Near-optimal Algorithms for Online Linear Programming Zizhuo Wang - - PowerPoint PPT Presentation

near optimal algorithms for online linear programming
SMART_READER_LITE
LIVE PREVIEW

Near-optimal Algorithms for Online Linear Programming Zizhuo Wang - - PowerPoint PPT Presentation

Near-optimal Algorithms for Online Linear Programming Zizhuo Wang Department of Industrial and Systems Engineering University of Minnesota Joint work with Shipra Agrawal and Yinyu Ye July 7th, 2014 Zizhuo Wang Online Linear Programs, TOLA


slide-1
SLIDE 1

Near-optimal Algorithms for Online Linear Programming

Zizhuo Wang

Department of Industrial and Systems Engineering University of Minnesota Joint work with Shipra Agrawal and Yinyu Ye

July 7th, 2014

Zizhuo Wang Online Linear Programs, TOLA

slide-2
SLIDE 2

Introduction

Zizhuo Wang Online Linear Programs, TOLA

slide-3
SLIDE 3

Introduction

In traditional optimization problems, the input data is given (either in deterministic form or stochastic form).

Zizhuo Wang Online Linear Programs, TOLA

slide-4
SLIDE 4

Introduction

In traditional optimization problems, the input data is given (either in deterministic form or stochastic form). For example, in a linear program maximizex πTx subject to Ax = b x ≥ 0,

  • ne solves the decision variables all at once.

Zizhuo Wang Online Linear Programs, TOLA

slide-5
SLIDE 5

Introduction

In traditional optimization problems, the input data is given (either in deterministic form or stochastic form). For example, in a linear program maximizex πTx subject to Ax = b x ≥ 0,

  • ne solves the decision variables all at once.

However, in many practical problems, the input information is not available at the start, but reveals sequentially. Decisions have to be made in an online fashion.

Zizhuo Wang Online Linear Programs, TOLA

slide-6
SLIDE 6

Introduction

In traditional optimization problems, the input data is given (either in deterministic form or stochastic form). For example, in a linear program maximizex πTx subject to Ax = b x ≥ 0,

  • ne solves the decision variables all at once.

However, in many practical problems, the input information is not available at the start, but reveals sequentially. Decisions have to be made in an online fashion.

◮ Online linear programming problems

Zizhuo Wang Online Linear Programs, TOLA

slide-7
SLIDE 7

Introduction

Zizhuo Wang Online Linear Programs, TOLA

slide-8
SLIDE 8

Introduction

In this talk, we consider linear programs of the following format: maximizex n

t=1 πtxt

subject to n

t=1 aitxt ≤ Bi,

∀i = 1, ..., m 0 ≤ xt ≤ 1, ∀t = 1, ..., n

Zizhuo Wang Online Linear Programs, TOLA

slide-9
SLIDE 9

Introduction

In this talk, we consider linear programs of the following format: maximizex n

t=1 πtxt

subject to n

t=1 aitxt ≤ Bi,

∀i = 1, ..., m 0 ≤ xt ≤ 1, ∀t = 1, ..., n In the online version of the problem: we only know Bi’s at the start.

Zizhuo Wang Online Linear Programs, TOLA

slide-10
SLIDE 10

Introduction

In this talk, we consider linear programs of the following format: maximizex n

t=1 πtxt

subject to n

t=1 aitxt ≤ Bi,

∀i = 1, ..., m 0 ≤ xt ≤ 1, ∀t = 1, ..., n In the online version of the problem: we only know Bi’s at the start.

◮ The constraint matrix is revealed column by column

sequentially along with the corresponding objective coefficient.

Zizhuo Wang Online Linear Programs, TOLA

slide-11
SLIDE 11

Introduction

In this talk, we consider linear programs of the following format: maximizex n

t=1 πtxt

subject to n

t=1 aitxt ≤ Bi,

∀i = 1, ..., m 0 ≤ xt ≤ 1, ∀t = 1, ..., n In the online version of the problem: we only know Bi’s at the start.

◮ The constraint matrix is revealed column by column

sequentially along with the corresponding objective coefficient.

◮ An irrevocable decision must be made as soon as a column

arrives without observing or knowing the future data.

