Also Well thank you I'm
genuinely excited to be here.
Appreciates you spending your lunch
time with me and thank you Ben wall for
inviting me today so
my idea today to talk about is I want
to spend the beginning of my lecture to
really motivate why I think we need to
rethink supply chain design and then I
have one example of how I think we can do
that there are many of them in the many of
them I spent this morning talking with Ben
walk about things he's doing that are very
much in line with that and then ultimately
you know really thinking this is.
The beginning of hopefully a long journey
and so there's definitely lots of future
research very specific to what I'm talking
about but also I'm hoping that there's
some discussion about what
are ideas to build on this.
To get started I want to ask
you a bunch of questions so
how many of you have purchased something
online in the last week is anyone
all right what types of
stuff have you purchased
you know anyone willing to say what
they've purchased shoes so OK.
First.
What you have your dog also OK Any other
bizarre things that
anyone's willing to say.
A dinosaur OK so
I would say that's quite a wide assortment
from a dinosaur to close right.
How long were you willing
to wait to get these items
are you willing to wait a week.
A little bit a couple days right any
one time am like one hour delivery.
Yeah OK And you know where did
you get that delivered to you
likely your house probably right and
where you're ordering a full.
Palette of those things or did you just Or
did you order a polyp of dog bones or
just like a couple right.
And how much were you were you willing to
pay for the use where you are willing to
pay a lot for shipping so the survey say
no I mean that there's not much there so
you really contributed to what I would say
is a fundamental change in how we need to
think about supply chain so Today's
customers expect a wide variety of request
to be fulfilled extremely quickly
with very little warning you didn't
forewarn what we need to do in small
units so at the peace level and
some of the pilot level and some many
dispersed locations so we're going to your
houses and sort of willing to come to your
store and at extremely low cost OK So
this is today's customer expectations
which I will say is fundamentally
different than yesterday so
yesterday we were willing to have fixed or
locations and the demand would
aggregate at those locations and so
my hypothesis is today supply chains
are optimized for yesterday's customers.
And so when you think about what
worked really well yesterday we have
very fixed networks and so we thought
about let's optimize where should we
locate those different locations and then
once we have those locations how should we
optimize the resources that we own and a
lot of those resources are big things and
we could even think innovatively
about that in terms of
you know Wal-Mart got famous about doing
cross stocking and things like that but
it was all built around the fact that
I knew where my demand was happening
it was very fixed and we had
aggregate demand at those locations.
And that's what worked well in the past
my hypothesis is that's not going to work
well in the future and
you don't have to take my word for
it these are all companies that
have either filed for bankruptcy or
have closed major numbers of
stores in the last six months.
And this is by no means
an exhaustive list and
I think there's a really interesting
survey that came out recently that found
that only ten percent of global So this is
not just a U.S. problem global Burke and
mortar retailers are actually profitably
fulfilling e-commerce orders so
to dissect that means if I'm
you know when I'm Macy's or
a gap that has physical stores and
I also have another channel
the online channel if they're using
the same supply chain network and
using the same decision making and
kind of optimization approaches.
They're not doing well so
ninety per cent of them are actually
not affordably making money.
So that brings me to the idea that we
really need to rethink supply chain and
was just the design and I think there's
you know two ways to really big categories
we can rethink supply chain design
we can rethink supply resources so
what do we mean by and network what do we
mean by resources to fulfill the customer
requests and we can also rethink how these
demand requests are going to occur I am
of the belief that our now now now and we
want this fast is not going to go away but
we can be innovative about
how we deliver that and so
today I'm going to talk about
rethinking supply resources but
I think there's other things you can
do in kind of both of these spaces.
The best way I think to think
about this is through picture.
Here's a picture of Seattle downtown and
you think
of the extra capacity on that road do
you think you could move more people or
goods on that roadway any
ideas any thoughts that full.
Any creative ideas of how I could move
more people or goods through the road.
Sidewalk So yeah we could use
the sidewalk sometimes when I
give this you can save maybe we have a
bike courier in there any other thoughts.
Put them in buses.
Vertical right you could have some
growing something going on right.
Crowd shipping.
