[00:00:06]
>> When I was asked if I wanted to give a presentation in this series nice year of course I said yes. And I had to give a topic 6 or 7 months ago. And I decided that I'm surely would be able to say something about. The sharing economy which is of topic that has intrastate me for a while.

[00:00:36]
When I actually started preparing to talk decided to. Do a bit of a combination all from my perspective all the developments I see in these areas. And focus also a little bit but not very much on some of the work that has been done here and you were to take in this context.

[00:01:01]
So it's a mixture it's a bit of a perspective view introduction maybe to some of the questions that I think are relevant and important and a tiny bit also reflecting on work that we've done here with some of the companies that are actively involved in these questions. So what I want to do is 1st talk a little bit about last mile believe 3 I mentioned believe in the title but time really thinking about last mile believe free here then I want to introduce the concept of sharing and geek economy.

[00:01:39]
And then I want to focus a bit more on meal belief 3 specifically where both of these are coming together. And it's also a very popular and interesting topic and that's an area that we take have worked on extensively I have to say but as I say if I want to start by just looking at beat that last mildly free which was also the starting point for us I think here at tape in terms of what we wanted to work on or have been working on and.

[00:02:16]
Surprisingly Things have changed to be around I'm not exactly sure why in the presentation I had that one was in the upper left hand corner and it clearly shows that we're talking about Amazon here although most of you may have recognized the so little bit but so why a slight on Amazon I think Amos along certainly has been partly responsible for sort of the attention that last mile delivery has been getting and so I wanted to point out a few numbers.

[00:02:55]
So. The number that's now in the middle of about 150000000 mobile users x. aced the app in September so there are a few things I wanted to point out of course that it's a large number but also that this was purely only at Accies not on your home computer or your laptop it was through a mobile phone.

[00:03:20]
Right so and we'll get back to that later the other thing that I think is very important to realize is the number of products that these least at the upper right hand corner right I remember very well because I'm among the older ones here in the room that emotional and start selling books.

[00:03:42]
Right. Now their portfolio be. 119 or about 11020000000 different products that you can buy online right so that's a phenomenal change. Also at the your lower left hand corner Amazon Prime day we chased their big marketing employee to get you to buy things and you get discounts they sold 100000000 products right.

[00:04:19]
So one of the things that interest a kid in me or people like me is that and this song has really changed the scale of the beast right. It's also nice that here we have $4000.00 items sold per meaning. Right that leads to huge changes in how you can think about optimization take knowledge to support of Operation is like these which made it very interesting to all those there are a few more statistics one of them is prime Amazon Prime I expect that at least some of us in the room are Amazon Prime emerse I'm one.

[00:05:06]
What I want to point out mostly in this life or the information on the right if you look at why what do prime ministers like about the surface and in it captures 2 important pieces of information about last mile belief are right they like that it's free. And they like that it's going to be at their home fairly fast within 2 days right and this seems to characterize what's happening at least on the customer side in last mile the free people don't want to pay too much for it prefer a lead free and they want to half things as fast as possible right and this is where the logistics person starts to get kind of ichi because that makes things challenging.

[00:06:01]
That that seems to work right in shown here we have fairly recent clip from something that appeared in The Guardian newspaper that I tend to repeat I think this was last week or or at least not long ago profits have gone up tremendously as a result at least that's what Amazon things by offering makes They service instead of 2 days.

[00:06:30]
Right and they plan to roll this out hopefully to all of us right so again amas only seen this thing heavily in making believe very faster and faster and faster. Right now anybody in the logistics feels thay knows that that also means sure chances for consolidation will go down if you have less time of course they make it up by volume right so there are some interesting trade offs happening here.

[00:07:03]
So let's look at the big ethic growth in last mildly for what's happening here and clearly didn't the big dip be. Recent e-commerce so. Business to consumer we can purchase almost anything we want now online and get it delivered to our home right so that the bullet underneath about product available t. is very important here as well the growth of the sort of 2 parts it's more products you can buy online and people tend to buy more along the.

[00:07:44]
Growth of the middle class that's maybe not necessarily the case in the ways but certainly in other parts of the world China specifically with more people can afford to buy things and they tend to buy it online. The other part I mentioned that the already to smartphones right we saw what was it 150000000 people sort of actually purchased something online through their mobile phone.

[00:08:14]
He has played the big role in this we all know I saw some statistics that we look at least $25.00 times a day if it was per hour on our phone. Right. And then the other part and we'll get to that later of course it's not only am I so on that he sort of heavily involved in the online beast nice we also have really different business model and in part that was motivated by the sharing economy which we'll get to do a bit later so there's all this growth I want to highlight the few things about e-commerce per se and I'm sure most people have seen statistics like these what to me is the most interesting is the graph on the right and explicitly the blue line but the blue line indicates the fraction of retail sales that is going to be purchased online right now this is for China and China is much further ahead of us in that curve but they expect that more than half more than 60 percent is going to be purchased online right and that's an interesting statistic.

