Today. - Peoplesatishr/cs270/sp13/slides/lec-5.pdf · Today. Quick Review: experts framework/multiplicative weights algorithm Finish: randomized multiplicative weights algorithm for

Today.

Quick Review:

experts framework/multiplicative weights algorithmFinish:

randomized multiplicative weights algorithm for experts framework.

Equilibrium for two person games:using experts framework/MW algorithm.

Today.

Quick Review:experts framework/multiplicative weights algorithm

Finish:randomized multiplicative weights algorithm for experts framework.


Today.


Finish:

randomized multiplicative weights algorithm for experts framework.


Today.




Today.




Notes.

Got to definition of Approximate Equilibrium for zero sumgames.

The multiplicative weights framework.

Expert’s framework.

n experts.

Every day, each offers a prediction.

“Rain” or “Shine.”

Whose advise do you follow?

“The one who is correct most often.”

Sort of.

How well do you do?


n experts.





Sort of.

How well do you do?


n experts.





Sort of.

How well do you do?


n experts.





Sort of.

How well do you do?


n experts.





Sort of.

How well do you do?


n experts.





Sort of.

How well do you do?


n experts.





Sort of.

How well do you do?

Infallible expert.One of the expert’s is infallible!

Your strategy?

Choose any expert that has not made a mistake!

How long to find perfect expert?

Maybe..never! Never see a mistake.

Better model?

How many mistakes could you make? Mistake Bound.

(A) 1(B) 2(C) logn(D) n−1

Adversary designs setup to watch who you choose, and makethat expert make a mistake.

n−1!


Your strategy?




Better model?


(A) 1(B) 2(C) logn(D) n−1


n−1!


Your strategy?




Better model?


(A) 1(B) 2(C) logn(D) n−1


n−1!


Your strategy?




Better model?


(A) 1(B) 2(C) logn(D) n−1


n−1!


Your strategy?



Maybe..

never! Never see a mistake.

Better model?


(A) 1(B) 2(C) logn(D) n−1


n−1!


Your strategy?



Maybe..never!

Never see a mistake.

Better model?


(A) 1(B) 2(C) logn(D) n−1


n−1!


Your strategy?




Better model?


(A) 1(B) 2(C) logn(D) n−1


n−1!


Your strategy?




Better model?


(A) 1(B) 2(C) logn(D) n−1


n−1!


Your strategy?




Better model?

How many mistakes could you make?

Mistake Bound.

(A) 1(B) 2(C) logn(D) n−1


n−1!


Your strategy?




Better model?


(A) 1(B) 2(C) logn(D) n−1


n−1!


Your strategy?




Better model?


(A) 1(B) 2(C) logn(D) n−1


n−1!


Your strategy?




Better model?


(A) 1(B) 2(C) logn(D) n−1


n−1!

Concept Alert.

Note.

Adversary:makes you want to look bad.”You could have done so well”...

but you didn’t! ha..ha!

Analysis of Algorithms: do as well as possible!

Concept Alert.

Note.

Adversary:

makes you want to look bad.”You could have done so well”...



Concept Alert.

Note.

Adversary:makes you want to look bad.

”You could have done so well”...but you didn’t! ha..ha!


Concept Alert.

Note.




Concept Alert.

Note.


but you didn’t!

ha..ha!


Concept Alert.

Note.


but you didn’t! ha..

ha!


Concept Alert.

Note.




Concept Alert.

Note.




Back to mistake bound.

Infallible Experts.

Alg: Choose one of the perfect experts.

Mistake Bound: n−1Lower bound: adversary argument.Upper bound: every mistake finds fallible expert.

Better Algorithm?

Making decision, not trying to find expert!

Algorithm: Go with the majority of previously correct experts.

What you would do anyway!


Infallible Experts.



Better Algorithm?





Infallible Experts.


Mistake Bound: n−1

Lower bound: adversary argument.Upper bound: every mistake finds fallible expert.

Better Algorithm?





Infallible Experts.


Mistake Bound: n−1Lower bound: adversary argument.

Upper bound: every mistake finds fallible expert.

Better Algorithm?





Infallible Experts.


Mistake Bound: n−1Lower bound: adversary argument.Upper bound:

every mistake finds fallible expert.

Better Algorithm?





Infallible Experts.



Better Algorithm?





Infallible Experts.



Better Algorithm?





Infallible Experts.



Better Algorithm?





Infallible Experts.



Better Algorithm?





Infallible Experts.



Better Algorithm?




Alg 2: find majority of the perfect

How many mistakes could you make?

(A) 1(B) 2(C) logn(D) n−1At most logn!

When alg makes a mistake,|“perfect” experts| drops by a factor of two.

Initially n perfect experts mistake→ ≤ n/2 perfect expertsmistake→ ≤ n/4 perfect experts

...mistake→ ≤ 1 perfect expert

≥ 1 perfect expert→ at most logn mistakes!


How many mistakes could you make?(A) 1(B) 2(C) logn(D) n−1

At most logn!






How many mistakes could you make?(A) 1(B) 2(C) logn(D) n−1At most logn!














Initially n perfect experts

mistake→ ≤ n/2 perfect expertsmistake→ ≤ n/4 perfect experts






Initially n perfect experts mistake→ ≤ n/2 perfect experts

mistake→ ≤ n/4 perfect experts...

mistake→ ≤ 1 perfect expert












...

mistake→ ≤ 1 perfect expert



















≥ 1 perfect expert

→ at most logn mistakes!







Imperfect Experts

Goal?

Do as well as the best expert!

Algorithm. Suggestions?

Go with majority?

Penalize inaccurate experts?

Best expert is penalized the least.

1. Initially: wi = 1.2. Predict with weighted majority of experts.3. wi → wi/2 if wrong.

Imperfect Experts

Goal?



Go with majority?




Imperfect Experts

Goal?


Algorithm.

Suggestions?

Go with majority?




Imperfect Experts

Goal?



Go with majority?




Imperfect Experts

Goal?



Go with majority?




Imperfect Experts

Goal?



Go with majority?




Imperfect Experts

Goal?



Go with majority?




Imperfect Experts

Goal?



Go with majority?



1. Initially: wi = 1.

2. Predict with weighted majority of experts.3. wi → wi/2 if wrong.

Imperfect Experts

Goal?



Go with majority?



1. Initially: wi = 1.2. Predict with weighted majority of experts.

3. wi → wi/2 if wrong.

Imperfect Experts

Goal?



Go with majority?




Imperfect Experts

Goal?



Go with majority?