So this is the same picture
only without the cars OK so
it's very clear there's all kinds of
capacity there's all kinds of extra
capacity in this picture however we can't
think of capacity of the traditional
way we thought of capacity in the past
we can't think about adding more cars
instead we need to think about how can we
tap into inside of that car those smaller
granularity how do we tap into
all of this extra capacity and so
I call that underutilized capacity
these people are going there anyways
Can we somehow tap into that the other
thing about capacity is we need to not
think about it in terms of ownership so we
need to think about if we really want to
tap into the if it's not that I owned
that person's trunk you know so
we need to operationalize a little bit
more when we think about capacity so
this picture is coming at the heart of my
motivation for a lot of my research is I'm
interested in tapping into these
underutilized resources on demand and
I feel that is a potential way to get
away from this static fixed network and
have a much more dynamic
a laugh network OK And
there's a ton of analogies to supply
chains to my car picture one example
is a warehouse or distribution center
these by definition are extremely static
resources so once you build that it's
really hard to kind of make it go up or
down however this is commonly
what demand looks like or
inventory looks like and so
if you're building a static fixed resource
to fulfill something that looks like this
there's a mismatch so there's either
underutilized resources so if I build that
the blue I have underutilized resources or
if I could think about somehow
making my resources less fixed and
static that would be good.
And I actually have a project going on
right now that I'm not going to talk about
in terms of detail today looking at this
from a facility location perspective and
looking at how can we build a dynamic
elastic Resource Network flex
is just an example of this there really
on Demand platform kind of Air B.N. B.
for warehousing and our interests here is
in more the facility location problem and
how do you get access to scale
how do you compete with Amazon
without owning this number of
distribution centers that Amazon does.
But today I'm going to talk kind of more
generally about a different type of on
demand platform So Flex is an example
of that the two most common on demand
platforms that you've probably
heard of are an Air B.N. B.
So the idea here is that there
is a central mechanism or
a platform that owns no resources so
whoever owns no cars Air B.N. B.
owns no facilities hotels
instead their value is in
first of all tapping into
underutilized resources and
matching that with demand on demand so
in other words you click a button
if you want to say I want Hoover and
it finds that match and so their value is.
Really in that mapping of supply and
demand and
they're interested in kind of
a systematic perspective so
they want the maximum number of people to
participate in this platform and they also
want people to participate continuously so
I want to use the server service over and
over again so who actually owns the
resources there what I'm calling agents so
they are you know you're
if you're using Air B.N. B.
You're the owner of that house
that you're renting out and so
because nobody owns me
the platform doesn't own me
my hypothesis is that we need to make
sure we're providing discretion and
autonomy to these agents and these agents
really have an individual perspective you
know they own the resource and so
I'm going to provide access to that but
I want to I own that so
I have a preference for that and
so there's lots of different
players in the space a few come in
the supply chain around the bottom but
I would say what I'm really interested
in is how can we design these systems and
how can we operate these systems in such
a way that we tap into existing
underlies capacity on demand so
I'm going back to my car picture I'm
interested in underutilized resources and
how to tap into them so I could
make the case that you know over is
a different type of capacity but I don't
know if they're necessarily tapping into
underutilized capacity they're creating
different types of taxi drivers but
it's not that they were going there
anyway so I would say that's a little
bit different of what I'm interested in
is if I'm going from point A to Point B.
and I have extra capacity can I tap
into that any questions comments.
You want to.
Just.
Share Yes So I'm not saying I would agree.
Potentially I'm not preventing us from
creating extra but I want to have this
in mind when I'm designing and some of my
algorithms that I'll talk about this is at
the heart because I am actually very for
providing someone a ride if you're going
where I'm going but I don't necessarily
want to be an overdrive or in my spare
time you know that kind of thing so
that's where I would say my focus is yes
probably you will attract people that are
participating in the platform beyond that.
Any other questions or
comments All right so let me talk
about how this would work operationally so
please meet millennial Milli OK And
so she gets off of work and
she gets a text message and
ask her would you like to deliver
groceries to others in the community and
you know she's a good millennial she wants
to make a difference so she clicks us OK
and then at that point there
are two choices that appear OK so
who would you like to deliver to
Jane who lives near Milly's home or
would you like to deliver
to Will who lives downtown.
