product includes the service level this is a product that I want in four hours that's the product now given that that's what the consumer wants how are we going to organize a system to do that effectively and we're gonna focus primarily therefore on cost if that makes sense okay so what's one of the primary drivers of last model delivery costs one thing to think about is that since these order shipments are small we're consolidating them into shipments carried at least presently multiple shipments carried by single vehicles that may change as we move more towards smaller automated vehicles we'll talk about that a little bit later but right now typically consolidation is important we'll talk about why so here you can see outbound delivery tours from a distribution center making multiple deliveries to multiple delivery locations and the one on the upper left is making a smaller number of deliveries and one of the lower right is making a larger number of deliveries and hopefully you can see that the density of customers and the number of deliveries made allows you to decrease the cost per delivery in the lower right right so I'm somehow spreading cost over more customers in this slot in this part of the slide then in here right you can also see that roughly perhaps if you think about the travel distance of these two vehicles they're not that different for these two sequences and we're spreading that travel distance over more customers and there may also be fixed cost so let's talk about some of these cost components more specifically now so if you look at last mile costs they tend to be focused in fixed costs and variable costs fixed costs are basically related to how well we're using the vehicle if we're using the vehicle at a high level of utilization then that means that we're using fewer vehicles fewer dispatches let's say fewer vehicles per day and that helps us reduce costs but this sort of a fixed cost per vehicle similarly with drivers so drivers are an expensive part of last mile logistics they tend to be paid per day or per per hour and these costs are also in some sense fixed so if you can get more deliveries per driver hour then you're going to be doing a better job with cost and then variable costs usually more focused on vehicle costs although sometimes driver costs are variable with distance traveled and often times distance traveled is the primary determinant here so if we want to have an efficient system what we like are short short routes that don't use a lot of travel distance if if possible that have high levels of capacity utilization so that we're spreading the fixed cost over as many deliveries as possible so that's a basic the other thing you should think about though is that this idea that density is going to drive costs down is somewhat conflicting a little bit with service levels okay so in traditional even traditional e-commerce let's say you would have let's place an order and we'll receive it in five to seven business days or more recently let's place an order and it will be delivered in two days service or three day service or next day service okay the growth now is more in this type of area same-day service or very short window I order now and it's delivered in two hours okay now what the challenge here is that the time between when you place an order and the time that you want to receive the order is small and that limits the opportunities for consolidation right so it limits the opportunities that there's other orders that can can be combined with your order to drive down these last mile costs so you can think about that as lead time slack the time between order placement and the absolute deadline that that order needs to be dispatched from a facility to meet your service commitment as that grows shorter that puts a pressure on delivery density it's harder to create tours or to create operations that have high levels of density in that case one last idea before we move into some specifics on some innovations is that you can also drive improve density by improving supply location density okay so in this system I have a single supply point let's say a distribution center this Center and over here I may have multiple supply points one of these might be a distribution center but others may be retail stores or maybe these are all retail stores and I'm filling my e-commerce orders from stores by by increasing the density of supply locations I'm also able to improve logistics costs here by improving local density of customer deliveries around those around those supply points right so supply location density also provides flexibility and helps you drive down car and of course if you follow what Amazon has been up to recently by building out a much larger set of distribution centers their goal of course has been to increase their distribution center supply location density for customers okay so now let's talk with the basics in mind let's talk about some innovations and some of this is motivated by research that we do here at Georgia Tech and I want to focus the first part of the research part of this talk on same-day delivery so along with my colleague Alejandro toriello who I'm not sure is here yet and our PhD student Mateus clap over the last few years we've looked at a number of challenging problems in dynamic dispatching to support same-day delivery so in this section of the talk I'm going to talk about our definition of same-day delivery which is order today receive some time today okay so it's not necessarily a two-hour guarantee like Amazon Prime it's a it gets a little bit of a longer window but it's still the still this feature is still that you're ordering potentially on the same day that you're receiving the goods and from the logistics side a point beside the supply side what it means to the company is that the company may be receiving orders for your to deliver while you're making other deliveries okay so they're still receiving orders but you've already begun the distribution operation okay so same-day delivery is challenging again because order size is small and you still have this idea of many possible geographically dispersed delivery locations but there's extra things the short lead time slack here so now we have a much shorter time to consolidate orders potentially and we may use our vehicles multiple times during the operating day so we may also decide when to dispatch orders during the day so there's some more complicated decisions about when we might decide to dispatch an order given that it's ready to go and what we call that is maybe dynamic updates to this in order to drive cost savings I'll try to illustrate these ideas and show you some of the things that we've learned in the next ten