Analysis: weighted majority

1. Initially: wi = 1.

2. Predict withweightedmajority ofexperts.

3. wi → wi/2 ifwrong.

Goal: Best expert makes m mistakes.

Potential function: ∑i wi . Initially n.

For best expert, b, wb ≥ 12m .

Each mistake:total weight of incorrect experts reduced by−1? −2? factor of 1

2?each incorrect expert weight multiplied by 1

2 !total weight decreases by

factor of 12? factor of 3

4?mistake→ ≥ half weight with incorrect experts.

Mistake→ potential function decreased by 34 .

We have

12m ≤∑

iwi ≤

(34

)M

n.

where M is number of algorithm mistakes.

Analysis: weighted majority1. Initially: wi = 1.












We have

12m ≤∑

iwi ≤

(34

)M

n.














We have

12m ≤∑

iwi ≤

(34

)M

n.






Potential function:

∑i wi . Initially n.








We have

12m ≤∑

iwi ≤

(34

)M

n.






Potential function: ∑i wi .

Initially n.








We have

12m ≤∑

iwi ≤

(34

)M

n.














We have

12m ≤∑

iwi ≤

(34

)M

n.














We have

12m ≤∑

iwi ≤

(34

)M

n.








Each mistake:

total weight of incorrect experts reduced by−1? −2? factor of 1






We have

12m ≤∑

iwi ≤

(34

)M

n.








Each mistake:total weight of incorrect experts reduced by

−1? −2? factor of 12?

each incorrect expert weight multiplied by 12 !

total weight decreases byfactor of 1

2? factor of 34?

mistake→ ≥ half weight with incorrect experts.


We have

12m ≤∑

iwi ≤

(34

)M

n.








Each mistake:total weight of incorrect experts reduced by−1?

−2? factor of 12?



2? factor of 34?



We have

12m ≤∑

iwi ≤

(34

)M

n.








Each mistake:total weight of incorrect experts reduced by−1? −2?

factor of 12?



2? factor of 34?



We have

12m ≤∑

iwi ≤

(34

)M

n.









2?



2? factor of 34?



We have

12m ≤∑

iwi ≤

(34

)M

n.










2 !


2? factor of 34?



We have

12m ≤∑

iwi ≤

(34

)M

n.














We have

12m ≤∑

iwi ≤

(34

)M

n.












4?



We have

12m ≤∑

iwi ≤

(34

)M

n.














We have

12m ≤∑

iwi ≤

(34

)M

n.














We have

12m ≤∑

iwi ≤

(34

)M

n.














We have

12m ≤∑

iwi ≤

(34

)M

n.


Analysis: continued.

12m ≤ ∑i wi ≤

(34

)Mn.

m - best expert mistakes M algorithm mistakes.

12m ≤

(34

)Mn.

Take log of both sides.

−m ≤−M log(4/3)+ logn.

Solve for M.M ≤ (m + logn)/ log(4/3)≤ 2.4(m + logn)

Multiple by 1− ε for incorrect experts...

(1− ε)m ≤(1− ε

2

)M n.

Massage...

M ≤ 2(1+ ε)m + 2lnnε

Approaches a factor of two of best expert performance!


12m ≤ ∑i wi ≤

(34

)Mn.

m - best expert mistakes

M algorithm mistakes.

12m ≤

(34

)Mn.





(1− ε)m ≤(1− ε

2

)M n.

Massage...

M ≤ 2(1+ ε)m + 2lnnε



12m ≤ ∑i wi ≤

(34

)Mn.


12m ≤

(34

)Mn.





(1− ε)m ≤(1− ε

2

)M n.

Massage...

M ≤ 2(1+ ε)m + 2lnnε



12m ≤ ∑i wi ≤

(34

)Mn.


12m ≤

(34

)Mn.





(1− ε)m ≤(1− ε

2

)M n.

Massage...

M ≤ 2(1+ ε)m + 2lnnε



12m ≤ ∑i wi ≤

(34

)Mn.


12m ≤

(34

)Mn.





(1− ε)m ≤(1− ε

2

)M n.

Massage...

M ≤ 2(1+ ε)m + 2lnnε



12m ≤ ∑i wi ≤

(34

)Mn.


12m ≤

(34

)Mn.





(1− ε)m ≤(1− ε

2

)M n.

Massage...

M ≤ 2(1+ ε)m + 2lnnε



12m ≤ ∑i wi ≤

(34

)Mn.


12m ≤

(34

)Mn.



Solve for M.M ≤ (m + logn)/ log(4/3)

≤ 2.4(m + logn)


(1− ε)m ≤(1− ε

2

)M n.

Massage...

M ≤ 2(1+ ε)m + 2lnnε



12m ≤ ∑i wi ≤

(34

)Mn.


12m ≤

(34

)Mn.





(1− ε)m ≤(1− ε

2

)M n.

Massage...

M ≤ 2(1+ ε)m + 2lnnε



12m ≤ ∑i wi ≤

(34

)Mn.


12m ≤

(34

)Mn.





(1− ε)m ≤(1− ε

2

)M n.

Massage...

M ≤ 2(1+ ε)m + 2lnnε



12m ≤ ∑i wi ≤

(34

)Mn.


12m ≤

(34

)Mn.





(1− ε)m ≤(1− ε

2

)M n.

Massage...

M ≤ 2(1+ ε)m + 2lnnε



12m ≤ ∑i wi ≤

(34

)Mn.


12m ≤

(34

)Mn.





(1− ε)m ≤(1− ε

2

)M n.

Massage...

M ≤ 2(1+ ε)m + 2lnnε



12m ≤ ∑i wi ≤

(34

)Mn.


12m ≤

(34

)Mn.





(1− ε)m ≤(1− ε

2

)M n.

Massage...

M ≤ 2(1+ ε)m + 2lnnε



12m ≤ ∑i wi ≤

(34

)Mn.


12m ≤

(34

)Mn.





(1− ε)m ≤(1− ε

2

)M n.

Massage...

M ≤ 2(1+ ε)m + 2lnnε


Best Analysis?

Two experts: A,B

Bad example?

Which is worse?(A) A right on even, B right on odd.(B) A right first half of days, B right second

Best expert peformance: T/2 mistakes.

Pattern (A): T −1 mistakes.

Factor of (almost) two worse!

Best Analysis?

Two experts: A,B

Bad example?





Best Analysis?

Two experts: A,B

Bad example?





Best Analysis?

Two experts: A,B

Bad example?





Best Analysis?

Two experts: A,B

Bad example?





Best Analysis?

Two experts: A,B

Bad example?