And she actually has plans to go
downtown to meet a friend and so
she's going to feel like we'll she's
going to Ben get notifications on
her smartphone about where to go
pick up fees if it's a nonprofit and
then make the deliveries or this could be
a commercial setting where you're getting
groceries at your local grocery store and
I'm a member of kind of this rewards card
and so when I check out I say scan this
and they could see You also have these
options would you be willing to do the
last mile delivery in a kind of ad hoc or
crowd sourced way so this could work for
either kind of a nonprofit setting or
a commercial setting but I think dissects
kind of what's happening here OK So
first of all we're able to tap and
civilians under utilized capacity as
needed results in resource alas to city so
what I mean by that is if I didn't need
someone to make that delivery one tap into
her and this allows my
resources to be less static and
more dynamic both from a spatial
perspective as well as a temporal or
time perspective so that's great but
to do that I really need to entice
millet to do this so I am not
controlling her she could say no she could
have selected neither Jane nor will.
And really the platform so
who's sending these messages they
don't have control over her so
they can't force her to do this she's
not an employee of the platform.
And they also don't have perfect
knowledge of what she's doing so
most of the night she probably would
go home and therefore if they only
recommended Will she would have said no
when she won the participated OK And so
I provide choice for the platform provides
choices they're definitely estimated based
on some past history but they don't
have perfect knowledge about that.
Choices so that's good but
choice is also have consequences so
what happens to Jane you know if they need
their groceries to who's going to cover
that that would be a reject and because
I'm providing choice to someone that
participates on this platform that means
that they could say no and so to hedge
against the people saying no you might
want to recommend multiple people the same
request and so what happens if both people
pick this I only need to make one delivery
there are consequences there are others
but these are the two kind of major ones.
And then if you dissect this if you think
about it from kind of a decision making
perspective there are actually a lot of
decisions require to make this operation
happen what choice is why did I recommend
Jane and well how many of them should I've
recommended more of them what if
I compensate What if the platform
meaning you know if it's a grocery store
they likely have some of their own
delivery Caprica paper cap of capabilities
so how do I use that and tap into that but
there's tradeoffs if I wait too long then
I'm really have short time windows and
you know how does that all work and
many many more so
this is really at the heart of my research
is to try to think about I want to
tap into underutilized resources on
demand I want to provide choices
to multiple agents which are kind of
thought of as ad hoc suppliers but I also
need to balance that I need to make sure
these demand commitments get made OK So
that's the central challenge of a lot
of my research yeah one very very well.
For us.
Right there.
Were three years.
You know through.
You know only.
One.
So we're.
Very.
You know or.
Most.
Part here we're.
Learning here so.
Yes there's going.
To experiences very.
Very very past experience but I would say
the difference here is technology and
the reach of visibility so
sharing has started from very beginning of
humanity we shared but
we only shared with people I knew right so
I shared you know I know us all share
my resource you know this kind of thing
now all of a sudden you have much
more reach into who wants a ride and
potentially it could be a little bit
better tied together both spatial and
temporal I.E. potentially but
that wasn't really true.
Right here.
For years.
Already so.
I.
Look.
You know.
So.
Also I think has also helped this so
I can I think in the past it didn't work
now when I first started talking about
this I had to explain what it was some
people like I wouldn't get in a car
with somebody else now that's just
second nature so I also think
if we really try this experiment we're not
at the same point as we were initially.
But you're right there's lots
of potentially problems.
You.
Could.
Yeah.
OK.
So.
Let's.
See.
This.
So I would say first of all it's much
more complex than just the delivery but
I do think fundamentally demand
has changed so fundamentally we're
not going to physical stores anymore we're
getting stuff delivered and I think what's
telling is the statistic that they did
a survey of a bunch of retailers and
you take a given retailer and it's not
saying they're not making a profit but
they're not making a profit on
their e-commerce side so that J.D.
a survey they surveyed a bunch of
different retailers and some of them could
be doing really well financially but
they are actually losing money currently
on their e-commerce operations and
so they might still be around for
a really long time but I would say they
need to think differently about how they
fulfill e-commerce demand then how they
fill store demand does that answer.
Exactly.
Yes.
Yeah it's a great question.