[00:09:28]
The growth in China again just a figure about packages the growth facing enormous Just look at the rate bar state keep going up almost exploding right so this is big nice of course that comes at a price right and I do think we especially as a community logistics community and everybody is aware of this but need to really think about it and also think about what we can and should be doing right there is also make it a consequence whose off the scene norm is growth right the most obvious one is traffic and congestion right especially in urban areas 2nd lane parking right there's no space for them to leave 3 vehicles to park so old they still have to make their deliveries they cannot just drive around waiting for a parking spot and we get situations like we see in the picture obviously there is a question about emissions right climate change thinks we need to be mindful of space if more and more is going to be the lead for it holds we need to start thinking about how to mitigate the phase and I'll show you some ideas of how that's going to work.

[00:10:48]
I want to mention 2 trains the one I mentioned already is fat the train that everything needs to be delivered faster faster and faster right the statistics up there are for the u.s.. And it mentions instant believe hurry and instant believe very suspiciously relevant when we go to meal believe the idea of instant delivery is that one should place an order to fulfillment essentially starts right away it's instant it doesn't mean that you have to instantly but at least the process starts so anything where let's say your service guarantee within an hour within 2 hours after placing the order is now considered under the category instant the leaf or if you were not familiar with that and I mentioned already that Amazon wants to get things to us faster and faster.

[00:11:50]
The other train that they seem Porton also when it comes to meeting geisha in the negative externalities of last mile delivery in stairs new technology in here there's really a mixture of why people are thinking about these technologies one from the company's perspective is cost and service and part of fit is also a so I say it me to geisha especially electric vehicles is a combination right that it's better for the environment.

[00:12:22]
And that's an important part of why a lot of people are thinking about. Self driving vehicles the situation needs to be different bear the whole piece that that's really cost reduction you don't need to necessarily and will see a few examples have the least worry people on board or drivers on board so that safe should cost.

[00:12:47]
I'll give you a few examples in case you're not that familiar with Droid x. Here are some pictures right this is what we're thinking about they can go on sidewalks make small believe Amazon is testing them it's one of the new technology states out there I haven't seen them in my stream create but they may be coming soon.

[00:13:12]
Self driving delivery vans I find this interesting. In this particular picture it's not necessarily completely clear how things are going to work right but if there's still a person that this is going to take the packages out of the van and the leverage game to me. Here we see that their taste in terms of grocery delivery I find this one the most intriguing and the most likely to be if faked if this kind of partial books on week ends right issue can see from the design you can walk up to this vehicle type in your coat and one of the lockers is going to open.

[00:13:57]
Right so if you think about why are companies thinking about these well there are several reasons right one is this isn't the tone of this vehicle so there's no drive for so that safe zone cost the other thing that's interesting partial boxes take up space right and space is not always available right store stake up space here you have a mobile solution where all you need to use a place where you can park for a little bit and then move on so this is kind of a very interesting concept and will get big plater in that there are people that believe that this is what we're going to see more and more off and the dominant mode to all believe Parsons right the other thing that it does is it doesn't go to your home right so it still has a consolidation opportunity right so it's kind of meeting gaping.

[00:14:59]
Possibly emissions some time and also still making it convenient we could have come to the parking lot of the I asked why are you building which currently doesn't exist and we could go and pick up our packages there which is a good mode to all transport. Electric vehicles I think here the main motivation be climate change right Electric is better than diesel or whatever and I think this is going to be bank and it's also a very good thing for society to see that companies like Amazon are going to use these more and more right and like anything if Amazon order something they're going to order a few.

[00:15:47]
Here are some other examples of the electric vehicles the one thing to note here is that the endless old one is still pretty large that's more for possibly Home believe or ease in really downtown areas you want to have smaller versions that may be able if it's legal to go on the sidewalk it's true so there's a lot of different designs that are happening right now.

[00:16:12]
Of course in inner city especially in Europe where vehicles may not always be allowed to go in to see the same period certain times by couriers are in many ways ideal right to efficient there's no emissions you can have a combination of pure bicycle or a lick treat bicycle it's a very nice option.

[00:16:36]
The last thing I want to point out is this is I think another example of war things are going right so here the idea of what we're looking at is there is kind of a partial box where you can go to pick up your package but this is going to be supplied fully in a fully automated way right so a drone can come.

[00:17:01]
And then there is technology inside that can just painful and so everything can be done without human interference other than picking up your package right so so one thing that this shows that there are a lot of creative people out there that are trying to make the last mile believe very both easier cleaner.