Zizhuo Wang Online Linear Programs, TOLA

slide-12
SLIDE 12

Applications

Zizhuo Wang Online Linear Programs, TOLA

slide-13
SLIDE 13

Applications

maximizex n

t=1 πtxt

subject to n

t=1 aitxt ≤ Bi,

∀i = 1, ..., m 0 ≤ xt ≤ 1, ∀t = 1, ..., n

Zizhuo Wang Online Linear Programs, TOLA

slide-14
SLIDE 14

Applications

maximizex n

t=1 πtxt

subject to n

t=1 aitxt ≤ Bi,

∀i = 1, ..., m 0 ≤ xt ≤ 1, ∀t = 1, ..., n This model is frequently used in resource allocation problems.

Zizhuo Wang Online Linear Programs, TOLA

slide-15
SLIDE 15

Applications

maximizex n

t=1 πtxt

subject to n

t=1 aitxt ≤ Bi,

∀i = 1, ..., m 0 ≤ xt ≤ 1, ∀t = 1, ..., n This model is frequently used in resource allocation problems.

◮ {ait}m i=1 are the request of a bundle of goods by tth customer

Zizhuo Wang Online Linear Programs, TOLA

slide-16
SLIDE 16

Applications

maximizex n

t=1 πtxt

subject to n

t=1 aitxt ≤ Bi,

∀i = 1, ..., m 0 ≤ xt ≤ 1, ∀t = 1, ..., n This model is frequently used in resource allocation problems.

◮ {ait}m i=1 are the request of a bundle of goods by tth customer ◮ Bi is the inventory for the ith good.

Zizhuo Wang Online Linear Programs, TOLA

slide-17
SLIDE 17

Applications

maximizex n

t=1 πtxt

subject to n

t=1 aitxt ≤ Bi,

∀i = 1, ..., m 0 ≤ xt ≤ 1, ∀t = 1, ..., n This model is frequently used in resource allocation problems.

◮ {ait}m i=1 are the request of a bundle of goods by tth customer ◮ Bi is the inventory for the ith good. ◮ πt is the price that the tth customer is willing to pay.

Zizhuo Wang Online Linear Programs, TOLA

slide-18
SLIDE 18

Applications

maximizex n

t=1 πtxt

subject to n

t=1 aitxt ≤ Bi,

∀i = 1, ..., m 0 ≤ xt ≤ 1, ∀t = 1, ..., n This model is frequently used in resource allocation problems.

◮ {ait}m i=1 are the request of a bundle of goods by tth customer ◮ Bi is the inventory for the ith good. ◮ πt is the price that the tth customer is willing to pay. ◮ xt is the decision whether to accept or reject the tth customer

Zizhuo Wang Online Linear Programs, TOLA

slide-19
SLIDE 19

Applications

maximizex n

t=1 πtxt

subject to n

t=1 aitxt ≤ Bi,

∀i = 1, ..., m 0 ≤ xt ≤ 1, ∀t = 1, ..., n This model is frequently used in resource allocation problems.

◮ {ait}m i=1 are the request of a bundle of goods by tth customer ◮ Bi is the inventory for the ith good. ◮ πt is the price that the tth customer is willing to pay. ◮ xt is the decision whether to accept or reject the tth customer

Customers arrive sequentially and an irrevocable decision must be made without observing future customer arrivals.

Zizhuo Wang Online Linear Programs, TOLA

slide-20
SLIDE 20

Applications

maximizex n

t=1 πtxt

subject to n

t=1 aitxt ≤ Bi,

∀i = 1, ..., m 0 ≤ xt ≤ 1, ∀t = 1, ..., n This model is frequently used in resource allocation problems.

◮ {ait}m i=1 are the request of a bundle of goods by tth customer ◮ Bi is the inventory for the ith good. ◮ πt is the price that the tth customer is willing to pay. ◮ xt is the decision whether to accept or reject the tth customer

Customers arrive sequentially and an irrevocable decision must be made without observing future customer arrivals.

◮ Applications: revenue management, channel allocation in

communication networks, charging allocation for electric vehicles, etc.