All right so that was an example of just
like one person let's kind of you know
bring this hour a little bit more
abstractly and if you have multiple agents
examples so here I have you know
three smilies different agents and
we also have some preference we know some
preference about these alternatives and
just to make it easy we have a capacity
of one so each agent can fulfill one and
only one request and each alternative
only needs one agent to fulfill it
just to make life simple so the
traditional kind of you have agents and
alternatives you could think
OK let's use the hungry and
method just very centralized optimization
I care about how do these alternatives
gets matched with the agency and we
ignore anything about the preferences but
we basically have a dictatorship so
if we did that we could say OK based on
different utilities of the system
you can say I'm going to sign this.
This way but here one thing I want to
know at the end is the no choice so
at some point they have a threshold
maybe they're only willing to drive so
far out of the way if you
go to my driver rider or
driver situation and so there is this
no choice where if you offer me a or B.
I'm just not going to participate and so
be centralized systems are really not
applicable in this kind of ad hoc purpose
because you can end up I recommended
three to be and they didn't select.
You can say OK well we really care about.
The agent so I'm going to provide
every alternative to every agent so
that's a very decentralized approach so
each one can see a B.
and C.
agent two can see a B.
and C.
and so
forth the problem there is we
have my office decision making so
the price of anarchy takes happening and
so here you have two and
three both warning this one request and
then B.'s nobody picks.
Yes.
There is a difference in the sense.
This is a very important part of which I
explore is what is the value of how much
information they have but the one thing
here is I should probably add the five as
they have utilities here so thinking about
how traditionally didn't they just cared
about the wait time of the rider so that
maybe a different utility than the agents
which maybe you want to end up close to
their destination so what I'm using for
the centralized system even if you
assume they know everything about what's
happening then you could do this and if
you know exactly even though no choice you
can assign them things such that they
don't pick no choice so there is some
aspects of information and I will actually
do you tell more of that as I go along.
So one of the issues with
said Fly systems is if they
truly are kind of the very simplified
version is all I care about is my platform
totally ignoring everything else that's
probably not a good idea because you
have to use these people to participate
but will get back to yes information about
how they can estimate what the agents
are doing becomes really crucial.
So this is obviously a very
simplified example but so
what I'm proposing is
a much more integrated or
hierarchical approach which tries
to combine these two benefits so
the good thing about centralize is that
able to take a systematic perspective so
that you don't get my office behavior not
two people are picking the same one or
you're trying to kind of prevent
that on the flip side we like
to centralize in the sense that you take
into account that these have utilities and
they have agency and
we want to provide them so
the idea here is what if we could
personalize my recommendation So
number one get sure all of them but
you only get showed a three just showed C.
for example and in that case you know in
this very simple example you know then.
Each one of them are getting picked on and
so
they're all better than the no choice
their might not be their first choice but
they are willing to
participate in my platform.
Similar except stable Matt
marriage cares about do I really
care about my first choice and of somebody
else's so I would say we care more about
are you willing to participate it
might not be the best utility for
the agent it's similar to stable marriage
but not exactly so stable marriage and
will show I have actually compare results
to that in some ways prioritize the agents
a little too much and you could
reduce your systematic performance.
But there are similarities to
the actually we'll compare our
two stable matching
problem other questions.
OK So that is a ton of research questions
very much joint work with my Ph D.
student CD So when I think about how
do I design the system that I really
want to tap into ad hoc suppliers if we
think about well what alternatives should
a platform recommend to multiple agencies
given this finite resource capacity
how can we develop a new hierarchical
method to balance the need for
demand commitments and
choice what is the performance
of this compared to these kind of
more traditional approaches and
then one thing that's really at the heart
of what I think is exciting and
kind of as a proof of concept is
understanding choice in these systems.