[00:17:25]
And more efficient there's a lot of things happening which is why this is an exciting time. I wanted to show you something that was published in a recent study by McCain c. about different making the Sims for partial belief 3 and I'll flip through them quickly so the 1st.

[00:17:46]
One that you see there is pretty decent when you have a person driving around in a vehicle stops it to front of your door and hand to the package right drones we've all seen the little videos of a missile on. The other one and we'll get to that in the main it is being right so it's not the endless on employee that's going to make to believe 3 but somehow there is an app and when you say I need to believe every somebody in the to individual like you and me is going to make to believe will talk a lot more about that in a meeting then we have to draw instead we've seen.

[00:18:28]
By couriers semi autonomous ground for vehicles and the ones that I pointed out right kind of the mobile parcel boxes which also the study believes is going to be to future write. A columnist vehicle like a parcel box that can get to it strategically chosen locations and you can go there and pick up your package that's what they believe is going to be future All right so where does the sharing and geek economy coming.

[00:19:03]
Right so let's 1st make sure that we put the concepts there so that we know what we're talking about so if we look at the sharing economy right the name suggests we're sharing and states. Among individuals mostly right that's the big idea the geek economy it's more that you are no longer an employee but you work on getting this right so you for a short period of time you perform a task and that is again and you get paid for that right so that's so and clearly if we think about a company like this is where these 2 are coming together right many of the drivers that work for.

[00:19:56]
Use their own vehicle and so in that sense they're sharing an assett and. Some of them at least are performing geeks right to fuel the work for an hour it's considered to be I mean we've seen and this is a downside to the strength that there are lots of people that actually work for 8 hours so de facto they have become an employee but they're not treated as an employee and don't get the benefits that employee Skitt and this is of course where we've seen or at least some of us in California have trying to change their right to a b 5 legislation says if you work for 8 hours for Uber should consider you to be an employee.

[00:20:43]
Right and we'll have to see what happens. With that because they're going to $58.00 right something interesting dink economy seems to be growing rapidly I graphed some statistics here that. We might may find that in the near future most of us are working to get the economy rather than as an employee he also company.

[00:21:13]
Or the state like. All right so as I mentioned when we put sharing a geek economy together we're getting kind of the like 850 s. and if we sort of put that together again with last mile the leverage then the prime example of that he's. Right another important one is groceries that share similar characteristics although the pickup is typically in grocery stores or your Wal-Mart rather than a restore on a chair similar feature right and I want to specifically look at this a little bit because we have spent quite some time with one of the players in this area so this is not a u.s. phenomenon right we see it happening everywhere I had a chance to talk with people from my to one not so long ago and all.

[00:22:14]
The Chinese version of Mulally very and if I have time I'll come back to that particular example a tiny bit later. Here are some statistics for the u.s. basically it shows that it's growing. The other part that they seem Porton to is and I don't know if you can easily read that platform to consumer the leaf which is the blue part and that case kind of the growth pop that it's a trick that we're going to focus on right of course pizza delivery we've known that for ages what's new in the way are these aggregators that are really taking over these pieces.

[00:22:59]
If we think about market sizes not surprisingly China is the biggest market for me you believe. Of course they have a huge population there is b could option all fake knowledge she. It it's it's big but the u.s. is not that far behind certainly in terms of money.

[00:23:20]
All right so here are a few players in the u.s. and I'm going to focus on grow up which is one of the large players being nice arena. So we probably all have heard to all certainly the students here in the Orleans may have given you stick one of the things I want to point out here is that they and I checked their Web site yesterday if this number was still accurate they're making 500000 believe.

[00:23:55]
And if you think about this clearly that process can no longer be manual right when actually when I started working with growth pop. 4 or 5 years ago it was filming right and it's amazing how quickly and of course they realize that that is not maintainable if they continue to grow so you need to and that's partly what I'm going to talk about so of course we know how it works we have an app we look at what we want to eat we press a button and it gets the lever where we'll.

[00:24:35]
And if you think about it of course this is one of these examples where you want them to leave or right if you are hungry and you want to eat you don't want to wait 3 hours right so in this is the prime example of an environment where the service guarantee has to be sure and has to be instant right.

[00:25:02]
So I'm going to talk a little bit more about the o.r. side of things and there are basically 3 takeaways one is stat what happens in most of these situations now it's kind of repeatedly solving an assignment problem and we'll get into a little bit of detail in a minute of the other one which I think personally is the most intriguing.