Randomization

!!!!

Better approach?

Use?

Randomization!

That is, choose expert i with prob ∝ wi

Bad example: A,B,A,B,A...

After a bit, A and B make nearly the same number of mistakes.

Choose each with approximately the same probabilty.

Make a mistake around 1/2 of the time.

Best expert makes T/2 mistakes.

Rougly optimal!

Randomization

!!!!

Better approach?

Use?

Randomization!







Rougly optimal!

Randomization!!!!

Better approach?

Use?

Randomization!







Rougly optimal!

Randomization!!!!

Better approach?

Use?

Randomization!







Rougly optimal!

Randomization!!!!

Better approach?

Use?

Randomization!







Rougly optimal!

Randomization!!!!

Better approach?

Use?

Randomization!







Rougly optimal!

Randomization!!!!

Better approach?

Use?

Randomization!







Rougly optimal!

Randomization!!!!

Better approach?

Use?

Randomization!







Rougly optimal!

Randomization!!!!

Better approach?

Use?

Randomization!







Rougly optimal!

Randomization!!!!

Better approach?

Use?

Randomization!







Rougly

optimal!

Randomization!!!!

Better approach?

Use?

Randomization!







Rougly optimal!

Randomized analysis.

Some formulas:

For ε ≤ 1,x ∈ [0,1],

(1+ ε)x ≤ (1+ εx)(1− ε)x ≤ (1− εx)

For ε ∈ [0, 12 ],

−ε− ε2 ≤ ln(1− ε)≤−ε

ε− ε2 ≤ ln(1+ ε)≤ ε

Proof Idea: ln(1+x) = x− x2

2 + x3

3 −·· ·


Some formulas:

For ε ≤ 1,x ∈ [0,1],

(1+ ε)x ≤ (1+ εx)(1− ε)x ≤ (1− εx)

For ε ∈ [0, 12 ],

−ε− ε2 ≤ ln(1− ε)≤−ε

ε− ε2 ≤ ln(1+ ε)≤ ε


2 + x3

3 −·· ·


Some formulas:

For ε ≤ 1,x ∈ [0,1],

(1+ ε)x ≤ (1+ εx)(1− ε)x ≤ (1− εx)

For ε ∈ [0, 12 ],

−ε− ε2 ≤ ln(1− ε)≤−ε

ε− ε2 ≤ ln(1+ ε)≤ ε


2 + x3

3 −·· ·


Some formulas:

For ε ≤ 1,x ∈ [0,1],

(1+ ε)x ≤ (1+ εx)(1− ε)x ≤ (1− εx)

For ε ∈ [0, 12 ],

−ε− ε2 ≤ ln(1− ε)≤−ε

ε− ε2 ≤ ln(1+ ε)≤ ε


2 + x3

3 −·· ·


Some formulas:

For ε ≤ 1,x ∈ [0,1],

(1+ ε)x ≤ (1+ εx)(1− ε)x ≤ (1− εx)

For ε ∈ [0, 12 ],

−ε− ε2 ≤ ln(1− ε)≤−ε

ε− ε2 ≤ ln(1+ ε)≤ ε


2 + x3

3 −·· ·


Some formulas:

For ε ≤ 1,x ∈ [0,1],

(1+ ε)x ≤ (1+ εx)(1− ε)x ≤ (1− εx)

For ε ∈ [0, 12 ],

−ε− ε2 ≤ ln(1− ε)≤−ε

ε− ε2 ≤ ln(1+ ε)≤ ε


2 + x3

3 −·· ·

Randomized algorithmLosses in [0,1].

Expert i loses `ti ∈ [0,1] in round t.

1. Initially wi = 1 for expert i .

2. Choose expert i with prob wiW , W = ∑i wi .

3. wi ← wi(1− ε)`ti

W (t) sum of wi at time t . W (0) = n

Best expert, b, loses L∗ total. →W (T )≥ wb ≥ (1− ε)L∗ .

Lt = ∑iwi `

ti

W expected loss of alg. in time t .

Claim: W (t +1)≤W (t)(1− εLt) Loss→ weight loss.

Proof:W (t +1)≤∑

i(1− ε`t

i )wi = ∑i

wi − ε ∑i

wi`ti

= ∑i

wi

(1− ε

∑i wi`ti

∑i wi

)= W (t)(1− εLt)





3. wi ← wi(1− ε)`ti



Lt = ∑iwi `

ti




i(1− ε`t

i )wi = ∑i

wi − ε ∑i

wi`ti

= ∑i

wi

(1− ε

∑i wi`ti

∑i wi

)= W (t)(1− εLt)





3. wi ← wi(1− ε)`ti



Lt = ∑iwi `

ti




i(1− ε`t

i )wi = ∑i

wi − ε ∑i

wi`ti

= ∑i

wi

(1− ε

∑i wi`ti

∑i wi

)= W (t)(1− εLt)





3. wi ← wi(1− ε)`ti



Lt = ∑iwi `

ti




i(1− ε`t

i )wi = ∑i

wi − ε ∑i

wi`ti

= ∑i

wi

(1− ε

∑i wi`ti

∑i wi

)= W (t)(1− εLt)





3. wi ← wi(1− ε)`ti



Lt = ∑iwi `

ti




i(1− ε`t

i )wi = ∑i

wi − ε ∑i

wi`ti

= ∑i

wi

(1− ε

∑i wi`ti

∑i wi

)= W (t)(1− εLt)





3. wi ← wi(1− ε)`ti

W (t) sum of wi at time t .

W (0) = n


Lt = ∑iwi `

ti




i(1− ε`t

i )wi = ∑i

wi − ε ∑i

wi`ti

= ∑i

wi

(1− ε

∑i wi`ti

∑i wi

)= W (t)(1− εLt)





3. wi ← wi(1− ε)`ti

W (t) sum of wi at time t . W (0) =

n


Lt = ∑iwi `

ti




i(1− ε`t

i )wi = ∑i

wi − ε ∑i

wi`ti

= ∑i

wi

(1− ε

∑i wi`ti

∑i wi

)= W (t)(1− εLt)





3. wi ← wi(1− ε)`ti



Lt = ∑iwi `

ti




i(1− ε`t

i )wi = ∑i

wi − ε ∑i

wi`ti

= ∑i

wi

(1− ε

∑i wi`ti

∑i wi

)= W (t)(1− εLt)





3. wi ← wi(1− ε)`ti


Best expert, b, loses L∗ total.

→W (T )≥ wb ≥ (1− ε)L∗ .