minutes or so okay so to motivate this a little bit more suppose over here suppose I had all these orders for next day delivery and they happened to all be available and picked at this distribution center and I built two vehicle up tours for distributing those Goods today so these all these orders were placed sometime before today they were picked and packed and loaded and ready to go and I could create these nice to nice tours on the Left if we look at that same set of orders now and just imagine that they were not all placed in advance of operations right so in this case the green orders might have been placed between 12:00 and 2:00 p.m. the brown orders between 2:00 and 4:00 p.m. the blue orders between 4:00 and 6:00 p.m. and suppose they're all due the deadlines for this same day operation is something like 7:30 p.m. okay so they all need to get done by the end of today but some of them were placed quite close to the deadline right so if we waited until all the orders were placed let's say that the latest order came in at 6:00 p.m. I would love to operate these two sequences but I could never make all these deliveries in time to meet that 7:30 deadline so I need to do potentially something different one of the simplest things I could do is just say well let's just create three dispatch times let's say 2 p.m. 4 p.m. and 6 p.m. whatever orders have arrived by 2 p.m. I'll create operations to deliver them then whatever orders arrive by 4 p.m. I'll create some delivery operations to deliver them and so on and then you get towards that looks like this you can see there's a lot of overlapping and hopefully you can it's not too difficult to see that the transportation distance of these six tours is much higher than what you saw on the two tours there you might need additional vehicles to make this happen although maybe you can just reuse these same two vehicles three times per day to do this ok so this is sort of the different paradigm that we're looking at with same-day delivery so when it comes to decision support one thing that we wanted to look at in our research was what if we don't consider fixed dispatch time so what if we don't just say we're going to have a dispatch at 2 p.m. 4 p.m. and 6 p.m. let's instead dynamically adjust what we do every day so there still might be some discrete opportunities for dispatch but we get to select based on current conditions based on what we're seeing today in terms of demand what customers have ordered today where where products need to be delivered today we're going to dynamically decide when to dispatch vehicles and then of course at this time this is a dispatch time here I can dispatch any order that has is picked and ready to go before that time at this time I can dispatch orders that arrived between the first dispatch and now but I could also dispatch orders that I just decided not to dispatch in the first outbound dispatch and there might be a good reason why I want to delay some of these to a later dispatch if they don't mesh well geographically with the ones that needed to be dispatched earlier so that's sort of the system that we're looking at here ok and what you see here whose operations for a single vehicle it waits until it's this first deadline loaded dispatched back to the Depot reloaded dispatched on a shorter dispatch and then eventually completes a third dispatch before the end of the operating day ok so if we look at how to make these decisions there's always trade-offs in terms of cost and and the trade-offs in terms of when to dispatch a vehicle if we think about waiting if a vehicle waits before dispatching it allows us to accumulate more orders and then potentially improve that delivery density metric that's so important to help us reduce cost so waiting allows us to improve delivery density but it also wastes time because if you wait then the vehicle is not making productive deliveries in the field and it's in some sense an opportunity cost so that's the trade-off and then how about dispatching orders if I dispatch an order right away I know I get it done and I get the let's say the revenue or I don't any sort of penalty situation where I don't get an or dispatch meet service but if I waited I could have potentially found later orders that arrived that match up better with that order and crepe that are route sequences right so if i if i delay it maybe i can find better partners for that order later in the day so there's these types of trade-offs that we have to balance in these systems so we looked at this from a research perspective we looked at a single vehicle example and we've done some analysis using two types of two types of paradigms let's say and the first we look at a very simple system where we imagine here's a distribution center and all the customers are located on a line outbound from the distribution center we know this is not realistic but it's going to help us at least understand these types of systems okay and we can still understand when to dispatch and now it's a little bit of a simpler decision about which customers to serve because what we're going to assume now is that if we dispatch for example at this time w1 all we need to decide is really how far along this line to go and we're going to assume that we can serve any customer that arrived before then that is closer to the distribution center than the furthest that we go in the in the sequence so it sort of simplifies the decision we don't have to worry about routing and scheduling type sequencing decisions and our goal is going to be maximizing some reward for serving orders so maximizing some reward for getting orders done minus the costs related to the dispatch distances here and the number of dispatches let's that so what you should have shown in this picture is actually a plan let's say time goes forward left to right and distance from the distribution center goes up on the board well the red dots represent when customers show up and the blue lines represent a dispatching plan so at this time I dispatch but I only serve this one customer I come back to the Depot by this point and then I wait until w3 I dispatch and at that point sorry that it's not always working the laser but I can cover all the other red dots that I didn't cover in that blue square but the ones outside the Blue Square are not covered I get back to the depot at time w-4 and then I make a smaller dispatch and I cover this group in here the ones outside the blue are not served there's sort of missed opportunities and you can sort of then calculate a reward - service cost now suppose let's consider the following unrealistic situation now suppose customer orders do arrive in time and the red dots represent right along the time line when