[00:25:29]
Drive for capacity management right if you think about what the big change in the sharing economy and the use of crowd cheeping that where in the past companies owned to believe every capacity like u.p.s. owns the belief capacity you have private fleets that's all change right with the sharing economy you now have tons of people that are not your employee and you want to rely on them to make believe grandpop doesn't own or own.

[00:26:09]
Employ any drivers they're all individuals that use an amp like bird to say I'm happy to drive into lever for you in the next hour right how do you manage that capacity remember that to me a lever you want to be sure that the meal is delivered let's say within 45 meters how do you ensure that you have to delivery capacity to make that right this is a completely new kind of question that we didn't see 10 years ago and a very challenging one and we'll talk more about that in the future the other thing that I have observed and I'll spend a few minutes talking about that naturally when started.

[00:26:56]
The focus was on what they call the diner the customer that orders the meat right all their focus was on how can we make that customer that diner happy right how should we build our optimization or decision technology to make diners happy right what we see now is that there is a big change happening that they realize that they also need to make their drivers happy in part because they're not their employees right and unhappy drivers may instead of accepting tasks for grandpapa may actually say I prefer to work for or eat and so this is a big change and there are lots of interesting opportunities for people that like to think about logistics transportation and optimization to see how to go about that or pricing it's huge but we'll get back to that.

[00:28:02]
There's a lot of works here but we all understand the basic principle you place an order and that gets somehow delivered to your home. The bottom part is probably the most interesting part that says that there is an awful lot of uncertainty in this whole process right how many orders are placed because it all happens on the end it's not like I place an order and say well can you tomorrow evening at 630 deliver repeat it's not how it works it's instant the lever that matters.

[00:28:43]
There is some pretty action that can happen and we'll get to that in the end point we start talking a bit more about machine learning. Similarly meal preparation time right is uncertain right. Aggregators talk with the restaurant have some agreement on sort of what they can expect but many of the restaurants have also seating and people in the restaurant and they often prefer to keep them happy so there's a lot of uncertainty even when the mealies ready right and then there is traffic and all those other things parking are you in an apartment complex there so there's a lot of uncertainty in this particular environment.

[00:29:33]
So how do these companies the ones that were aware of. Pop actually sold this problem well the core is repeated assignment and we'll get to that in a minute so so what are the main. Entities or elements in this process right on the one hand you have the orders and on the other hand to have to drive right.

[00:30:01]
And solo an order also has a Ristorante associated with it so this is where the pick up takes place and of course an order also has had to leave 3 actors and a driver even though I didn't put it here in some cases you know a little bit about how long they're working and will get to that in the mean if they are purely signing along then you may not know how long they will be there for you but.

[00:30:30]
Case many of them have a fixed period of time when they work and then what I didn't least here is you know what their current least of tasks ease that they have already been assigned to them and where they are read they become available in game right. And then there is so Ok so we need to put these to give right and if you think about optimizing nice decisions then you need to have an objective and so it's actually already quite important to think about what the objectives are right and 2 that we have been using most of our click to door time which means I pushed a button when the meal actually arriving at my door and there is also ready to door which is when was the meal ready at the restaurant and did it the rifle home right and clearly the 1st one ace my total weight.

[00:31:27]
And the 2nd one is freshens right and so both are important. But of course there are also drivers. That your cost him to cease them so you want to utilize then. You can't so somehow they're all these different objectives stand you half that you need to put together in an optimization based approach.

[00:31:55]
And I wanted to look a little bit more at these 2 basic performance matrix at least the way did grow up thinks about then right so when you place an order internally they have a target let's say within 45 minutes we want that order to be placed and if you think about this this already is kind of it's not a strange concept.

[00:32:24]
That it is already something where you could have questions where the the right way to go right because clearly if you leave next door to to restore and it's going to be easier for me to make to believe every Then if you live far away from the restore So should the target be decided but even if you think about the target right shift it would be a function off the placement time right so shift it Target be different at 3 pm versus 6 pm we know that the order volumes at 6 pm are going to be quite different right should be different based on the particular restaurant right some restaurants tend to be base year than others should we take that into account when we use our internal target right and time of day day of the week all these things already.

[00:33:20]
In a way you need to think about it but it's usually just 45. Right but there are lots of questions about that and similarly you can think about ready to do or time what in the optimization the way they think about it is if you sort of exceed that target time then somehow something that's not good to happening and you want to avoid it right so this is how the optimization deals with these kind of questions.

[00:33:54]
So what are the challenges. There's a lot of uncertainty in the system right so basically what you do is you have a mostly reactive approach right at a certain point in time you look at what you know and do the based You came right. So here is that process right here is we sketch how it works sold at some point your day starts and at some point your day ends meaning the time that you accept your last orders and then during that day every few minutes you make decisions.

