and Matroids Soheil Ehsani January 2018 Joint work with M Hajiaghayi T Kesselheim S Singla The problem consists of an initial setting and a sequence of events We have to take particular actions ID: 727166
Download Presentation The PPT/PDF document "Prophet Secretary for Combinatorial Auct..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Prophet Secretaryfor Combinatorial Auctionsand Matroids
Soheil EhsaniJanuary 2018
Joint work with M.
Hajiaghayi
, T.
Kesselheim
, S.
SinglaSlide2
The problem consists of an initial setting and a
sequence of events.We have to take particular actions
during time.The goal is maximize expected reward or minimize expected cost.
Two important fundamental problems are
Secretary problem and Prophet Inequality.
Online Decision MakingSlide3
Secretary Problem
There is an
unknown
set of numbers
and an
unknown
random permutation
.In an online fashion and at each step we observe the value of .The goal is to select one value , such that is maximized.First introduced by Dynkin in 1960’s and first studied in mechanism design by Hajiaghayi, Kleinberg and Parkes, EC’04 (introduced the value version v.s. the classic rank version and obtained the first bound for the general multi-choice version).
Slide4
Prophet Inequalities
Given
known
probability
distributions
In an online fashion and at each step
,
we observe a value .The goal is to select one value , upon its arrival, such that is maximized.First introduced by Krengel and Sucheston in 1970’s and first studied in computer science by Hajiaghayi, Kleinberg and Sandholm, AAAI’07 (introduced the first bound for the general multi-choice version). Slide5
Compute a threshold (virtual price)
.
Select the first item
where
.
The above method
-approximation (guarantees
).And is tight! A Simple Tight Algorithm! with probability otherwise
with probability
Worst Order
Best Order
Slide6
Prophet Secretary
Similar to prophet inequality, but the boxes arrive in a random permutation.
Given known
distributions
and
unknown
random permutation
.At each time , we observe and randomly draw .The goal is to make an irrevocable and immediate selection such that is maximized.Can we achieve an approximation factor better than for prophet secretary?There is a approximation algorithm using different thresholds [EHLM ESA’15].There is a -approximation algorithm for i.i.d.
and large market instances [AEEHKL STOC’17].
Slide7
Our Contribution
Prophet Inequality has been generalized to many settings:
½-approximation for matroids [KW STOC’12].½-approximation for combinatorial auctions
[FGL SODA’15].
-approximation for best-order
matroids [YAN SODA’11
].
Is there a better than ½-approximation for prophet inequality in random order? Combinatorial Auctions: we show a approximation algorithm (incentive-compatible for unit demand).Matroids: we show a approximation algorithm.Single Items: we show a single threshold approximation algorithm (tight). YESSlide8
Generalization: Combinatorial Auctions
Combinatorial Auctions:There is a set
of items and of buyers, each with a valuation function
.
We
want
allocate items to buyers to
maximize social welfare, i.e. Valuation functions are XOS, i.e. for some additive functions .Prophet Inequality Variant:We know distributions for buyers’ valuations.Buyers arrive one by one and we have to allocate a bundle to them immediately. Slide9
Generalization: Matroids
Matroids:Set of elements
such that
where
there exists such that .E.g. graphic matroids, transversal matroids.Prophet Inequality:Given a set of elements, their probability distribution on their values, and a matroid .At each step we draw and should decide weather to collect it (add to set ).The goal is to collect an independent set A of matroid with maximum value, i.e. . Slide10
Our Approach
First we convert the problem into a continuous-time setting.
Assume each buyer arrives at time
(independently at random).
Then we focus on designing dynamic-pricing algorithms: if we serve buyer
with assignment
Her gain is
Her virtual payment is This approach allows us to split the objective function into revenue and utility:
Achieving
for
Prophet Secretary
time-dependent
discount function
time-independent
base priceSlide11
Lower bounding Social Welfare
The goal is to maximize social welfare, i.e.
We define a
residual function
to capture the expected remaining value at time
.
Residual function is defined for a given instance of prophet secretary and a dynamic-pricing algorithm.Definition: a residual function has following properties:
Slide12
Proof Template
Let
be a prophet secretary instance
Propose a dynamic pricing algorithm
based on
and
Show there exists a residual function for , , and any Using a good discount function prove is approximation Slide13
Tuning the Discount Function
Proof:
Lemma: If an algorithm has a residual function then s
etting
results in a
-approximation algorithm.
Second property of residual functions says: .From integration by parts we know: .Therefore,
Now by adding Revenue and Utility we get:
Slide14
Algorithm: Combinatorial Auctions
Preprocess:Set a base price for each item based on its contribution to the optimum,
.
Draw
numbers
independently
and uniformly at
random from .Assignments:Buyer selects each remaining item if . Lemma: The following function is a residual for the algorithm: . Slide15
Algorithm: Single Item
Algorithm:Preprocess the base price
.
Draw
values
for arrival times.
The item is assigned to the first buyer
such that .The existence of a residual function in conjunction with the above tuning of the discount function results in a approximation algorithm.We need to show is a residual function.
Slide16
Analysis: Single Items
First property
:
Second property
: note that
is
probability
of selling item at time (for price ).Third property:
For the conditional utility of buyer
given arrival time
we have
Since
we have
Slide17
Conclusion
Present
approximation algorithms for most important generalizations of single item.Can we beat this barrier? [ACK arXiv 2017] improve by
for single item.
What about the i.i.d. case? [AEEHKL STOC’17] present 0.74 approximation algorithm for single item.
Mohamad
Hajiaghayi (UMD)Sahil Singla (CMU)
Thomas
Kesselheim
(TUD)
Thanks to my co-authors!