And so many new sizer how many choices
I should give to someone I think is
an interesting kind of central driving
a lot of the development of these models
so the current state of the knowledge
in terms of resource sharing platforms
there is a definitely growing
set of research in this area so
it started with more
descriptive economy but
there are definitely prescriptive models
thinking about how can we develop resource
sharing systems more specifically in
terms of dynamic ride sharing crowdsource
delivery there is actually a growing
growing population of literature with some
of the authors in the room and if you
think about these approaches you can
put them into kind of broad
categories centralized approaches and
I don't mean just the hunger in method
more complicated than that there are kind
of two ways to think about it there are
some papers that say I want my suppliers
to have preferences but you have to bid
day in advance so it's not on demand and
then they take into account these
preferences in these bids but
there's a lag or
there's some latency and then
there are ones that taking a centralized
optimization to match supply and
demand and they are motivated I would say
by similar ideas that I'm motivated by but
they use constraints to enforce that so
you may say I'm only going to have a max
Tor distance so I add a constraint for my
optimization model or a number of stops or
you want to enforce that there's this
is a stable match when it's made and
I'm There's also a paper that fits
really well with what I'm doing
that's getting older from Powell that had
nothing to do with resource sharing but
they're interested in
accepting rejecting it and
you know doing some this patching that
actually has a lot of similarities
the difference here is they have a single
dispatcher they don't care as much
about the multiple editions and
there's also decentralized approaches this
is just one of them there's actually many
literature there so
the gap in terms of this specifically on
demand distribution platforms is no
one's really looked at the hierarchical.
Roaches to this process so looking at that
there is a central decision maker that
makes maybe recommendations but there's
also a decision being made by these
suppliers and so that's where our research
is contributing to this body of literature
and you can say well there's been a lot of
literature in by Level Optimization words
by far not the first one to do that
sort of an optimization from traditional
network design so you design a road and
then people decide how to travel on that
road those are all by level
higher fickle literature but
existing work typically fits into
two of these things you have either
you know the probably most common is this
one you decided you were assortment so
think about a traditional store what is
put on the shelves you decide at once and
then individual shoppers go but
they all have the same assortment
similar to a road network you design a
road once and then yes there's individuals
that make decisions but you don't have a
personalized road for each person right so
a lot of the Buy level
work is actually one.
Aggregate decision followed by everyone
sharing that there's also work that looks
at OK I'm going to give you
an independent decision but
there's no interaction between them.
Versus the stuff that we're looking
at is that we're interested in making
personalized recommendations to
multiple agents but it matters not just
the decisions that I send from my platform
it also matters if you select the same one
as this person selects the same one
there's interaction OK What we really care
about is we want to prevent rejections We
basically want to make sure that we serve
these requests and so it is OK if you
say no as long as he says yes right and
so there is interaction between
this agent's selections of
what are interdependent that influences
the performance of our system.
One.
So.
So.
There's.
Yes.
Yes.
Or only yes so the stuff I'm presenting
today I'm assuming a snapshot but
that definitely future research is
that OK if you have a duplicate
I just right now have a tie breaker
that says OK which one gets this but
you can have this in
a very dynamic setting and
that you would say that then gets
starts back as a request and
that is the future work of this
definitely So right now I assume there is
one round and we deal with the
consequences of that round based on some.
But future research is to make that point.
And I won't participate in the future
exactly so and I think that part to me is
really interesting is how do you make
sure that this is a viable system and
in some ways I'm trying to encourage
participation as much as I mean thing and
the participation is decided if I want to
participate and no one's there that's not
good and if I have people there and
no one wants to participate but also that.
So this is really what I'm interested in
is trying to understand how can I engage
participation through this idea of choice
so to give you a framework on the X.
axis is the number of choices so in my
example I gave two choices I gave well and
Jane you could think about more choices
and so what I think would be interesting
is this black line OK So initially this
is just kind of hypothesized if I only
give you one choice and the platform is
not exactly sure what I'm going to do so
they have some information about you but
there's error in that estimation if I only
give you one choice is a high chance
that's a choice I don't want and I reject.
So if I give you another choice
there's a higher chance that I
would actually accept something
however that only works up to a point
because you lose performance and the Y.
axis here is the platform's objective or
the system's objective because I'm willing
to maybe give you one that you want but
the platform really wants you to
take another one that you don't take
also at some point you get this aspect
of you're offering them to everybody and
then you go you have a lot of duplicate
and you have myopic decision making so
I'm really interested in this like black
curve that's kind of a proof of concept
is first of all is there kind of
an optimal number of choices what
influences those Does this
increase participation and
then comparing it to a kind of
centralized essentially solution.