they arrive but suppose that at the time zero here I know in advance exactly when court orders will be ready to be dispatched so it's I sort of have a perfect forecast of the future so I know exactly that this order is going to arrive at this time it's not going to arrive before that time so I can't serve it before that time but I know it's gonna arrive first with certainty at that time well that's called a deterrent that's a deterministic problem on the line and if you look at this type of problem we can show that optimal solutions always look like this there's always waiting for a while at the distribution center and then at some point vehicles are dispatched and furthermore the vehicles are dispatched such that the first dispatch is the longest the second dispatch is no longer than the first dispatched and this so they're non increasing over time so this is the longer one and then there's a shorter one and you can show that an optimal solution always has this form and furthermore because in this simple line setting we can show that it's not difficult to have an algorithm to find the best possible way of trading off this reward maximization and cost minimization in this setting there's an efficient algorithm for finding an optimal solution here what's interesting is because of course in the real world problem the arrival times are not known right you don't know at this time all you know is you have these two orders that are ready to go you don't know when these other orders are going to arrive necessarily so suppose these future order arrival when x or you use the word arrival time but it's really the time that the order is picked and ready to be dispatched suppose they're modeled is some random variables in this case we might want to come up with a system that optimizes the expected reward minus cost that we would earn in this in this system and this is a very difficult problem this is among the hardest class of computer science problems in the np-hard class you can formulate it as a dynamic program and solve it recursively but you can't get very far when the problems get of reasonable size there's another approach called a priori optimization that we take advantage of in this research which is the following suppose I am going to develop a fixed plan so I'm going to decide that I'm going to dispatch at this time and I'm gonna go this far and I'm gonna decide at this time and go this far in advance I'm gonna fix that in advance and I'm going to select it such that it minimizes expected cost you might have some very simple updating rules like never go beyond why travel further than the furthest order you need to deliver that kind of thing but other than that it's basically a fixed plan and is this problem any easier and it turns out yes it turns out this problem maximizing this expected revenue minus cost is no more difficult than the deterministic problem and I sort of describe how that works here you just replace uncertain customer orders with multiple copies representing each possible time the customer order might show up with a probability and then it's you can use the similar algorithm or the same algorithm to optimize that problem so it's actually not difficult to solve optimal opry re plans that build these plans in advance that minimize or maximize the expected revenue minus cost not only does this hold for the simplified geometry but we also show in our research that this holds when customers are not just located on a line so you can show that as well ok last thing and then I'll show you some results here of course developing one fixed plan at time zero here is a is also not necessarily ideal and what if things are quite different today than you expected what if there's an opportunity today to take advantage of changes that happen in reality adapt to them and change our plan so what we actually are proposing is that we're going to use develop dynamic plans for these systems where I'm going to continuously think about changing my plan in real time in response to conditions and one very effective approach is actually using that a priori optimization strategy in what we call a roll out scheme so when the roll out scheme is very easy think about this at any time the vehicles at the distribution center remember this is a single vehicle problem I'm just going to compute a new optimal a priori solution from now until the end of the horizon and again I told you that that isn't too difficult to do so you have a new set of known orders that might change things because you know at 2 p.m. things might look different to you than they did at 8:00 a.m. okay so use that new information and compute a new plan and then you just dispatch the vehicle on the first route that's proposed by that a priori plan do what it's said to do you're back at the distribution center and then you repeat that's called a roll out of this simple strategy and we find that this approach works very well so using this idea of opera optimization in this repeated rollout setting so here's some results for example for problems on the line what you see here the gray line represents just looking at the AA priori solution and the axis here from 0 to 18 this is the percent gap between the solutions that we find by our dynamic policies and the best one could ever do and the way we estimate the best one could ever do is we imagine that you had perfect information so we use the perfect information case to decide what's the best we could ever do so this are gaps with an unrealistic lower bound and actually shows pretty well that a priori solutions are not so good with small numbers of orders but one you get large number of orders a priori solutions work almost the same as dynamic solutions and then the gap between the grey and the red line here or the blue line either one they're sort of minor technical variations that we don't need to worry about it represents the savings that you can get by finding a dynamic plan right so for example for about ten customers I can reduce the penalty from 17 percent over optimal down to like 7 percent so I've dramatically improved operations by rolling out the opera plan rather than just computing it once in advance so to finish up this part of the talk I do want to talk a little bit about more realistic geographies right because we all know that customers are not located on the line and to be honest to be one problem with problems on the line is that there's a lot it's probably the maximum amount of benefit for pooling of customers in that case because we assume that if you travel you know this far that you can serve all the customers that are closer with no marginal cost and that's a little bit of a strong assumption so we wanted to look at more