Zizhuo Wang Online Linear Programs, TOLA

slide-21
SLIDE 21

Our Objective

Zizhuo Wang Online Linear Programs, TOLA

slide-22
SLIDE 22

Our Objective

maximizex n

t=1 πtxt

subject to n

t=1 aitxt ≤ Bi,

∀i = 1, ..., m 0 ≤ xt ≤ 1, ∀t = 1, ..., n

Zizhuo Wang Online Linear Programs, TOLA

slide-23
SLIDE 23

Our Objective

maximizex n

t=1 πtxt

subject to n

t=1 aitxt ≤ Bi,

∀i = 1, ..., m 0 ≤ xt ≤ 1, ∀t = 1, ..., n We call the above problem the offline problem. And we denote the

  • ptimal value of it by OPT.

Zizhuo Wang Online Linear Programs, TOLA

slide-24
SLIDE 24

Our Objective

maximizex n

t=1 πtxt

subject to n

t=1 aitxt ≤ Bi,

∀i = 1, ..., m 0 ≤ xt ≤ 1, ∀t = 1, ..., n We call the above problem the offline problem. And we denote the

  • ptimal value of it by OPT.

Our objective: to find a decision rule for the online problem such that it achieves near-optimal performance, i.e., an algorithm that can achieve values close to OPT.

Zizhuo Wang Online Linear Programs, TOLA

slide-25
SLIDE 25

Model Assumptions

Zizhuo Wang Online Linear Programs, TOLA

slide-26
SLIDE 26

Model Assumptions

Main Assumptions

Zizhuo Wang Online Linear Programs, TOLA

slide-27
SLIDE 27

Model Assumptions

Main Assumptions

◮ We know the total number of columns n a priori

Zizhuo Wang Online Linear Programs, TOLA

slide-28
SLIDE 28

Model Assumptions

Main Assumptions

◮ We know the total number of columns n a priori ◮ The columns at arrive in a random order, i.e., the set of

columns together with their objective coefficients πt can be adversarily picked at the start. However, their arriving order is uniformly distributed over all the permutations

Zizhuo Wang Online Linear Programs, TOLA

slide-29
SLIDE 29

Model Assumptions

Main Assumptions

◮ We know the total number of columns n a priori ◮ The columns at arrive in a random order, i.e., the set of

columns together with their objective coefficients πt can be adversarily picked at the start. However, their arriving order is uniformly distributed over all the permutations The algorithm is evaluated on the expected performance over all the permutations comparing to the offline optimal solution, i.e., an algorithm A is c-competitive if and only if Eσ n

  • t=1

πtxt(σ, A)

  • ≥ c · OPT

Zizhuo Wang Online Linear Programs, TOLA

slide-30
SLIDE 30

Comment on the Assumptions

Zizhuo Wang Online Linear Programs, TOLA

slide-31
SLIDE 31

Comment on the Assumptions

Random permutation of input:

Zizhuo Wang Online Linear Programs, TOLA

slide-32
SLIDE 32

Comment on the Assumptions

Random permutation of input:

◮ An intermediate path between worst-case and i.i.d model.

Zizhuo Wang Online Linear Programs, TOLA

slide-33
SLIDE 33

Comment on the Assumptions

Random permutation of input:

◮ An intermediate path between worst-case and i.i.d model. ◮ Worst case model: Too conservative, no algorithm can

achieve better than O(1/n) approximation of the optimal

  • ffline solution [Babaioff et al 2008].

Zizhuo Wang Online Linear Programs, TOLA

slide-34
SLIDE 34

Comment on the Assumptions

Random permutation of input:

◮ An intermediate path between worst-case and i.i.d model. ◮ Worst case model: Too conservative, no algorithm can

achieve better than O(1/n) approximation of the optimal

  • ffline solution [Babaioff et al 2008].

◮ i.i.d. model: Need distribution information. Performance

might suffer if the actual input distribution is not as assumed.

Zizhuo Wang Online Linear Programs, TOLA

slide-35
SLIDE 35

Comment on the Assumptions

Random permutation of input:

◮ An intermediate path between worst-case and i.i.d model. ◮ Worst case model: Too conservative, no algorithm can

achieve better than O(1/n) approximation of the optimal

  • ffline solution [Babaioff et al 2008].