Lt = ∑iwi `

ti




i(1− ε`t

i )wi = ∑i

wi − ε ∑i

wi`ti

= ∑i

wi

(1− ε

∑i wi`ti

∑i wi

)= W (t)(1− εLt)





3. wi ← wi(1− ε)`ti



Lt = ∑iwi `

ti




i(1− ε`t

i )wi = ∑i

wi − ε ∑i

wi`ti

= ∑i

wi

(1− ε

∑i wi`ti

∑i wi

)= W (t)(1− εLt)





3. wi ← wi(1− ε)`ti



Lt = ∑iwi `

ti




i(1− ε`t

i )wi = ∑i

wi − ε ∑i

wi`ti

= ∑i

wi

(1− ε

∑i wi`ti

∑i wi

)= W (t)(1− εLt)





3. wi ← wi(1− ε)`ti



Lt = ∑iwi `

ti


Claim: W (t +1)≤W (t)(1− εLt)

Loss→ weight loss.


i(1− ε`t

i )wi = ∑i

wi − ε ∑i

wi`ti

= ∑i

wi

(1− ε

∑i wi`ti

∑i wi

)= W (t)(1− εLt)





3. wi ← wi(1− ε)`ti



Lt = ∑iwi `

ti




i(1− ε`t

i )wi = ∑i

wi − ε ∑i

wi`ti

= ∑i

wi

(1− ε

∑i wi`ti

∑i wi

)= W (t)(1− εLt)





3. wi ← wi(1− ε)`ti



Lt = ∑iwi `

ti



Proof:

W (t +1)≤∑i(1− ε`t

i )wi = ∑i

wi − ε ∑i

wi`ti

= ∑i

wi

(1− ε

∑i wi`ti

∑i wi

)= W (t)(1− εLt)





3. wi ← wi(1− ε)`ti



Lt = ∑iwi `

ti




i(1− ε`t

i )wi

= ∑i

wi − ε ∑i

wi`ti

= ∑i

wi

(1− ε

∑i wi`ti

∑i wi

)= W (t)(1− εLt)





3. wi ← wi(1− ε)`ti



Lt = ∑iwi `

ti




i(1− ε`t

i )wi = ∑i

wi − ε ∑i

wi`ti

= ∑i

wi

(1− ε

∑i wi`ti

∑i wi

)= W (t)(1− εLt)





3. wi ← wi(1− ε)`ti



Lt = ∑iwi `

ti




i(1− ε`t

i )wi = ∑i

wi − ε ∑i

wi`ti

= ∑i

wi

(1− ε

∑i wi`ti

∑i wi

)

= W (t)(1− εLt)





3. wi ← wi(1− ε)`ti



Lt = ∑iwi `

ti




i(1− ε`t

i )wi = ∑i

wi − ε ∑i

wi`ti

= ∑i

wi

(1− ε

∑i wi`ti

∑i wi

)= W (t)(1− εLt)

Analysis

(1− ε)L∗ ≤W (T )≤ n ∏t(1− εLt)

Take logs(L∗) ln(1− ε)≤ lnn +∑ ln(1− εLt)

Use −ε− ε2 ≤ ln(1− ε)≤−ε

−(L∗)(ε + ε2)≤ lnn− ε ∑Lt

And

∑t Lt ≤ (1+ ε)L∗+ lnnε

.

∑t Lt is total expected loss of algorithm.

Within (1+ ε) ish of the best expert!

No factor of 2 loss!

Analysis

(1− ε)L∗ ≤W (T )≤ n ∏t(1− εLt)


Use −ε− ε2 ≤ ln(1− ε)≤−ε

−(L∗)(ε + ε2)≤ lnn− ε ∑Lt

And


.




Analysis

(1− ε)L∗ ≤W (T )≤ n ∏t(1− εLt)


Use −ε− ε2 ≤ ln(1− ε)≤−ε

−(L∗)(ε + ε2)≤ lnn− ε ∑Lt

And


.




Analysis

(1− ε)L∗ ≤W (T )≤ n ∏t(1− εLt)


Use −ε− ε2 ≤ ln(1− ε)≤−ε

−(L∗)(ε + ε2)≤ lnn− ε ∑Lt

And


.




Analysis

(1− ε)L∗ ≤W (T )≤ n ∏t(1− εLt)


Use −ε− ε2 ≤ ln(1− ε)≤−ε

−(L∗)(ε + ε2)≤ lnn− ε ∑Lt

And


.




Analysis

(1− ε)L∗ ≤W (T )≤ n ∏t(1− εLt)


Use −ε− ε2 ≤ ln(1− ε)≤−ε

−(L∗)(ε + ε2)≤ lnn− ε ∑Lt

And


.




Analysis

(1− ε)L∗ ≤W (T )≤ n ∏t(1− εLt)


Use −ε− ε2 ≤ ln(1− ε)≤−ε

−(L∗)(ε + ε2)≤ lnn− ε ∑Lt

And


.




Analysis

(1− ε)L∗ ≤W (T )≤ n ∏t(1− εLt)


Use −ε− ε2 ≤ ln(1− ε)≤−ε

−(L∗)(ε + ε2)≤ lnn− ε ∑Lt

And


.


Within (1+ ε)

ish of the best expert!


Analysis

(1− ε)L∗ ≤W (T )≤ n ∏t(1− εLt)


Use −ε− ε2 ≤ ln(1− ε)≤−ε

−(L∗)(ε + ε2)≤ lnn− ε ∑Lt

And


.


Within (1+ ε) ish

of the best expert!


Analysis

(1− ε)L∗ ≤W (T )≤ n ∏t(1− εLt)


Use −ε− ε2 ≤ ln(1− ε)≤−ε

−(L∗)(ε + ε2)≤ lnn− ε ∑Lt

And


.




Analysis

(1− ε)L∗ ≤W (T )≤ n ∏t(1− εLt)


Use −ε− ε2 ≤ ln(1− ε)≤−ε

−(L∗)(ε + ε2)≤ lnn− ε ∑Lt

And


.




Gains.

Why so negative?

Each day, each expert gives gain in [0,1].

Multiplicative weights with (1+ ε)gti .

G ≥ (1− ε)G∗− lognε

where G∗ is payoff of best expert.

Scaling:

Not [0,1], say [0,ρ].

L≤ (1+ ε)L∗+ρ logn

ε

Gains.

Why so negative?





Scaling:

Not [0,1], say [0,ρ].


ε

Gains.

Why so negative?





Scaling:

Not [0,1], say [0,ρ].


ε

Gains.

Why so negative?