realistic geographies in these types of settings there may be travel costs that are defined by a network or they're defined by like a metric on the on like an l2 metric for example you can also incorporate times where the vehicle has to stop at a delivery here which also increases realism so we've looked at this more realistic case and in this case notice that I still might need to decide when to dispatch a vehicle like at time 3 here and again at time 2 I apologize that all the times are always counting down to time zero it's something that Matthias liked and I I always had trouble remembering but the higher time is earlier in the day okay so you still need to decide when the vehicles are gonna be dispatched but now when the vehicle is dispatched you need to decide which subset of orders to serve and in which sequence to serve them okay so if you look at the plan here for example at time 3 I dispatched this red route and I've decided to serve this customer this customer this customer this customer the labels in the the nodes represent the time that the order was picked and ready to go so notice that for the route that was dispatched at time 3 all of these numbers are greater than or equal to 3 which means they were already somewhere greater time were available at time 6 some at time 4 and time 3 remember time is counting down just to confuse you once you get to time 2 you dispatch the blue route and notice that you can now see some twos in there right but by the time I complete the blue route I have more time left so the ones that showed up at time one I just can't serve in this case so there I don't get the reward for serving them this problem is much more difficult than the problem of dispatching on the line because it also combines for the students who are here I mean you might recognize that there's some traveling salesperson type problem going on in here as well as everything else I've been talking about so the deterministic dispatching problem is more complex but if I did again know the orders and the arrival times of the orders in advance there is something I can do we formulated this as a modified variant of an extension of that traveling salesperson problem called the prize collecting traveling salesperson problem that's a fairly well known problem and we we changed that again to another extension where it has order release times in other words the orders are only available to be dispatched at specific times during the day and you can use multiple dispatches from this distribution center but they can't overlap in time because this is going to be one vehicle solutions okay we have a formulation that solves this problem fairly effectively using integer programming methods and I'm not going to go into the details but it it uses some advanced integer programming methods doing some customization on the branch and bound and heuristics again we're gonna use the same paradigm to solve these problems we can show that this same price collecting modified price collecting TSP which we can solve for the deterministic problem we can again solve operatory problems with that right so you go from deterministic to a priori and then we're gonna roll them out again now the difficulty here is that it does take more time to compute solutions here so one thing one modification that we've done is that we don't resolve the operator problem until the time that the Opera or a solution decides the first vehicle dispatch but in general it's roughly the same and I just wanted to show you as we as I finish up here how this again this roll out policy performs again it performs much better than the uh priori strategy so here the gap of the operator strategy is 23% again to that best possible case and we basically cut the gap half right so you get a 50% reduction in the penalty due to uncertainty and these we believe these solutions are pretty good it's hard to say that this how far this is from the optimal solution but we believe that it's much closer than 12% probably more like within the 5% range you can also see that the policies on the right here and sorry the laser isn't working fr is the fill rate this is the fill rate here so more orders are getting fulfilled when we use the rollout policy versus the operator policy and notice that the next line the duration poor order that's sort of the cost per delivery okay and notice that that's not increasing very much but we're driving up the fill rate so we're basically making better vehicle utilization decisions with our rollout strategy there's another strategy that's somewhere in between it's a little bit simpler than the full rollout I'm not going to talk about it all it works about half as well it's basically halfway between last point I want to make is one thing that we found out in this study is that if your goal is to deliver it make as many deliveries as possible it carries a stiff cost so maximizing fill rate is something we do here by increasing this parameter alpha as alpha gets bigger the objective focuses mainly on just getting as much reward possible and ignoring the transportation cost and you can see that if I want to drive fill rate from somewhere between about 85 percent to 91 percent I had to almost incur or incur more than a 50 percent increase in my travel cost per order and that's a fairly interesting result so six percent increase in fill rate requires a 50 percent increase in cost per order we found that interesting and here's looking at two different plans the plan on the left is the balanced plan that is sort of balancing fill rate with transportation cost the one on the right is trying to drive as much fill rate as possible and you can see they're only different they only serve the what the plan on the right only serves one additional customer but it costs a lot more to do so so one as we've looked at furthering this research or going to the next step looking at problems about deciding which customers to offer same-day delivery service - or not we called and that's probably the most important bullet point here we're also looking at multiple vehicle problems and problems of adding capacity dynamically like search type capacity but I really want to focus on the first one you know in an online ordering environment when a customer is browsing you can decide what types of delivery options to show them so if you're just if you it you can respond dynamically to conditions on the ground and say right now based on what types of capacity I have left to make same-day delivery orders today maybe it's wise if I don't even show them the same day delivery option right it's better not to offer that right so that's what we've looked at in these accept reject problems using similar frameworks anyway so