◮ i.i.d. model: Need distribution information. Performance

might suffer if the actual input distribution is not as assumed.

◮ Our assumption is strictly weaker than the i.i.d. model

Zizhuo Wang Online Linear Programs, TOLA

slide-36
SLIDE 36

Main Results

Zizhuo Wang Online Linear Programs, TOLA

slide-37
SLIDE 37

Main Results

Theorem

We propose an algorithm such that for any fixed ǫ > 0, our online algorithm is 1 − O(ǫ) competitive for online linear programming on all inputs when B = mini Bi ≥ Ω

  • m log (n/ǫ)

ǫ2

  • Zizhuo Wang

Online Linear Programs, TOLA

slide-38
SLIDE 38

Main Results

Theorem

We propose an algorithm such that for any fixed ǫ > 0, our online algorithm is 1 − O(ǫ) competitive for online linear programming on all inputs when B = mini Bi ≥ Ω

  • m log (n/ǫ)

ǫ2

  • Theorem

For any online algorithm for the online linear program in random permutation model, there exists an instance such that the competitive ratio is less than 1 − O(ǫ) if B = min

i

Bi ≤ log(m) ǫ2

Zizhuo Wang Online Linear Programs, TOLA

slide-39
SLIDE 39

Compare to Previous Results

Zizhuo Wang Online Linear Programs, TOLA

slide-40
SLIDE 40

Compare to Previous Results

k-secretary problem [Kleinberg 2005]: special case when m = 1, and at = 1 for all t. His algorithm is 1 − O(ǫ)-competitive for B ≥ 1

ǫ2

Zizhuo Wang Online Linear Programs, TOLA

slide-41
SLIDE 41

Compare to Previous Results

k-secretary problem [Kleinberg 2005]: special case when m = 1, and at = 1 for all t. His algorithm is 1 − O(ǫ)-competitive for B ≥ 1

ǫ2 ◮ We extend the result to multi-products case.

Zizhuo Wang Online Linear Programs, TOLA

slide-42
SLIDE 42

Compare to Previous Results

k-secretary problem [Kleinberg 2005]: special case when m = 1, and at = 1 for all t. His algorithm is 1 − O(ǫ)-competitive for B ≥ 1

ǫ2 ◮ We extend the result to multi-products case. ◮ The ǫ part of the bound is still unchanged, thus must be

  • ptimal

Zizhuo Wang Online Linear Programs, TOLA

slide-43
SLIDE 43

Compare to Previous Results

k-secretary problem [Kleinberg 2005]: special case when m = 1, and at = 1 for all t. His algorithm is 1 − O(ǫ)-competitive for B ≥ 1

ǫ2 ◮ We extend the result to multi-products case. ◮ The ǫ part of the bound is still unchanged, thus must be

  • ptimal

◮ We show, for the first time, that the dimension of the problem

m indeed adds to its difficulty (at least log m).

Zizhuo Wang Online Linear Programs, TOLA

slide-44
SLIDE 44

Compare to Previous Results

k-secretary problem [Kleinberg 2005]: special case when m = 1, and at = 1 for all t. His algorithm is 1 − O(ǫ)-competitive for B ≥ 1

ǫ2 ◮ We extend the result to multi-products case. ◮ The ǫ part of the bound is still unchanged, thus must be

  • ptimal

◮ We show, for the first time, that the dimension of the problem

m indeed adds to its difficulty (at least log m). Our algorithm is quite different from theirs due to the higher dimension

Zizhuo Wang Online Linear Programs, TOLA

slide-45
SLIDE 45

Compare to Previous Results

Zizhuo Wang Online Linear Programs, TOLA

slide-46
SLIDE 46

Compare to Previous Results

Online adwords problem [Devanur and Hayes 2009]: special case of the online linear programming problem. They have an algorithm that achieves 1 − ǫ-competitiveness if OPT ≥ Ω m2 log(mn) ǫ3

  • Zizhuo Wang

Online Linear Programs, TOLA

slide-47
SLIDE 47

Compare to Previous Results

Online adwords problem [Devanur and Hayes 2009]: special case of the online linear programming problem. They have an algorithm that achieves 1 − ǫ-competitiveness if OPT ≥ Ω m2 log(mn) ǫ3