[00:34:35]
Right and it grew up up this happens every meeting right but there are already interesting questions who said mean it really important Could it be 2 minutes right then we maybe came back to war and there are lots of questions and each time you sold an assignment model and then you have to compete and I'll be more specific about what that means right so we for those of you that have seen our models we have the drivers on one side and what they do is they look at all drivers that are active at that time right so for all drivers you can look at what time and what location do they finish their last assign times.

[00:35:19]
So that's the time that matters and then the other thing issue have orders and you have choices to make already day or do you look at individual orders or do you already grouplet say orders that are close to each other into a single kind of you need to which you want to assign to drive right and then the biggest biggest challenge is how do you actually put a weight or sort of a measure of how good these particular assignment that's where in a way old and all h. and proprietary information should go right every company.

[00:36:03]
Uses these kind of methods and they think they do better because so few out they choose kind of this evaluation about what are good measures to you. That it's not easy I tried to illustrate we waved a few small examples right so let's look at the particular order which is the blue circle and this is the time when it gets ready at the restaurant and note that this is already uncertain but being these optimization models it's assumed that we know these and you have 2 drivers one that's a little bit further away but is ready with this last task or completes his last task of being earlier so it can drive a longer distance and get there before the order recent ready or you have a driver that is much closer but only finishes a little bit later and therefore he can get to the restaurant only a little bit after the order is ready.

[00:37:05]
Right now which one should you prefer. Right it's not completely over use right in a way maybe to dry for that these nearby because when he drives through to restore auntie's driving empty and doing something useful would be preferrable But if you think about the customer may be speaking up the order when it's ready may be preferred Of All right so there are a lot of tradeoffs that have to be made.

[00:37:34]
I mentioned commitment strategies earlier the reason for talking about commitment is to look at all the drivers and the orders that are in the system and you make the best possible assignments that doesn't necessarily mean that those assignments have to be implemented. Right there is so to say can stay up where you look at all the assignments and say Should I implement this assignment or not and the simplest case in a way the.

[00:38:05]
If I may to an assignment to a driver that they still actively working on the tasks. And it takes 3 minutes at least for that driver to finish that task and I'm going to run my optimization again in 2 minutes maybe I don't need to commit right because I have another chance to see if more information these have a little and I can make a better assignment for that driver right so some assignments we simply ignore then there are also assignments state you say Ok.

[00:38:41]
You have to start driving now to pick up that order this is just nothing we should do otherwise it's going to be very late. You can also make decisions well start driving toward the wrist wrong. But when you get to the restaurant we may actually are there change or all come into the assignment that we get a few early right so into commitment part there are also a number of different issues that come up.

[00:39:14]
Another important aspect is bundling right I already mentioned that maybe sometimes you want to group orders were bundled orders and assign a bundle to a driver right and again if you think about face the tradeoffs are not obvious right bundling can be more effective in this say in static going believe or come back to the restaurant and then the lever again of course involves partly driving takes up more time so it might be better to pick up 2 orders at the same time then go deliver one and only for the other one right it's a more efficient use off your driver capacity on the other hand is this example shows is that if we have 3 orders that we want to put together in a bundle of course I can only pick up the bundle when the last order is rate.

[00:40:12]
So Order Number 3 determines when I can pick up these bundle but that means that order one which was already ready earlier losing freshens right and then there's also to sequence seeing of how I'm going to deliver a damn right if will top off that I'm 1st doing 3 been 2 then one then that order is also in the vehicle right so even though bundling is necessary at some times during the day how to do it is not necessarily an obvious question.

[00:40:50]
So here are another aspect that we looked at we've grown pop is this is some curves that show what happens during a day right so. Issue we expect there is a little bit of $5050.00 during the lunch hour and a big kick to 50 happens during the dinner hour right now what's interesting or the most interesting thing is the green line the green line shows how many drivers become available so finish a task.

[00:41:24]
Who are their last assigned tasks in the next 15 minutes right so you can just do a plot and create such a graph and then the other areas of course you can also plot how many orders sort of were placed at a particular time which is the blue curve at least to think about right and so if if you look at the green graph and the blue graph you see that there is some challenges happening at dinner time right the number of driver that becomes available in the next 15 minutes is relatively small compared to the number of orders so you have to start bundling if you want to have any chains of succeed so one of the kind of things we work dong is trying to design a mic trick that helps recognize when it's time to start bundling more aggressively.

[00:42:17]
Here is how bundling works we've seen this graph I mentioned this already in a way if you look at it at lunchtime on this particular day they had too many drivers they didn't need that many but at dinner time they had to shoot right and this already points to this capacity management question that I believe is so important.