that's a summary of some of the research that we've done here in same-day dispatching it's been fun to work on and later on I'd be happy to talk to you about possible ideas that you might have along those lines all right I will get started since I have more slides and less time so we'll do some online scheduling to see how we deal with that so I'm not gonna go into much detail on any of the research that is related to the topics that I'm gonna discuss my goal here is primarily to introduce some of the ideas that we've been working on and how they relate to the topic of last-mile logistics so when I'm may go into one slightly deeper so of course companies have realized I mean the slide that Ellen showed about Amazon that shows that as a percentage of revenue the amount that they're paying for shipment is going up is the part that worries them right so in a way they yes they increase their revenues which is great but as a percentage they're spending more money to actually make the deliveries and that's the part that concerns them and so and it's not only true for Amazon that's of course true for any company so all of them are looking at cost reduction strategies in this business to consumer space right and we're gonna look at a few of them that I listed here and the two main ideas here are that delivery density is important but we've argued that last mile delivery is challenging because we have all these changing delivery locations that can be anywhere as opposed to simply delivering to stores that don't move are always there and want big loads right so one of the ideas is can we somehow change where we actually make the delivery right so that's the first one and the second one is can we sort of re-engineer our business processes or start thinking about how we might deliver differently right and so the ones that we're going to look at is box delivery trunk delivery drones and crouch shipping right and there are others but these are the ones that we're going to look at right so one of the ideas is that maybe rather than going to a consumers home if we have a parcel box that they pass on their way to work on their way to sports on their way to church at the subway station right we can use that to make deliveries to individuals right and note that here of course you make deliveries to many people in the same location which from a logistics perspective is much better plus if you manage to place them closer to your Pillman Center you safe as well right so this is the first idea many companies are looking at these so why are we interested in it quantity variation is reduced right I mean I'm always putting deliveries in the the box the box is always in the same location so I can sort of anticipate that and if I locate them strategically they will be closer to my supply point so lots of advantages of course why is it interesting to us because it leads to a set of new questions that we haven't addressed before right how many of these boxes should you have where should you locate them what should the design be and how do you actually route packages to that and I'm not going to touch upon this very much but obviously in the dynamic setting that Ellen spent a lot of time on this gets even more interesting right you swipe your credit card to pick up your package out of the box now that box is available for somebody else but if so you need to resupply right so the dynamic settings here are always more challenging and in a way more interesting the other idea is to rather than introducing parcel boxes what if we deliver to somebody's car right other manufacturers now all start including technology that can allow you to give access to the trunk of your car controlled a one-time access right so companies like Amazon and others are quite intrigued by this option for for two reasons right one if 10 percent of us had ordered something on Amazon yesterday and we would have all come by car they could make the delivery right here on campus as opposed to go into all of our homes which is probably cheaper the other thing and this is more important in Europe and Asia is if you make it delivery somebody has to sign for it and so miss deliveries is a major problem in the business to consumer space and if the legislation is there that if you go to some of the trunk of somebody's car it's considered as an accepted delivery because you gave them the access code then that problem may go away right so this is certainly somewhat something that people are looking at again what are we saving here location flexibility right because we don't have to go to somebody's home necessarily we might decide to go to a location that is more convenient to us closer to the fulfillment center for example so here is the idea if you think a bit more about sort of the research questions so the brown dots are home locations the blue dots are locations where the car is going to be during the day let's say at work at soccer practice at the grocery store and I have a depot or fulfillment center so rather than having to go to the homes I could decide to deliver like this right so some I deliver at the home location other idly Fleur closer and this would be if I only had home locations as an option right so you see that the blue looks a bit shorter than the brown one that's the idea what are the challenges well clearly if you think about this I would have to know where the car is going to be during the day right and the other thing and that may not be such a challenge in the future a lot of companies are betting on knowing your travel behavior and if I go to work every day probably they know my car is gonna be in the parking lot here so but that's certainly a challenge and then there is some uncertainty right sometimes I go home at 6:00 sometimes I go home at 4:00 right you have to be able to deal with that and the routing problem gets more complex which of course is interesting for me the other option is drones we've all seen some of these pictures videos and that is still something that is unclear exactly where things are going to go but it's an interesting opportunity and again there what are we trying to do it is one eliminating the driver expense that you have you don't have a person on a truck making a delivery and what they promote primarily is faster delivery time right you can get something in half an hour instead of an hour or two hours again there are lots of interesting question if you decide that this is something that you want to pursue how many should you have which order should you assign to drones etc etc another thing that is interesting and this is just one particular example that at least intrigued me is that of crowd shipping right and here the idea is something that was proposed