And seeing what what happens does that
make sense you know what's going on.
So you know kind of when resources
are finite for this example and
we'll see this with some of our
preliminary results is the centralized
solution gives you only one and
then if it's not the one you want then
you reject it some of them will pick the
decent allies one gives you all ten and
everyone sees all ten and therefore
it can be myopic So can we do better
than those two systems dictatorship here
is an upper bound in the sense that might
not be achievable in the sense that I just
care about what's the best for the system
assuming I can force you to actually do
that which is why it's an upper bound for.
This.
Group.
So I.
Don't.
Want to yes so inside of here I'm
also caring about which choices so
I care about which ones
as well as how many and
so if you count which ones you
know count that's how many so
I am definitely interested in both and our
method actually optimizes which ones but
kind of a more fundamental
research question and
I think this is interesting but
yes I'm optimizing which ones as well.
So our very kind of preliminary model is
a bi level optimization framework it's
also a meter follower game so
the leader makes X.
I.J. decisions which are binary that says
I'm going to recommend you this option.
And there are different constraints
here and we can add more but
the one thing I want to emphasize
is in the objective function these.
Are rejections and
duplicates at this moment so
there the performance is influenced
by what a bunch of people select and
the lower level so you can think about it
network wise the X.'s are kind of showing
what am I showing to the agents and then
the agents responds on the second level
they are only able to select something
that was actually recommended and
we assume they maximize
the utility of their selection but
their utility U.I. J could be different
than the systems utility see I.J.
which I'll explain in a few other slides.
This is definitely a first step
model a major assumptions with
all of these being planned to be relaxed
if we are taking a snapshot view so
we assume we have a set of available
requests and a set of available agents we
are only caring about one sided autonomy
right now so requests will take any
deliveries from any supplier so suppliers
are the only ones that have autonomy and
here we are assuming these U.I.
J.s the platform can fully understand and
estimate which is a major assumption
that we relax in the future slides
so one thing is OK typical
this is a discrete
buy level optimization program known to
be extremely computationally expensive.
And so
we graph it is a short level as it is so
what we do is we transform this
by Level Optimization model into
a single level optimization model
using logical statements so if we know
the preference profile of an agent so this
agent prefers one over to over three and
so forth then we can write
statements that say if I recommend
you one then you would select one
because it's the one you want the most
where assuming they're picking the best
one if I don't recommend you one but
I recommend you to you with flex
to if I recommend you to and
three you would pick two you know you
could think about there's all these
combinations there is exponentially many
of them but you can write them all out.
If then statements are obviously
not linear constraints but
you can introduce new decision variable T.
i j which is a binary decision variable
and you can see there's coefficients
will either be plus one negative one or
zero and so you can take this for
every supplier you can then create
the constraints and so about
allows us to instead of having a buy level
optimization approach we have a single
level and we've taken out the objective
function and then built in the constraint.
OK And so the performance of this red
line is using take eighty conditions
the purple line is what I just showed you
and then so it's definitely an improvement
we were able to solve much bigger problem
sizes much more quickly however you know
in an on demand system we need to be fast
and so because of time I won't explain but
the Green Line is a stick approach
that we're still working on but
we're getting some promising results
of a much faster recommendation and
comparing that the main idea is can we
somehow project the to you I js and
C.I. J's the alignment between the system
and the user and can we use that
to solve a problem there so
we're giving some promising results there.
But due to time or Mathematica.
So we now basically have a model that
we're able to use to kind of get at some
insights and to me about what is exciting
about research as I find that the model is
useful to get insights into the real
system the other things we employ
is we input C.I. J's and you are just C.I.
J's are the platforms benefit of
making this match and you I.J. is that
the agent's benefit of making that match.
This is definitely ongoing work but we can
think about a data model that you would
have information because people are
participating in your platform you would
know the current location you want to have
some predicted destinations of where this
person is going they're going home they're
going downtown those types of things
they're definitely predicted in some
probabilities you could have from past
history rating and some other character
sets for each arriving request so
you would know their origin their
destinations the time window and
other characteristic information and you
know you can see something like this and
you can start estimating using the data.
And you edges the main thing out to
point out is the stuff in yellow
are things that you.