  • ◮ We consider a much more general problem

Zizhuo Wang Online Linear Programs, TOLA

slide-48
SLIDE 48

Compare to Previous Results

Online adwords problem [Devanur and Hayes 2009]: special case of the online linear programming problem. They have an algorithm that achieves 1 − ǫ-competitiveness if OPT ≥ Ω m2 log(mn) ǫ3

  • ◮ We consider a much more general problem

◮ Their results is weaker by a factor or ǫ and also depend on

OPT which is not known until the problem is solved. (Our condition can be verified before solving the problem)

Zizhuo Wang Online Linear Programs, TOLA

slide-49
SLIDE 49

Key Ideas of Our Algorithm

Zizhuo Wang Online Linear Programs, TOLA

slide-50
SLIDE 50

Key Ideas of Our Algorithm

Dual Optimal Price

◮ For the offline linear program, there exists a dual price vector

p∗ for each goods such that, x∗

t = 1 if πt > aT t p∗ and x∗ t = 0

  • therwise, is near optimal

Zizhuo Wang Online Linear Programs, TOLA

slide-51
SLIDE 51

Key Ideas of Our Algorithm

Dual Optimal Price

◮ For the offline linear program, there exists a dual price vector

p∗ for each goods such that, x∗

t = 1 if πt > aT t p∗ and x∗ t = 0

  • therwise, is near optimal

Learning the price:

◮ Our online algorithm works by learning a price vector ˆ

  • p. The

price vector is determined by solving the dual problem using existing arrival data.

Zizhuo Wang Online Linear Programs, TOLA

slide-52
SLIDE 52

Key Ideas of Our Algorithm

Dual Optimal Price

◮ For the offline linear program, there exists a dual price vector

p∗ for each goods such that, x∗

t = 1 if πt > aT t p∗ and x∗ t = 0

  • therwise, is near optimal

Learning the price:

◮ Our online algorithm works by learning a price vector ˆ

  • p. The

price vector is determined by solving the dual problem using existing arrival data. We first show a one-time learning algorithm to illustrate this idea. Then we show that we can improve the algorithm by updating the price vector more frequently.

Zizhuo Wang Online Linear Programs, TOLA

slide-53
SLIDE 53

One-Time Learning Algorithm

Zizhuo Wang Online Linear Programs, TOLA

slide-54
SLIDE 54

One-Time Learning Algorithm

  • 1. Set xt = 0 for all t ≤ ǫn

Zizhuo Wang Online Linear Programs, TOLA

slide-55
SLIDE 55

One-Time Learning Algorithm

  • 1. Set xt = 0 for all t ≤ ǫn
  • 2. Solve the ǫ part of the problem

maximizex ǫn

t=1 πtxt

subject to ǫn

t=1 aitxt ≤ (1 − ǫ)ǫBi

i = 1, ..., m 0 ≤ xt ≤ 1 t = 1, ..., ǫn Get the optimal dual solution ˆ p;

Zizhuo Wang Online Linear Programs, TOLA

slide-56
SLIDE 56

One-Time Learning Algorithm

  • 1. Set xt = 0 for all t ≤ ǫn
  • 2. Solve the ǫ part of the problem

maximizex ǫn

t=1 πtxt

subject to ǫn

t=1 aitxt ≤ (1 − ǫ)ǫBi

i = 1, ..., m 0 ≤ xt ≤ 1 t = 1, ..., ǫn Get the optimal dual solution ˆ p;

  • 3. Determine the future allocation xt(ˆ

p) as: xt(ˆ p) = if πt ≤ ˆ pTat 1 if πt > ˆ pTat If aitxt(ˆ p) ≤ Bi − t−1

j=1 aijxj, set xt = xt(ˆ

p); otherwise, set xt = 0.

Zizhuo Wang Online Linear Programs, TOLA

slide-57
SLIDE 57

Analysis Idea

Zizhuo Wang Online Linear Programs, TOLA

slide-58
SLIDE 58

Analysis Idea

By the complementarity conditions of LP, if we can show that n

t=1 aitxt(ˆ

p) = Bi for each i, then xt(ˆ p) is the optimal solution to the offline problem.