by Walmart is what if customers that come to my store are willing to make a delivery of an order that was placed online right how could I exploit that right so again the motivation is it could be cheaper because I may not have to reward them as much as my own drivers but obviously there are also challenges right so the idea here in terms of pictorially what you can do is brown dots represent locations where I have to make a delivery blue dots represent locations where somebody that's currently in my store is going and that have indicated that they're willing to make a delivery for me then I could have a combination right I could have a some order still done by company vehicles but some done by the in-store customers and this is the idea of crowd shipping right so very interesting questions there especially in this part particular said thing they're interesting questions about compensation right how much should i reward an in-store customer what flexibility does the in-store customer have and obviously the two are probably related right and we hear a lot about uber Freight they're all in the same realm of research questions that we're studying reduce fleet right fewer company vehicles fewer company drivers is what's driving the decision right or part of the decision-making but of course there is uncertainty in terms of these flexible drivers whether they will be available or not how do you balance that fleet sizing number of drivers which should be fulfilled by in-store customers a whole range of interesting questions that were researching I want to spend a little bit of time think talking about what I label here as the holy grail of last mile logistics in the sense that the service requirements here are probably the tightest from anything that you can think of right so we've been working together we group up who also operates here in Atlanta where you can go on their app say I want sushi they show you a bunch of restaurants you pick your favorite one and they make sure it will be delivered at the location that you want right and of course since its food we don't want to delete to place the order at 6:00 and then have to wait until 9:30 typically we would like to see it arrive within let's say 45 minutes right so there is an awful lot of pressure time pressure to make this happen right what's interesting to us and this is a statistic that was probably I don't know a few two months old maybe three months old right the scale of this problem gets larger and larger right they're currently making at least in this particular instant I looked at their website 250,000 deliveries per day now of course they're not all in the same city but the scale is another interesting question for us from a research perspective how do you velop technology that can in a few seconds deal with that many orders and still be cost-effective so I am going to talk a little bit about the challenges that you face and some of the issues that come up right so they're high surface expectations right and also from the company this it so first of all it's important to realize that they make most of their money from the restaurant not from the consumers or the diners right the restaurants have a contract with them and say we'd like to reach a larger market if you can do the delivery for us we can right so but both of them are interested in what they call click to door which is I push the button I want my meal when does it actually arrive at my home and of course ready to door freshness is important it's ready at the restaurant when does it reach my home it's highly dynamic right this is what makes it interesting to us right all the placements can come at any time from any location from any restaurants that we have in our portfolio right and these things differ by day of week by hour of the day by weather conditions etc right so it's it's a fantastic place to experiment with last-mile logistics technology so what are the challenges right how many drivers should I have on a particular day right is there a Braves game does that change anything right how long should the shift be should I employ drives and by the way these are never company employees right these are more the uber type drivers that say I have a few hours available I'm willing to make deliveries for you right so should we do that in chunks of two hours of four hours three hours what is the best strategy when should they start should we actually deliver multiple orders in a single trip right by a single driver and we'll come back to that in a minute if we do multiple orders in a single trip should they come from one restaurant or from multiple restaurants so typically if you think about the problem that we're facing right we have orders that specify a restaurant we know when the click happens right that's the placement time as soon as the click happens there is something in the back that says it said this restaurant that restaurant typically at this time of day takes 15 minutes to prepare so we have a planned ready time for the order we have drivers and we have different objectives that's also an important aspect so here again is the performance metrics most of surface that is related to click of door we would we have a target right drop up has a target that says about 45 minutes if we have a click to the door of an hour we are not making sort of our target and similarly with freshness and a lots of questions about that driver utilization of course it's important if we have low utilization the only way to serve all the orders is to have more drivers and drivers are only happy if they have a high utilization because that's their compensation scheme they are they tend to be I mean in a simplistic view paid just per order that they deliver now they're sort of incentive schemes and bonuses etc but that's the basis so they too want to have a high utilization the underlying optimization methodology that we use is a matching problem or an assignment problem where you can think of drivers on one side orders on the other side we look at the combination we somehow have an evaluation about the attractiveness of having a particular driver do a particular order and then we solve and this is a standard optimizations problem that we can solve in seconds even for large problems and of course Alan talked about this rollout strategy we do make these decisions every minute every two minutes right so the information is constantly updated and our decisions reflect the latest state of the system what are some of the challenges fairness right so we like probably to be fair to each of the diners even though one may be in a very busy area the other may be in a less busy area we cannot simply say we only serve the people in the busy area