[00:42:40]
You can look at some other things I mean when driver is one thing that you've seen here is that you had a large number of drivers during the lunch period you have to even larger number of drivers during the dinner period and fewer in between but the work that they do especially the one so a lot of them came on duty at 5 right but they spent a lot of time driving to the 1st restore.

[00:43:10]
Right this is the green area and it's not very effective if you're not using them if you so can you do something about that when you think about driver management. All right there's a lot more things that you can do I'm going to skip that I do want to point out a few of the practical complications you may actually lose communication with your drivers right it's one thing to say I optimize every minute and I'm going to commit and then you say well there is things drop that you can do or days past that you can do 1st of all tell me whether you are going to accept this right they're not your employees so they may actually say no.

[00:43:53]
And they say actually no very often which is frustrating of course for a company like Groupon I was surprised when I heard that that's a new order all 35 to 40 percent. Imagine how to deal with these that's challenging. And I mean of course uncertainty it's true. All right so I say at the end this is one of the things I want you to take away right from an academic perspective this is I think where most so far opportunities are to do really something innovative right how do you manage this uncertain capacity.

[00:44:40]
This is totally new we didn't have u.p.s. doesn't worry about it and muscle and worries about the little beat but certainly companies like grow up and sees a big big thing and we're currently also working together with Rhodey which is sort of a local start up that actually is doing home grocery delivery for Wal-Mart relying again on crapshoot being crowdsourcing How do you manage to make sure that you can deliver in an hour or 2 hours how do you know you have enough drivers right so how to ensure that the sufficient number of drivers is available to deliver the placed orders and to meet to service problems right the letter parties very important if you promise the livery within an hour or within 45 minutes how can you make sure that you have the right number of people.

[00:45:38]
So I split things up in 2 questions one How do you even know how many you need to write that mean it's one thing to start thinking about maybe how should die we wore thin but the 1st thing is how many do I actually need right and that's not an easy question because it depends on your dispatching strategy it depends on the order patterns it depends on.

[00:46:07]
Did. Geography the safety in which you operate these questions are not easy to Ames right in with roading and a Ph d. student here we're working on using machine learning take me to try and estimate how many drivers we need for a particular store at particular times of day.

[00:46:29]
Or for particular order patterns and then they are the question is. Even if you know how many dry for shoes should have how can you make sure you get it right and soul the latter question. We looked at the little bit also we've grown up because of course the only way that you can control them somewhat by means of compensation right and so in the case which is a reasonable approach is face a we're going to actually request that you can mean 2 working for us for a certain period off the day let's say between 5 and 7 tomorrow.

[00:47:15]
And in return for that commitment we want to we give you a guaranteed minimum pay. Right and hopefully that attracts enough people to sign up right otherwise you're completely dependent on what order so I assign to you. But here at least you have a minimum pay guarantee on top of that we may also ask that you can only per hour reject $1.00 of our tasks that we offer right so you can structure this compensation in different ways.

[00:47:55]
But what that means now is if you start planning such a system there are 2 different types of drivers in your system one the ones that we label scheduled which half accepted this minimum pay guarantee and name return tell me or or are available for me to assign But you also still have the people that just open the app and say I'm here if you have a task for me I'm going to do it right so we looked a little bit at the end developing technology to determine how many schedule drivers you would meet so what these blocks would look like.

[00:48:36]
How many schedule drivers are needed and what should their blocks look like right so we develop something Knology for that and if you think about it so that's a tactical planning problem that we solved the night before and of course during the day there is dispatching dynamic routing and you have these unscheduled drive for that also show up.

[00:49:00]
So this is quite a different difficult problem and you want to maintain some service guarantee right 95 percent of the orders are delivered at or before 10 argot that's kind of how we incorporate that. We use continuous approximation value function approximation if people are interested in more detail we can talk about that a little bit later.

[00:49:30]
We tried it to all meal delivery data from Iowa that's publicly available for different strategies and our strategy seems to work in this sayings that we can make drivers available at the right time for the right amount of time better than other strategies which say if you own cost for the company.

[00:49:57]
So as I mentioned the 3rd thing that I wanted to talk briefly about is the fact that now suddenly companies are starting to worry more and more about the drivers as opposed to only to diners. 11 situation that we encountered early on in our work we've grown pup was.

[00:50:24]
What we labeled driver drainage. Right and what that means is drivers are not stupid people right over time they get a sense of where am I going to be making the most money right should I be in Buckhead or should I be an elf arena should I be downtown should I be in for Ginia Highlands they somehow start recognizing that if they were in a particular area either dictates for higher or they got more orders per hour or the distance between the restaurant and the delivery address was shorter which is good for them so dry 1st saw us start to convert all on areas where they believe they're going to make more money.