Zizhuo Wang Online Linear Programs, TOLA

slide-59
SLIDE 59

Analysis Idea

By the complementarity conditions of LP, if we can show that n

t=1 aitxt(ˆ

p) = Bi for each i, then xt(ˆ p) is the optimal solution to the offline problem.

Lemma

With probability 1 − ǫ, (1 − 3ǫ)Bi ≤

n

  • t=1

aitxt(ˆ p) ≤ Bi, ∀i = 1, . . . , m given B ≥ 6m log(n/ǫ)

ǫ3

.

Zizhuo Wang Online Linear Programs, TOLA

slide-60
SLIDE 60

Analysis Idea

By the complementarity conditions of LP, if we can show that n

t=1 aitxt(ˆ

p) = Bi for each i, then xt(ˆ p) is the optimal solution to the offline problem.

Lemma

With probability 1 − ǫ, (1 − 3ǫ)Bi ≤

n

  • t=1

aitxt(ˆ p) ≤ Bi, ∀i = 1, . . . , m given B ≥ 6m log(n/ǫ)

ǫ3

. The proof uses intensively concentration inequalities (Hoeffding-Bernstein Inequalities) and union bound arguments

Zizhuo Wang Online Linear Programs, TOLA

slide-61
SLIDE 61

Analysis Idea Continued

Zizhuo Wang Online Linear Programs, TOLA

slide-62
SLIDE 62

Analysis Idea Continued

However, by the complementarity conditions, x(ˆ p) is optimal to

Zizhuo Wang Online Linear Programs, TOLA

slide-63
SLIDE 63

Analysis Idea Continued

However, by the complementarity conditions, x(ˆ p) is optimal to maximizex

  • t πtxt

subject to

  • t aitxt ≤

t aitxt(ˆ

p) i = 1, ..., m 0 ≤ xt ≤ 1 t = 1, ..., n

Zizhuo Wang Online Linear Programs, TOLA

slide-64
SLIDE 64

Analysis Idea Continued

However, by the complementarity conditions, x(ˆ p) is optimal to maximizex

  • t πtxt

subject to

  • t aitxt ≤

t aitxt(ˆ

p) i = 1, ..., m 0 ≤ xt ≤ 1 t = 1, ..., n Then by the above lemmas, it is easy to show that with high probability n

t=1 πtxt(ˆ

p) ≥ (1 − 3ǫ)OPT.

Zizhuo Wang Online Linear Programs, TOLA

slide-65
SLIDE 65

Analysis Idea Continued

Zizhuo Wang Online Linear Programs, TOLA

slide-66
SLIDE 66

Analysis Idea Continued

Another thing we need to take care of is that the algorithm does not make allocation in the first ǫn period.

Zizhuo Wang Online Linear Programs, TOLA

slide-67
SLIDE 67

Analysis Idea Continued

Another thing we need to take care of is that the algorithm does not make allocation in the first ǫn period.

Lemma

Let OPT(S) denote the optimal value of the linear program maximizex

  • t∈S πtxt

subject to

  • t∈S aitxt ≤ ǫBi,

i = 1, ..., m 0 ≤ xt ≤ 1, t ∈ S.

  • ver random sample S ⊂ N where |S| = ǫ|N|, and OPT(N) denote

the optimal value of the original offline linear program. Then, E[OPT(S)] ≤ ǫOPT(N)

Zizhuo Wang Online Linear Programs, TOLA

slide-68
SLIDE 68

Analysis Idea Continued

Zizhuo Wang Online Linear Programs, TOLA

slide-69
SLIDE 69

Analysis Idea Continued

Summarizing all the lemmas above:

Zizhuo Wang Online Linear Programs, TOLA

slide-70
SLIDE 70

Analysis Idea Continued

Summarizing all the lemmas above:

◮ With high probability, we never violate the inventory constraint

Zizhuo Wang Online Linear Programs, TOLA

slide-71
SLIDE 71

Analysis Idea Continued

Summarizing all the lemmas above:

◮ With high probability, we never violate the inventory constraint ◮ With high probability, the objective value is near-optimal if we

include the initial ǫ portion

Zizhuo Wang Online Linear Programs, TOLA

slide-72
SLIDE 72

Analysis Idea Continued

Summarizing all the lemmas above:

◮ With high probability, we never violate the inventory constraint ◮ With high probability, the objective value is near-optimal if we

include the initial ǫ portion

◮ With high probability, the first ǫ portion of the objective

value, a learning cost, doesn’t contribute too much.