because it's cheaper to us right so how do we balance that utilization and diner experience right if we put five orders in a single trip we probably make better use of the driver but automatically four of the five orders are not going to be as early as they could be right so how do you make that trade-off order information is revealed over time the other thing that is very big is the system intensity saying the number of orders that are in the system of course varies over time right there is kind of a bump at lunchtime and then there is a big pump bump at dinner time how do you dynamically adjust to that and if it rains suddenly the bump is even bigger right so system intensity is a big issue I think I'm gonna skip some of the details here as I mentioned system state is important what characterizes the system the number of currently unassigned orders right that have just come in for example or I haven't assigned yet the drivers and the intensity because our decision ideas or technology depends very much on the system state for example in very busy periods we have no choice but to put more orders in a single trip just because we need to utilize the drivers better and we sacrifice a little bit in sort of the freshness for example so bundling of orders is important if the system is very busy there is no way we can serve all the orders if we don't bundle orders so then there are questions about how do you do that what's the effective way to do it should you have single restaurant order bundles or multi restaurant order bundles we generate them in a certain way in a way they're very small vehicle routing problems we also concatenate some bundles into a bigger bundle just some ideas that I wanted to share with you we also distinguish between the orders in the system right orders that have been longer in the system that's the way to think about this become more and more important right we may not be able at any time to treat every order the same way some will have to remain in the system a little bit longer we want to recognize that they've been in the system right so we actually think about this driver to order assignment in sort of a staged organization there are also some issues about commitments we always make a complete plan at every point that we run an optimization we make a complete plan we don't necessarily have to immediately start the execution of that plan right and that depends a little bit on these situations right sometimes because we optimize every two minutes if my current decision doesn't really affect anything in the next two minutes I might as well say okay it looks good now I'll revisit in two minutes when i optimize again so that has to do with commitment strategies alright a few practical complications communication with drivers might be lost right this happens in practice they may be out of the car making it early and are not looking at their little screen if I have a task for them they have to say yes or no if they're not answering within a minute what do I do right do I assign it to a different driver do I wait that's a practical issue drivers since they're not company employees do what they think is best for them not what is good for the company and of course you have to try with sort of the compensation scheme and incentives to align these as best as you can but there are some major issues right intuitively the driver believes that he's if he's making deliveries in Midtown that may be better than somewhere out in the suburbs where the density is much more right whether that's true or not is a different question but it means that drivers tend to sort of move into the same area which actually is very bad for the drivers themselves but certainly bad from a delivery perspective because I may not have drivers where I need them right and that's a big problem and then there is bunches of uncertainty that we have to deal with all right so this is a phenomenally interesting problem where we get lots of new research questions that were investigating and they take some of the ideas that we generate and put them in their production systems I did want to switch topics a little bit and talk about something that I also find interesting which is a little bit of a different flavor right in all the ones that I talked about up to now really in essence the research that we do is creating better routing and scheduling decisions right that's the focus whether it's vehicle routing algorithms dispatching technology matchings the focus is on improving the decisions that we make here the idea is the complete opposite and so first of all this is probably very well known more and more people are going to live in urban areas closer together more in people more and more people will order online maybe we should rethink how we do this business to consumer delivery and so I want to think about innovative delivery strategies or at least practical delivery strategies that work and this work was motivated by Amazon China and how they organize deliveries in for example Beijing right so of course as a researcher you say ok I have an area I have a distribution center I have satellite or intermediate distribution facilities this is a phenomenal fantastic interesting routing problem right a to echelon vehicle routing problem and there is a bunch of research on how to solve these what happens in practice is something completely different right they don't run routing algorithms they say ok I have a distribution center a city distribution center that's on the outside of the region I divide my sort of Beijing up into 13 or 14 regions I have a satellite in each of these regions within a region I divide everything up in cells and the idea is I give a cell to a driver and let the driver go off and figure out himself what to do right no fancy routing but in a way a bit of a design right how do you design the regions how do you do the cells and cells are usually relatively easy because boundaries have to be wrote etc right and so we wanted to know if that's actually a reasonable strategy or not right and I put it there our intuition is that if you have a very high density this actually is quite good you don't need fancy routing software to do it right but we wanted to see if that's really true right and in the process we came up with some interesting thing so maybe the most important thing to mention is what it says there that so at the moment Amazon's thinking is reaching and then sell and one driver personal right we found that you can do much better by adding one bit of intelligence and that is actually combining a few cells and calling we call that the block and then assigning two or three drivers to the block and still on each individual day I look at the orders in