Scaling:

Not [0,1], say [0,ρ].


ε

Gains.

Why so negative?





Scaling:

Not [0,1], say [0,ρ].


ε

Gains.

Why so negative?





Scaling:

Not [0,1], say [0,ρ].


ε

Gains.

Why so negative?





Scaling:

Not [0,1], say [0,ρ].


ε

Summary: multiplicative weights.

Framework: n experts, each loses different amount every day.

Perfect Expert: logn mistakes.

Imperfect Expert: best makes m mistakes.

Deterministic Strategy: 2(1+ ε)m + lognε

Real numbered losses: Best loses L∗ total.

Randomized Strategy: (1+ ε)L∗+ lognε

Strategy:Choose proportional to weights

multiply weight by (1− ε)loss.

Multiplicative weights framework!

Applications next!











Applications next!











Applications next!











Applications next!











Applications next!











Applications next!











Applications next!








Strategy:

Choose proportional to weightsmultiply weight by (1− ε)loss.


Applications next!











Applications next!











Applications next!











Applications next!











Applications next!











Applications next!

Two person zero sum games.m×n payoff matrix A.

Row mixed strategy: x = (x1, . . . ,xm).Column mixed strategy: y = (y1, . . . ,yn).

Payoff for strategy pair (x ,y):

p(x ,y) = x tAy

That is,

∑i

xi

(∑j

ai ,jyj

)= ∑

j

(∑i

xiai ,j

)yj .

Recall row minimizes, column maximizes.

Equilibrium pair: (x∗,y∗)?

(x∗)tAy∗ = maxy

(x∗)tAy = minx

x tAy∗.

(No better column strategy, no better row strategy.)


Row mixed strategy: x = (x1, . . . ,xm).

Column mixed strategy: y = (y1, . . . ,yn).


p(x ,y) = x tAy

That is,

∑i

xi

(∑j

ai ,jyj

)= ∑

j

(∑i

xiai ,j

)yj .



(x∗)tAy∗ = maxy

(x∗)tAy = minx

x tAy∗.





p(x ,y) = x tAy

That is,

∑i

xi

(∑j

ai ,jyj

)= ∑

j

(∑i

xiai ,j

)yj .



(x∗)tAy∗ = maxy

(x∗)tAy = minx

x tAy∗.





p(x ,y) = x tAy

That is,

∑i

xi

(∑j

ai ,jyj

)= ∑

j

(∑i

xiai ,j

)yj .



(x∗)tAy∗ = maxy

(x∗)tAy = minx

x tAy∗.





p(x ,y) = x tAy

That is,

∑i

xi

(∑j

ai ,jyj

)= ∑

j

(∑i

xiai ,j

)yj .



(x∗)tAy∗ = maxy

(x∗)tAy = minx

x tAy∗.





p(x ,y) = x tAy

That is,

∑i

xi

(∑j

ai ,jyj

)= ∑

j

(∑i

xiai ,j

)yj .



(x∗)tAy∗ = maxy

(x∗)tAy = minx

x tAy∗.





p(x ,y) = x tAy

That is,

∑i

xi

(∑j

ai ,jyj

)= ∑

j

(∑i

xiai ,j

)yj .



(x∗)tAy∗ = maxy

(x∗)tAy = minx

x tAy∗.





p(x ,y) = x tAy

That is,

∑i

xi

(∑j

ai ,jyj

)= ∑

j

(∑i

xiai ,j

)yj .



(x∗)tAy∗ = maxy

(x∗)tAy = minx

x tAy∗.


Equilibrium.


p(x ,y) = (x∗)tAy∗ = maxy

(x∗)tAy = minx

x tAy∗.


No row is better:mini A(i) ·y = (x∗)tAy∗. 1

No column is better:maxj(At)(j) ·x = (x∗)tAy∗.

1A(i) is i th row.

Equilibrium.



(x∗)tAy = minx

x tAy∗.




1A(i) is i th row.

Equilibrium.



(x∗)tAy = minx

x tAy∗.




1A(i) is i th row.

Best Response

Column goes first:

Find y , where best row is not too low..

R = maxy

minx

(x tAy).

Note: x can be (0,0, . . . ,1, . . .0).

Example: Roshambo. Value of R?

Row goes first:Find x , where best column is not high.

C = minx

maxy

(x tAy).

Agin: y of form (0,0, . . . ,1, . . .0).

Example: Roshambo. Value of C?

Best Response

Column goes first:Find y , where best row is not too low..

R = maxy

minx

(x tAy).

Note: x can be (0,0, . . . ,1, . . .0).



C = minx

maxy

(x tAy).

Agin: y of form (0,0, . . . ,1, . . .0).


Best Response


R = maxy

minx

(x tAy).

Note: x can be (0,0, . . . ,1, . . .0).



C = minx

maxy

(x tAy).

Agin: y of form (0,0, . . . ,1, . . .0).


Best Response


R = maxy

minx

(x tAy).

Note: x can be (0,0, . . . ,1, . . .0).

Example: Roshambo.

Value of R?


C = minx

maxy

(x tAy).

Agin: y of form (0,0, . . . ,1, . . .0).


Best Response


R = maxy

minx

(x tAy).

Note: x can be (0,0, . . . ,1, . . .0).



C = minx

maxy

(x tAy).

Agin: y of form (0,0, . . . ,1, . . .0).


Best Response


R = maxy

minx

(x tAy).

Note: x can be (0,0, . . . ,1, . . .0).



C = minx

maxy

(x tAy).

Agin: y of form (0,0, . . . ,1, . . .0).


Best Response


R = maxy

minx

(x tAy).

Note: x can be (0,0, . . . ,1, . . .0).



C = minx

maxy

(x tAy).

Agin: y of form (0,0, . . . ,1, . . .0).


Best Response


R = maxy

minx

(x tAy).

Note: x can be (0,0, . . . ,1, . . .0).



C = minx

maxy

(x tAy).

Agin: y of form (0,0, . . . ,1, . . .0).


Best Response


R = maxy

minx

(x tAy).

Note: x can be (0,0, . . . ,1, . . .0).



C = minx

maxy

(x tAy).

Agin: y of form (0,0, . . . ,1, . . .0).

Example: Roshambo.

Value of C?

Best Response


R = maxy

minx

(x tAy).

Note: x can be (0,0, . . . ,1, . . .0).



C = minx

maxy

(x tAy).

Agin: y of form (0,0, . . . ,1, . . .0).


Duality.

R = maxy

minx

(x tAy).

C = minx

maxy

(x tAy).

Weak Duality: R ≤ C.Proof: Better to go second.