We're going to estimate using the data
from the platform but there are things
some error term you're not going to be
sure exactly what these people are doing
and so we are currently working on
their skull for your travel data
sets available and trying to use them to
kind of start estimating some of this
from real data that's available but from
our perspective today think about this
yellow is some number that the platforms
able to estimate plus some error term.
And so I actually made it even simpler so
we actually ran an optimization
simulation set up using our data
are using our model and the idea here
is we randomly generated zero D.
pairs and then what we care about
the platform is we're just simplified we
care about how long does this rider
have to wait so we care about
what time where the driver is now how
long does it take to get to them so
that's what we're going to use as
the platforms benefit of making a match
the drivers benefit is how much do they
need to detour after I drop them off so
if this person is going there but
they really need to end up here
that's the minimum of that distance.
But the one thing to realize is
the platform has a really good
understanding of what the rider is doing
because they put in their origin and
destination but the driver Remember we're
trying to entice them to participate I
don't tell them exactly my plans for the
evening so there is some error term and
so here we thought they
were going to go here but
their destination is actually here so
using kind of a random utility
model framework the red part is
the actual utility of that person
the blue part is what the platform
estimated user was going to do.
The other thing that's important
is remember we're trying to
entice participation so
the way we put this in there is we said
there is some threshold that I'm willing
to go out of the way the platform can as.
To make up but there's also an error and
if the distance there is larger than that
threshold I'm not going to participate so
if you recommend me something and
it's longer than that I'm going to
leave the system without words.
So to do that we ran this
optimization simulation
set up I know there's a lot of information
here but we basically generate origin and
destination and we have some other input
values and then the optimization engine
we test for different types so
we test our by Level Optimization
formula I just showed you we test
a hunger in method that only use a C.I.
J So ignores that there's agents and
they have utilities.
We have a decentralized which really is
an optimization we just show everything
to everyone and then we also do
many to many stable matching and so
we do all cave to cave choices so of one
is one to one stable matching all the way
here we're doing up to ten so that's
what's going on in the green part so
out of the green part we run four
different ways of doing this and
out of that we basically
get a recommendation set so
we get a set of things we recommend to
people and then what we do is we actually
simulate what are people actually going to
pick and we're using maximize utility so
the part that simulation is really simple
but we're just generating these errors and
then saying What are they actually
picking not what to the system
think they were going to pick and
then get the actual performance out of.
That makes sense.
Right so
I won't go through all the details but
we do a full factorial analysis
I think about different
correlation potentially with riders and
drivers destinations and origin.
Different penalties and
different error rates.
And here are some preliminary results
probably like anyone my grad students gave
me because last night I heard them so
they're definitely preliminary and
I actually have a legend is incorrect.
What is let's walk through some of
the lines that are solid are the ones
that are coming out of the optimization
and we haven't done the simulation so
you could think of those as if I had
perfect knowledge about exactly what
everyone did those are the solid lines and
the red is the centralized system the blue
is the stable matching and the reason it
goes off like this is because that K.
The K.
we assume you can make a choice miss but
then in the simulation we enforce
that you only can make one so
that's why we're up the green
line is our model and
yes of the green line is our model we
are maximizing So we want to big number.
And then the dotted lines are what
comes out of the simulation so
we're actually simulating choice so
one thing the thing here is the red
line centralize does really well if we
know exactly what what the person is going
to do that would make sense to do
that the minute the platform is
not able to perfectly estimate what these
agents are doing then all of a sudden
which is the dotted lines all of
a sudden our model does better so
if you look at this one this has low error
this is the same instance with high error.
So the number of choices that's best for
ours is a one over here so the green
is the biggest here it's around six
the green where is the biggest line here.
If you measure the best here the green
versus table matching this one is
pretty close the green is
a little bit better but
not by much so in this instance
stable matching approach would do
really well this one the green is
way better than either of those two.
The other things that we
thought were interesting so
there are around three thousand
different instances and so
what we did is out of our integrated
approach when does it make sense
to provide choice and so a large portion
of them you know there are you does
it make sense to provide choice
even in our integrated model.
And the large portion of them say OK
give a decentralized approach but
if you sum up everything between two and
nine that's forty three percent
of the instances so this is kind of
a proof of concept to me to say that
there is some value in trying to
understand choice in these systems and
they could potentially improve
the platform's performance.