Zizhuo Wang Online Linear Programs, TOLA

slide-73
SLIDE 73

Analysis Idea Continued

Summarizing all the lemmas above:

◮ With high probability, we never violate the inventory constraint ◮ With high probability, the objective value is near-optimal if we

include the initial ǫ portion

◮ With high probability, the first ǫ portion of the objective

value, a learning cost, doesn’t contribute too much. Therefore, we proved that our one-time learning algorithm is 1 − O(ǫ)-competitive if B ≥ 6m log(n/ǫ)

ǫ3

.

Zizhuo Wang Online Linear Programs, TOLA

slide-74
SLIDE 74

Dynamic Price Updating Algorithm

Zizhuo Wang Online Linear Programs, TOLA

slide-75
SLIDE 75

Dynamic Price Updating Algorithm

The one-time learning algorithm is simple, but the condition required on the size of B is stronger than the main theorem claims

Zizhuo Wang Online Linear Programs, TOLA

slide-76
SLIDE 76

Dynamic Price Updating Algorithm

The one-time learning algorithm is simple, but the condition required on the size of B is stronger than the main theorem claims

◮ ǫ3 versus ǫ2

Zizhuo Wang Online Linear Programs, TOLA

slide-77
SLIDE 77

Dynamic Price Updating Algorithm

The one-time learning algorithm is simple, but the condition required on the size of B is stronger than the main theorem claims

◮ ǫ3 versus ǫ2

The one-time learning algorithm only computes the price once.

Zizhuo Wang Online Linear Programs, TOLA

slide-78
SLIDE 78

Dynamic Price Updating Algorithm

The one-time learning algorithm is simple, but the condition required on the size of B is stronger than the main theorem claims

◮ ǫ3 versus ǫ2

The one-time learning algorithm only computes the price once.

◮ Potential improvement might be made by updating the price

dynamically during the process.

Zizhuo Wang Online Linear Programs, TOLA

slide-79
SLIDE 79

Dynamic Price Updating Algorithm

The one-time learning algorithm is simple, but the condition required on the size of B is stronger than the main theorem claims

◮ ǫ3 versus ǫ2

The one-time learning algorithm only computes the price once.

◮ Potential improvement might be made by updating the price

dynamically during the process. Question

Zizhuo Wang Online Linear Programs, TOLA

slide-80
SLIDE 80

Dynamic Price Updating Algorithm

The one-time learning algorithm is simple, but the condition required on the size of B is stronger than the main theorem claims

◮ ǫ3 versus ǫ2

The one-time learning algorithm only computes the price once.

◮ Potential improvement might be made by updating the price

dynamically during the process. Question

◮ How often should we update the price?

Zizhuo Wang Online Linear Programs, TOLA

slide-81
SLIDE 81

Dynamic Price Updating Algorithm

Zizhuo Wang Online Linear Programs, TOLA

slide-82
SLIDE 82

Dynamic Price Updating Algorithm

In the dynamic price updating algorithm, we update the price at time ǫn, 2ǫn, 4ǫn...

Zizhuo Wang Online Linear Programs, TOLA

slide-83
SLIDE 83

Dynamic Price Updating Algorithm

In the dynamic price updating algorithm, we update the price at time ǫn, 2ǫn, 4ǫn... At time ℓ ∈ {ǫn, 2ǫn, ...}, the price is the optimal dual solution to the following linear program:

Zizhuo Wang Online Linear Programs, TOLA

slide-84
SLIDE 84

Dynamic Price Updating Algorithm

In the dynamic price updating algorithm, we update the price at time ǫn, 2ǫn, 4ǫn... At time ℓ ∈ {ǫn, 2ǫn, ...}, the price is the optimal dual solution to the following linear program: maximizex ℓ

t=1 πtxt

subject to ℓ

t=1 aitxt ≤ (1 − hℓ) ℓ nbi

i = 1, ..., m 0 ≤ xt ≤ 1 t = 1, ..., ℓ where hℓ = ǫ n ℓ

Zizhuo Wang Online Linear Programs, TOLA

slide-85
SLIDE 85

Dynamic Price Updating Algorithm

In the dynamic price updating algorithm, we update the price at time ǫn, 2ǫn, 4ǫn... At time ℓ ∈ {ǫn, 2ǫn, ...}, the price is the optimal dual solution to the following linear program: maximizex ℓ

t=1 πtxt

subject to ℓ

t=1 aitxt ≤ (1 − hℓ) ℓ nbi

i = 1, ..., m 0 ≤ xt ≤ 1 t = 1, ..., ℓ where hℓ = ǫ n ℓ And this price is used to determine the allocation for the next immediate period.

Zizhuo Wang Online Linear Programs, TOLA

slide-86
SLIDE 86

Geometric Pace of Price Updating

Zizhuo Wang Online Linear Programs, TOLA

slide-87
SLIDE 87

Dynamic Price Updating Algorithm

Zizhuo Wang Online Linear Programs, TOLA

slide-88
SLIDE 88

Dynamic Price Updating Algorithm

◮ In this algorithm, we update the price log2 (1/ǫ) times during

the entire time horizon.

Zizhuo Wang Online Linear Programs, TOLA

slide-89
SLIDE 89

Dynamic Price Updating Algorithm

◮ In this algorithm, we update the price log2 (1/ǫ) times during

the entire time horizon.

◮ The numbers hℓ play an important role in improving the

condition on B in our main theorem. It balances the probability that the inventory constraint ever gets violated and the lost of objective value due to the factor 1 − hℓ.

Zizhuo Wang Online Linear Programs, TOLA

slide-90
SLIDE 90

Dynamic Price Updating Algorithm

◮ In this algorithm, we update the price log2 (1/ǫ) times during

the entire time horizon.

◮ The numbers hℓ play an important role in improving the

condition on B in our main theorem. It balances the probability that the inventory constraint ever gets violated and the lost of objective value due to the factor 1 − hℓ.

◮ Choosing large hℓ (more conservative) at the beginning

periods and smaller hℓ (more risk neutral) at the later periods, we can control the loss of objective value by an ǫ order while the required size of B can be weakened by an ǫ factor.

Zizhuo Wang Online Linear Programs, TOLA

slide-91
SLIDE 91

Proof Outline of the Dynamic Price Updating Algorithm

The proof is similar to the proof of the one-time learning

  • algorithm. We first show the following lemma:

Lemma

For any ǫ > 0, with probability 1 − ǫ:

2ℓ

  • t=ℓ+1

aitxt(ˆ pℓ) ≤ ℓ nbi, for all i ∈ {1, . . . , m}, ℓ ∈ {ǫn, 2ǫn, ...} given B = mini bi ≥ 10m log (n/ǫ)

ǫ2

. This lemma states that for each time period, with high probability, the inventory consumed by x(ˆ p) is less than the proportional total inventory.

Zizhuo Wang Online Linear Programs, TOLA

slide-92
SLIDE 92

Proof Outline of the Dynamic Price Updating Algorithm

Next we need to show that by introducing a factor hℓ, the loss of

  • bjective value is small.

It is easy to see that at period ℓ, the lost of objective value is about hℓ · ℓ

  • nOPT. Then the total loss of objective value is about
  • ℓ∈{ǫn,2ǫn,...} hℓ ℓ

nOPT

= ǫ

ℓ∈{ǫn,2ǫn,...}

nOPT

≤ ǫOPT

1 2i

≤ O(ǫOPT) Therefore, the loss of objective value is very small. And we can conclude that this algorithm gives a near-optimal solution.

Zizhuo Wang Online Linear Programs, TOLA

slide-93
SLIDE 93

Summary

We propose a dynamic near-optimal algorithm for a class of online linear programming problems under the random permutation model

◮ It solves linear programs for dual price based on revealed data,

and uses these prices to make future allocations

◮ The algorithm has the feature of “learning-while-doing”, and

the pace the price is updated is neither too fast nor too slow

◮ The application includes various online resource allocation and

revenue management problems

Zizhuo Wang Online Linear Programs, TOLA