At Equilibrium (x∗,y∗), payoff v :row payoffs (Ay∗) all ≥ v =⇒ R ≥ v .column payoffs ((x∗)tA) all ≤ v =⇒ v ≥ C.=⇒ R ≥ C

Equilibrium =⇒ R = C!

Strong Duality: There is an equilibrium point! and R = C!

Doesn’t matter who plays first!

Duality.

R = maxy

minx

(x tAy).

C = minx

maxy

(x tAy).






Duality.

R = maxy

minx

(x tAy).

C = minx

maxy

(x tAy).






Duality.

R = maxy

minx

(x tAy).

C = minx

maxy

(x tAy).


At Equilibrium (x∗,y∗), payoff v :

row payoffs (Ay∗) all ≥ v =⇒ R ≥ v .column payoffs ((x∗)tA) all ≤ v =⇒ v ≥ C.=⇒ R ≥ C




Duality.

R = maxy

minx

(x tAy).

C = minx

maxy

(x tAy).


At Equilibrium (x∗,y∗), payoff v :row payoffs (Ay∗) all ≥ v

=⇒ R ≥ v .column payoffs ((x∗)tA) all ≤ v =⇒ v ≥ C.=⇒ R ≥ C




Duality.

R = maxy

minx

(x tAy).

C = minx

maxy

(x tAy).


At Equilibrium (x∗,y∗), payoff v :row payoffs (Ay∗) all ≥ v =⇒ R ≥ v .

column payoffs ((x∗)tA) all ≤ v =⇒ v ≥ C.=⇒ R ≥ C




Duality.

R = maxy

minx

(x tAy).

C = minx

maxy

(x tAy).


At Equilibrium (x∗,y∗), payoff v :row payoffs (Ay∗) all ≥ v =⇒ R ≥ v .column payoffs ((x∗)tA) all ≤ v

=⇒ v ≥ C.=⇒ R ≥ C




Duality.

R = maxy

minx

(x tAy).

C = minx

maxy

(x tAy).


At Equilibrium (x∗,y∗), payoff v :row payoffs (Ay∗) all ≥ v =⇒ R ≥ v .column payoffs ((x∗)tA) all ≤ v =⇒ v ≥ C.

=⇒ R ≥ C




Duality.

R = maxy

minx

(x tAy).

C = minx

maxy

(x tAy).






Duality.

R = maxy

minx

(x tAy).

C = minx

maxy

(x tAy).






Duality.

R = maxy

minx

(x tAy).

C = minx

maxy

(x tAy).




Strong Duality: There is an equilibrium point!

and R = C!


Duality.

R = maxy

minx

(x tAy).

C = minx

maxy

(x tAy).






Duality.

R = maxy

minx

(x tAy).

C = minx

maxy

(x tAy).






Proof of Equilibrium.

Later.

Still later...

Aproximate equilibrium ...

C(x) = maxy x tAy

R(y) = minx x tAy

Always: R(y)≤ C(x)

Strategy pair: (x ,y)

Equilibrium: (x ,y)

R(y) = C(x)→ C(x)−R(y) = 0.

Approximate Equilibrium: C(x)−R(y)≤ ε.

With R(y)≤ C(x)→ “Response y to x is within ε of best response”→ “Response x to y is within ε of best response”


Later. Still later...


C(x) = maxy x tAy

R(y) = minx x tAy



Equilibrium: (x ,y)

R(y) = C(x)→ C(x)−R(y) = 0.






C(x) = maxy x tAy

R(y) = minx x tAy



Equilibrium: (x ,y)

R(y) = C(x)→ C(x)−R(y) = 0.






C(x) = maxy x tAy

R(y) = minx x tAy



Equilibrium: (x ,y)

R(y) = C(x)→ C(x)−R(y) = 0.






C(x) = maxy x tAy

R(y) = minx x tAy



Equilibrium: (x ,y)

R(y) = C(x)→ C(x)−R(y) = 0.






C(x) = maxy x tAy

R(y) = minx x tAy



Equilibrium: (x ,y)

R(y) = C(x)→ C(x)−R(y) = 0.






C(x) = maxy x tAy

R(y) = minx x tAy



Equilibrium: (x ,y)

R(y) = C(x)→ C(x)−R(y) = 0.






C(x) = maxy x tAy

R(y) = minx x tAy



Equilibrium: (x ,y)

R(y) = C(x)→ C(x)−R(y) = 0.






C(x) = maxy x tAy

R(y) = minx x tAy



Equilibrium: (x ,y)

R(y) = C(x)

→ C(x)−R(y) = 0.






C(x) = maxy x tAy

R(y) = minx x tAy



Equilibrium: (x ,y)

R(y) = C(x)→ C(x)−R(y) = 0.






C(x) = maxy x tAy

R(y) = minx x tAy



Equilibrium: (x ,y)

R(y) = C(x)→ C(x)−R(y) = 0.






C(x) = maxy x tAy

R(y) = minx x tAy



Equilibrium: (x ,y)

R(y) = C(x)→ C(x)−R(y) = 0.


With R(y)≤ C(x)

→ “Response y to x is within ε of best response”→ “Response x to y is within ε of best response”




C(x) = maxy x tAy

R(y) = minx x tAy



Equilibrium: (x ,y)

R(y) = C(x)→ C(x)−R(y) = 0.


With R(y)≤ C(x)→ “Response y to x is within ε of best response”

→ “Response x to y is within ε of best response”




C(x) = maxy x tAy

R(y) = minx x tAy



Equilibrium: (x ,y)

R(y) = C(x)→ C(x)−R(y) = 0.






C(x) = maxy x tAy

R(y) = minx x tAy



Equilibrium: (x ,y)

R(y) = C(x)→ C(x)−R(y) = 0.



Proof of approximate equilibrium.

How?

(A) Using geometry.

(B) Using a fixed point theorem.(C) Using multiplicative weights.(D) By the skin of my teeth.

(C) ..and (D).Not hard. Even easy. Still, head scratching happens.


How?

(A) Using geometry.(B) Using a fixed point theorem.

(C) Using multiplicative weights.(D) By the skin of my teeth.



How?

(A) Using geometry.(B) Using a fixed point theorem.(C) Using multiplicative weights.

(D) By the skin of my teeth.



How?

(A) Using geometry.(B) Using a fixed point theorem.(C) Using multiplicative weights.(D) By the skin of my teeth.



How?


(C)

..and (D).Not hard. Even easy. Still, head scratching happens.


How?


(C) ..and (D).