And then we ran a number of a no
vote here of the dependent variable
is the platforms objective function so
we're trying to understand what influences
the platforms objective and
all factors are significant but
what's interesting is the things that
really matter so these are ways to think
about if I'm going to design a platform
system what do I need to think about so
what we found is that the threshold
of no choice has a huge influence and
that should make sense if you're kind of
willing to do anything if I show you.
That alternative that's not the best for
you but it's the best for
the system I'll take it anyways but
if you're really picky then I really need
to be careful about how I'm
making these recommendations and
how many estimation errors definitely play
a big role and then if you really care
about making sure you have participation
so rejection is what you care about
then that's also the most important if you
don't care about rejection and you just
have plenty of people participating
then just do basically dictatorship.
And I think just one or two more.
Of this is looking at
the performance gap so
here's the integrated model objective
function versus table matching and
the integrated versus
the centralized approach.
Some preliminary insights is that this
performance gap is also influenced by
all of our factors but
what I think is really interesting is what
influences those so the stable matching
if you're going to do that you better be
really good at estimating utilities and
that should make sense if I'm going
to recommend something to you that
you feel as stable as the you know you're
not willing to change you better make
sure you know what those are versus
a centralized approach what is interesting
is your detour threshold is really
the most important so if you're willing to
participate versus not if you're willing
to participate you have a huge structural.
Then just use a centralized approach but
if that is more sensitive then it makes
sense to use a more integrated approach.
Right so in terms of contributions and
preliminary insights.
We created a new approach to really look
at the supplier choice so that's what I'm
very much interested in and we transformed
it from a programming perspective.
And Illustrated performance
improvement and
I really you know when I view this is
a proof of concept that I think providing
choice to agents hand improve a platform's
objective and the value of and
choice increases as these different things
happen I will say perfectly concept is
important here because in terms of future
research this is basically what my N.S.F.
career said was what I just presented to
you was basically preliminary insights and
then future research is expanding
upon many many more ideas so
from a recommendation model the most
obvious thing is I'm doing a deterministic
optimization and then simulating the
utilities put that into the optimization
in terms of compensation and I think there
are some interesting things is how can we
influence suppliers utilities so if
there's one that's way out in the boonies
no one wants is there ways to
personalize that incentives so
bad is something I'm interested in
oftentimes if you really care about making
sure that the demand gets fulfilled
you likely will also have some sort of
resource that is owned by the platform and
so where the tradeoffs with that.
And then there's all the stuff
with dynamic model and so when and
how to update decisions so obviously you
have dynamically arriving requests but
there's also things I think
are super interesting about how
do you generate trust in this network and
how do you update the information once
a supplier is done how do I update both
should I have them participate again and
second of all how do I update their
view of them from the platform so
all my data needs to be updated and
can I be.
Smart about what I recommend such
that I get good data and estimate my
utilities in a really good way which we're
finding is you know preliminary results
is a really important thing if you
want to get better performance.
And then in terms of evaluation I'm
interested in kind of I'm calling
the three E.'s efficiency effectiveness
and equity and one reason for
that is one of my use cases is thinking
about how do we make deliveries
to nonprofits so
this is a map put out by the U.S.D.A.
of people living in what they call
a food desert which is defined here
as having no car and
then no supermarket within a mile and so
if you think about right now there
are definitely people trying to solve this
problem things like Meals on Wheels
regex's a local nonprofit but
when you talk to them they say OK I'm glad
to get your volunteer support but you have
to show up every Monday morning and every
monday when you make those deliver It's so
that's obviously limiting the supply
capacity that is available so
you know thinking about the bone
when you're old generation you know
can we create a more on demand volunteer
basis where you want to help but
you want to help on your own terms and so
I think that's something that's also very
exciting from a nonprofit perspective so
with that I think we need to rethink
supply chain design I think there's
lots of interesting things happening I put
this up here is that I'm very interested
in you know new mathematical models and
algorithms to valuate new systems and
processes to get ultimately insights.
And that's my presentation I would love
any questions concerns comments now or
offline that have.
Very.
Low.
Form.
For children from.