[00:51:17]
Of course it doesn't work like that because you have too many drivers there then you can't assign but that for the company the biggest worry is now you have areas where people place orders and I don't have any drivers. Right that's the worst part. So how do you deal with such situation right so one opportunity date you can consider is actually restricting dry 1st to particular regions right so we only assign you orders from these particular areas right on the other hand we can also see that that can be limiting right if you don't need that many drivers in that particular area for half an hour an hour and a day how do you start thinking about sharing right so should there be.

[00:52:14]
Sharing of these kind of Fi Dia's. The other thing that you notice is drive 1st tamed to like operating in one in the same area they familiarise their cell suite the delivery address use where they can park etc So in especially if they leaf in that area they don't want to wander around.

[00:52:36]
All of New York City for example but rather work in only the Bronx or whatever so how do you start incorporating these ideas in your decision take knowledge is a big thing that the companies are now thinking about and as I mentioned before you've probably seen it especially in the leaf situation these drivers work for multiple platforms at the same time right so how do you create loyalty How do you make that happen.

[00:53:10]
That's not an easy thing. Another thing that we have been thinking about not really developing technology that grew up up decided to use but the other thing that this quite interesting and intriguing again to me that you control what shown on the app. Right if let's take a simple example if a customer wants to order a pizza.

[00:53:41]
I decide which pizza restaurants that person sees on the 1st page on the screen. Right now clearly probably you pick restaurants that are not 24 far away from his to leverage address because that's one what he expects or she but also probably what works best for you but if you happen to have a driver already at a particular pizza restaurant maybe that one should be at the top of the least right a pizza is a pizza at least to me so how how do you manage which of the restaurants you display now of course your contracts are with the wrist or office and so you don't have complete flexibility but you have some flexibility and they're interesting questions of how do you manage orders to minimize your logistics calls while still status finding the surface promises that you might be very interesting questions that.

[00:54:51]
Students hear that still looking for interesting topics to work on Certainly this is one of them. So I mentioned research opportunities in the space there are plenty around capacity management. Order Management and there are still lots of more things that we can do even in the drive for order assignment how can you look ahead how can you incorporate predictions about where so if the next batch of orders is going to come from.

[00:55:26]
This also indicates that there are great opportunities potentially for machine learning right all of us in the room I'm sure you have heard the bus around machine learning and that it's going to sold for all our problems once you can souls go when should Can we go you can probably soul for any problem in the world right of course it's not that simple but there are certain types of things that could be very helpful here right predicting future orders is not easy and a standard forecasting take me may not be appropriate right because what you really need to know or maybe want to know need piece of big big statement but want to know.

[00:56:14]
Delivery location specific right so there's not that much samples from a particular location that allow forecasting to work with great accuracy and are lots of things that influenced these the time of day the weather off today I mean I mean traffic conditions so it's possible that machine learning can do better in the traditional forecasting techniques.

[00:56:42]
Order acceptance is a very interesting one we try to do some basic regression modeling with grandpop to see what influences that decision right we thought maybe if a driver is safety and he saw her shift he wants to make sure that he aims. Near his home is that important it's the team in if we had 2 whole range of features it turned out that the best predictor was if he had actually often rejected orders before.

[00:57:15]
Still useful information maybe but that was a very basic simple model maybe there's more that we can do so I think there are certainly opportunities for machine learning what is the right time to start trying somehow to get more drive for seem to see still right as we said we in our decision take knowledge she developed at least one metric that we use to be more aggressive in bundling but maybe something similar is needed to try and attract more drivers right search pricing or whatever of these right and in machine learning might be helping you to make decisions like that.

[00:57:59]
All right I wanted to in. I mentioned that I had a chance to talk with people from Matewan when I was in Beijing in December. And they're one of the biggest believe every platforms that exist and also remember that time mentioned that. Grandpop now delivers $500000.00 orders per day issue can a mansion in China things are always a bit larger.

[00:58:32]
Right so here are the numbers that you half so they actually half $27000000.00 orders per day that they need to deal with. Thousands hundreds of thousands per means that come into their system right and typically they have $600000.00 drivers working for them on a particular day. Right now I mentioned when I started talking about amas along that the numbers be right and the time to make decisions get smaller and smaller.

[00:59:11]
This is another great opportunity for our community the optimization logistics community to start really thinking differently about how we might think about even addressing these problems when you get these numbers. You need to work heart. So they have what they call themselves the May 20th super brain and a late to see what kind of techniques they use parallel computing that's not a big surprise machine learning data mining prediction lots and lots of things the biggest thing I think that is helping d.m. how day so late for each restraint.

[00:59:56]
The acceptable belief for a location the right one way to scale down your problem is to say well I only believe for 2 person that leaves right next to the wrist of course you don't make money doing that but that makes the least the logistics problem quite easy.