Not hard. Even easy. Still, head scratching happens.


How?


(C) ..and (D).Not hard.

Even easy. Still, head scratching happens.


How?


(C) ..and (D).Not hard. Even easy.

Still, head scratching happens.


How?



Games and experts

Again: find (x∗,y∗), such that

(maxy x∗Ay)− (minx x∗Ay∗)≤ ε

C(x∗) − R(y∗)≤ ε

Experts Framework: n Experts, T days, L∗ -total loss.

Multiplicative Weights Method yields loss L where

L≤ (1+ ε)L∗+ lognε

Games and experts



C(x∗) − R(y∗)≤ ε




Games and experts



C(x∗) − R(y∗)≤ ε




Games and experts



C(x∗) − R(y∗)≤ ε




Games and experts



C(x∗) − R(y∗)≤ ε




Games and experts



C(x∗) − R(y∗)≤ ε




Games and Experts.

Assume: A has payoffs in [0,1].

For T = lognε2 days:

1) m pure row strategies are experts.Use multiplicative weights, produce row distribution.Let xt be distribution (row strategy) xt on day t .

2) Each day, adversary plays best column response to xt .Choose column of A that maximizes row’s expected loss.Let yt be indicator vector for this column.

Let y∗ = 1T ∑t yt and x∗ = argminxt

xtAyt .

Claim: (x∗,y)∗ are 2ε-optimal for matrix A.

Proof Idea:xt minimizes the best column response is chosen. Clearly good for row.

column best response is at least what it is against xt . Total loss, L is at leastcolumn payoff. Best row payoff, L∗ is roughly less than L due to MW anlysis.Combine bounds. Done!

Games and Experts.Assume: A has payoffs in [0,1].





xtAyt .









xtAyt .






1) m pure row strategies are experts.

Use multiplicative weights, produce row distribution.Let xt be distribution (row strategy) xt on day t .



xtAyt .






1) m pure row strategies are experts.Use multiplicative weights, produce row distribution.

Let xt be distribution (row strategy) xt on day t .



xtAyt .









xtAyt .







2) Each day, adversary plays best column response to xt .

Choose column of A that maximizes row’s expected loss.Let yt be indicator vector for this column.


xtAyt .







2) Each day, adversary plays best column response to xt .Choose column of A that maximizes row’s expected loss.

Let yt be indicator vector for this column.


xtAyt .









xtAyt .








Let y∗ = 1T ∑t yt

and x∗ = argminxtxtAyt .









xtAyt .









xtAyt .









xtAyt .


Proof Idea:

xt minimizes the best column response is chosen. Clearly good for row.column best response is at least what it is against xt . Total loss, L is at least

column payoff. Best row payoff, L∗ is roughly less than L due to MW anlysis.Combine bounds. Done!






xtAyt .


Proof Idea:xt minimizes the best column response is chosen.

Clearly good for row.column best response is at least what it is against xt . Total loss, L is at least

column payoff. Best row payoff, L∗ is roughly less than L due to MW anlysis.Combine bounds. Done!






xtAyt .









xtAyt .



column best response is at least what it is against xt .

Total loss, L is at leastcolumn payoff. Best row payoff, L∗ is roughly less than L due to MW anlysis.Combine bounds. Done!






xtAyt .



column best response is at least what it is against xt . Total loss, L is at leastcolumn payoff.

Best row payoff, L∗ is roughly less than L due to MW anlysis.Combine bounds. Done!






xtAyt .



column best response is at least what it is against xt . Total loss, L is at leastcolumn payoff. Best row payoff, L∗ is roughly less than L due to MW anlysis.

Combine bounds. Done!






xtAyt .



column best response is at least what it is against xt . Total loss, L is at leastcolumn payoff. Best row payoff, L∗ is roughly less than L due to MW anlysis.Combine bounds.

Done!






xtAyt .




Approximate Equilibrium!Experts: xt is strategy on day t , yt is best column against xt .


xtAyt .


Column payoff: C(x∗) = maxy x∗Ay .Loss on day t , xtAyt ≥ C(x∗) by the choice of x .Thus, algorithm loss, L, is ≥ TC(x∗).

Best expert: L∗- best row against all the columns played.

best row against ∑t Ayt and Ty∗ = ∑t yt→ best row against TAy∗.→ L∗ ≤ TR(y∗).

Multiplicative Weights: L≤ (1+ ε)L∗+ lnnε

TC(x∗)≤ (1+ ε)TR(y∗)+ lnnε→ C(x∗)≤ (1+ ε)R(y∗)+ lnn

εT→ C(x∗)−R(y∗)≤ εR(y∗)+ lnn

εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.


Let y∗ = 1T ∑t yt

and x∗ = argminxtxtAyt .








εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



xtAyt .








εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



xtAyt .








εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



xtAyt .


Column payoff: C(x∗) = maxy x∗Ay .

Loss on day t , xtAyt ≥ C(x∗) by the choice of x .Thus, algorithm loss, L, is ≥ TC(x∗).






εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



xtAyt .


Column payoff: C(x∗) = maxy x∗Ay .Loss on day t , xtAyt ≥ C(x∗) by the choice of x .

Thus, algorithm loss, L, is ≥ TC(x∗).






εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



xtAyt .








εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



xtAyt .








εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



xtAyt .




best row against ∑t Ayt

and Ty∗ = ∑t yt→ best row against TAy∗.→ L∗ ≤ TR(y∗).




εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



xtAyt .




best row against ∑t Ayt and Ty∗ = ∑t yt

→ best row against TAy∗.→ L∗ ≤ TR(y∗).




εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



xtAyt .




best row against ∑t Ayt and Ty∗ = ∑t yt→ best row against TAy∗.

→ L∗ ≤ TR(y∗).




εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



xtAyt .








εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



xtAyt .








εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



xtAyt .





Multiplicative Weights:

L≤ (1+ ε)L∗+ lnnε



εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



xtAyt .








εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



xtAyt .






TC(x∗)≤ (1+ ε)TR(y∗)+ lnnε

→ C(x∗)≤ (1+ ε)R(y∗)+ lnnεT

→ C(x∗)−R(y∗)≤ εR(y∗)+ lnnεT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



xtAyt .







εT


T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



xtAyt .








εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



xtAyt .








εT .

T = lnnε2 , R(y∗)≤ 1

→ C(x∗)−R(y∗)≤ 2ε.



xtAyt .








εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.

Approximate Equilibrium: notes!Experts: xt is strategy on day t , yt is best column against xt .