the block I do something very simple if I have two drivers per block I give half of them to one and half of them to the other but it helps you deal with the variability that you see on a day to day basis it's also extremely important when you start thinking about growing a market right if you don't know the exact densities that are gonna be in each cell and so we developed well we did a bunch of tests on a sort of a part of Beijing where we have four regions eighty-seven cells the region is listed over there the other something interesting that comes up as I mentioned the block concept seemed to be turned out to be very important so then there is a new optimization problem that comes into play that is how you are going to create the blocks right and that leads to relatively simple optimization problem so what we do here is we have a target number of orders or a target density for a block something that we can handle with either two or three drivers and now we of course we also want the cells that we put in a block to be somehow contiguous right I mean you don't want to assign a drive for two cells that are in in different parts of the city they need to be contiguous and so you can see at the bottom that based on how you do this optimization they may start to look a little bit different in shape right some are the center here is the satellite is where the products come from and you can see that shape will have some impact on the driver picks up can immediately start delivering or first he drives a little bit and which one is more important so there were a number of interesting aspects to this problem the remainder of the slides are mostly result so I'm going to skip these the one that so the one maybe that I will point out is the one on the left if you do things purely by cells it doesn't necessarily work that well in every conditions as soon as you add blocks which is a very simple simple idea bang you can do almost as good as using sophisticated vehicle routing algorithms for almost all densities which is interesting so because we're already a bit on the late side I want to end by saying that last my logistics research is a lot of fun is very important and it's all about effective and sustainable delivery today and tomorrow right and as I said one thing is better decision technology but the other is really rethinking about what strategies are gonna work right and we try to do both all right I think Ellen and I have given you at least an overview of the kind of questions that we've been working on we feel that there is still an awful lot to do in this area and we will continue to do so and we are happy to answer a few questions if there are questions and of course if there are no questions then we hope you all enjoy the rest of your afternoon orders always to restaurants and so if we know that certain restaurants are likely to be busy attendance for now we may already steer our decisions towards making sure that drivers will be at that location so all these are good questions to ask in this summation is a big part of the research that we're studying what is it about combining a few cells into a block that makes it perform so much better than just aggregating some of the variability so the simplest way to think about it is if you are very versed and that still today and so drivers can typically they think let's let these be workers they never cease to be if you're very restrictive and say here's his help that's your cell that's what you do every day some days there will be 30 words in that sells home base that will be 40 someplace that would even be tasty topic that it's not necessarily be takes a week to make it excel that might just be a little different so you either way reduce variability find English mr. yes it's also the case especially you say they didn't see at the moment that were there well that cell only this 25 but tomorrow with these you you're able to either their humility did I just make one we're coming about this type of this world it's basically a concept there's been a lot of research and a lot of different areas some of its transportation logistics of its production and flexibility and shows that just a little bit of flexibility like assigning a region to two drivers or the three drivers provides almost as much benefit as if you shared all your resources entirely over an entire system and that which I think so I even some research that I did 20 years ago I saw something very similar and stochastic erratically you care enough to work together there are also some hidden benefits if you will so you have a driver who's specializing in one cell or in our case they are specializing that their driver area knowledge of how to get from point A to D C and what you know so you want to keep that job for several years I mean they are experts on this area so when they go on vacation right or they're out sick or something happens and they can't do that or you're throwing them out there and they have no area knowledge so if you if you design your racks so there is some overlap or flexibility over time you have this kind of team of drivers who kind of know the area they can cover for one another and so that can really show up in the algorithms within a real world it can be a lifesaver you guys talk about pricing efficiency right oh sorry any discussion for efficiency to talk about diving costumes what about the other side the revenue stream like what studies in the done around maximizing revenue with the efficiency and customer satisfaction because Amazon's using the billions of dollars in here because the median customer expectations that came up especially complicated because airlines of course are a prime example there I mean it was an interesting reason responsive a philosophy certainly helped enact and at least changing need to be is growing marketing if that's costing us money being market size is gonna be eternal and for example the whole idea of eventual prize two days [Music] so doesn't that sort of comes back to Christ exactly so in all of these problems there's our Houston so can you drive up demand in times that could take advantage of times we may have excess capacity there's a lot of work to be done these tests promise especially especially integrating it with with the logistics costs I think there's less research that doesn't do chocolate air thanks to you it's up to me yes sure yes Madison yeah we know that the today's table spoke experimental for given service no I just needed to see what was happening [Applause]

Channel: GTSCL

Views: 33,899

Rating: 4.9248824 out of 5

Keywords: last mile logistics, supply chain, logistics, Georgia Tech

Id: SChmVn8A33w

Length: 71min 37sec (4297 seconds)

Published: Fri Oct 06 2017