[01:00:14]
Right and so this is a big thing thinking about the surface area for particular restaurant rank and even that is not the same problem to answer right and that again depends on your dispatching to believe every capacity and it you have is really for normally interesting problems but scale to me is certainly an interesting one not only in we'll believe 3 we see that also just problems we can yes but more and more packages are going to be delivered the problems get bigger and bigger we have lists and lists time to make decisions how do we deal with that and that requires some creative new thinking.

[01:00:58]
All right and that's it for today I am happy to answer questions if you have 10. Yes. No. I think this is a I mean a valid comment and a valid point you're making. We work primarily. On the transportation saw it we are transportation experts. And but I think it's a very legitimate question to say hey.

[01:01:49]
This Ready time right place impacting the time you have to make to believe every So focusing some of your attention and effort into streamlining maybe Neil production would be a very valid to and useful thing to do it is something that was never mentioned by a group pup is something that they want to help with but I fully agree with you that that's part of off the answer possibly to the big picture question yes you're absolutely right.

[01:02:25]
Yes. This is. Where. It's. Happening. It just means. That it almost seems like you're making a lot of money in this. I'm. So I think this is Main a morgue I mean. General question that somehow society has to think about right to value all these companies is really determined by the stock market not by their actual beasties or even Amazon is making money they're only making money on their computer server beasties in the package delivery they're losing money.

[01:03:23]
And so it's an interesting question I mean part of it comes back to what I said very early on we discussed of mergers are not willing to pay for the service right and in a way if you think about it. What we're asking for and again this is partly the fault of the companies themselves what they are offering to us instant believe free.

[01:03:52]
Delicious time you have to make the lever east the more costly it is for you in the same poll logistics. Soul. I think at the moment so one way of course that you can try to do this is by exploiting being where you don't pay for been a feat and it's nice to call space the smaller but even then it's probably not small enough to recover the calls.

[01:04:22]
I don't know it's something that. People have to think about I mean certainly a ton of this believe every might be getting there which is why I think at least the package delivery companies are looking at these kind of mobile partial boxes and the future. And I think we need to again try and think creatively if we can actually make it profitable.

[01:04:53]
Yes. There. Are. Areas. That are. Right. For. One. Thing. More likely because also in it so so I mentioned that the predictors that we found to be the most helpful are amounts and I mean just earlier previous behavior I think another I do think here and I'm making a broad straight statement here is that the other thing that really shown here the stat we need to integrate our optimization more human behavior modeling right that is not something that we have done in the past and I don't necessarily think that machine learning.

[01:06:03]
Can figure it all out by itself so talking with psychologists and thinking about if it's going to be quite important. That they're half I mean the other thing which of course is hard to understand just from let's say grow up update if a person doesn't accept an order doesn't mean that it was somehow an Ok order for the person but if her team something better at the same time he may just have chosen the.

[01:06:35]
Right and. The moment as far as I know they're not tracking or cannot track that information. So this makes it very hard to understand exactly what is happening the order to pass that we offer or are there other things that play the one thing that I do know and you seem to remember is well this is why day start creating Compensation Scheme dad discourage you from not accepting orders right I think now has something where you get kind of a bonus if you do a certain number of troops during the month or during a week right and the main reason for doing that is to say hey if we offer you something.

[01:07:25]
If you want to reach the bonus you might as well better accept it right and it's very hard for any individual to see let's say at the beginning of the month what should I do right they have faced target now you have to start thinking once they reach that target do you if you say that wrong do you need to start putting in extra compensation schemes their looks of questions around this particular spade and some of them are not easy to answer because only now people start trying to really investigate the impact of all people working force.

[01:08:03]
Several of these platforms at the same time. Yes. Yes. And I mean yes. So this is what you have seen a little between the pool right pull didn't work very well at the start. And again if you think about 8 The reasons were fairly obvious right if people just say I want to go from a to b. now and they expect a vehicle to show up within 34 minutes the chance that you can also rerouted to find somebody exactly on the route to where you go or small right so the people at the pool have started to think about different things right where you announced earlier to give them more time to find consolidation options.

[01:09:33]
So that's an example of where exactly this is happening you need to somehow incentivise right if you were Poulos a good example they said well it's going to be cheaper for you and they didn't get anything in return and then couldn't make it work right now they say it's going to be cheaper for you but you have to tell us a bit earlier and we will seem to him a situation a person is going to be at your door it's a trick so clearly working on trying to get information earlier is very important and we've done studies also for grew up pup to see what the impact would be if they knew 15 minutes earlier and there is a big impact how to make it happen it's less clear you.

[01:10:23]
All right thank you very much I hope you enjoyed it.