Let x∗ = 1T ∑t xt and y∗ = 1

T ∑t yt .


Column payoff: C(x∗) = maxy x∗Ay .Let yr be best response to C(x∗).Day t ,yt best response to xt → xtAyt ≥ xtAyr .Algorithm loss: ∑t xtAyt ≥ ∑t xtAyr

L≥ TC(x∗).






εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.


Let x∗ = 1T ∑t xt

and y∗ = 1T ∑t yt .



L≥ TC(x∗).






εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



T ∑t yt .



L≥ TC(x∗).






εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



T ∑t yt .



L≥ TC(x∗).






εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



T ∑t yt .


Column payoff: C(x∗) = maxy x∗Ay .

Let yr be best response to C(x∗).Day t ,yt best response to xt → xtAyt ≥ xtAyr .Algorithm loss: ∑t xtAyt ≥ ∑t xtAyr

L≥ TC(x∗).






εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



T ∑t yt .


Column payoff: C(x∗) = maxy x∗Ay .Let yr be best response to C(x∗).

Day t ,yt best response to xt → xtAyt ≥ xtAyr .Algorithm loss: ∑t xtAyt ≥ ∑t xtAyr

L≥ TC(x∗).






εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



T ∑t yt .


Column payoff: C(x∗) = maxy x∗Ay .Let yr be best response to C(x∗).Day t ,yt best response to xt → xtAyt ≥ xtAyr .

Algorithm loss: ∑t xtAyt ≥ ∑t xtAyrL≥ TC(x∗).






εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



T ∑t yt .



L≥ TC(x∗).






εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



T ∑t yt .



L≥ TC(x∗).






εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



T ∑t yt .



L≥ TC(x∗).






εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



T ∑t yt .



L≥ TC(x∗).


best row against ∑t Ayt

and Ty∗ = ∑t yt→ best row against TAy∗.→ L∗ ≤ TR(y∗).




εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



T ∑t yt .



L≥ TC(x∗).


best row against ∑t Ayt and Ty∗ = ∑t yt

→ best row against TAy∗.→ L∗ ≤ TR(y∗).




εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



T ∑t yt .



L≥ TC(x∗).


best row against ∑t Ayt and Ty∗ = ∑t yt→ best row against TAy∗.

→ L∗ ≤ TR(y∗).




εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



T ∑t yt .



L≥ TC(x∗).






εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



T ∑t yt .



L≥ TC(x∗).






εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



T ∑t yt .



L≥ TC(x∗).



Multiplicative Weights:

L≤ (1+ ε)L∗+ lnnε



εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



T ∑t yt .



L≥ TC(x∗).






εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



T ∑t yt .



L≥ TC(x∗).




TC(x∗)≤ (1+ ε)TR(y∗)+ lnnε

→ C(x∗)≤ (1+ ε)R(y∗)+ lnnεT


T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



T ∑t yt .



L≥ TC(x∗).





εT


T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



T ∑t yt .



L≥ TC(x∗).






εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.



T ∑t yt .



L≥ TC(x∗).






εT .

T = lnnε2 , R(y∗)≤ 1

→ C(x∗)−R(y∗)≤ 2ε.



T ∑t yt .



L≥ TC(x∗).






εT .

T = lnnε2 , R(y∗)≤ 1→ C(x∗)−R(y∗)≤ 2ε.

Comments

For any ε, there exists an ε-Approximate Equilibrium.

Does an equilibrium exist? Yes.

Something about math here? Fixed point theorem.

Later: will use geometry, linear programming.

Complexity?T = lnn

ε2 → O(nm lognε2 ). Basically linear!

Versus Linear Programming: O(n3m) Basically quadratic.(Faster linear programming: O(

√n +m) linear solution solves.)

Still much slower ... and more complicated.

Dynamics: best response, update weight, best response.

Also works with both using multiplicative weights.

“In practice.”

Comments


Does an equilibrium exist?

Yes.



Complexity?T = lnn







“In practice.”

Comments





Complexity?T = lnn







“In practice.”

Comments



Something about math here?

Fixed point theorem.


Complexity?T = lnn







“In practice.”

Comments





Complexity?T = lnn







“In practice.”

Comments





Complexity?T = lnn







“In practice.”

Comments





Complexity?

T = lnnε2 → O(nm logn

ε2 ). Basically linear!






“In practice.”

Comments





Complexity?T = lnn

ε2

→ O(nm lognε2 ). Basically linear!






“In practice.”

Comments





Complexity?T = lnn

ε2 → O(nm lognε2 ).

Basically linear!






“In practice.”

Comments





Complexity?T = lnn







“In practice.”

Comments





Complexity?T = lnn


Versus Linear Programming: O(n3m)

Basically quadratic.(Faster linear programming: O(





“In practice.”

Comments





Complexity?T = lnn


Versus Linear Programming: O(n3m) Basically quadratic.

(Faster linear programming: O(√

n +m) linear solution solves.)Still much slower ... and more complicated.



“In practice.”

Comments





Complexity?T = lnn







“In practice.”

Comments





Complexity?T = lnn




Still much slower

... and more complicated.



“In practice.”

Comments





Complexity?T = lnn







“In practice.”

Comments





Complexity?T = lnn







“In practice.”

Comments





Complexity?T = lnn







“In practice.”

Comments





Complexity?T = lnn







“In practice.”

Toll/Congestion

Given: G = (V ,E).Given (s1, t1) . . .(sk , tk ).Row: choose routing of all paths.Column: choose edge.Row pays if column chooses edge on any path.

Matrix:row for each routing: rcolumn for each edge: e

A[r ,e] is congestion on edge e by routing r

Offense: (Best Response.)Router: route along shortest paths.Toll: charge most loaded edge.

Defense: Toll: maximize shortest path under tolls.Route: minimize max loaded on any edge.

Toll/Congestion


Matrix:row for each routing: r

column for each edge: e




Toll/Congestion






Toll/Congestion






Toll/Congestion




Offense: (Best Response.)

Router: route along shortest paths.Toll: charge most loaded edge.


Toll/Congestion




Offense: (Best Response.)Router: route along shortest paths.

Toll: charge most loaded edge.


Toll/Congestion






Toll/Congestion





Defense: Toll: maximize shortest path under tolls.

Route: minimize max loaded on any edge.

Toll/Congestion






Toll/Congestion






Two person game.

Row is router.

An exponential number of rows.

Two person game with experts won’t be so easy to implement.

Version with row and column flipped may work.

Next Time.

Two person game.

Row is router.




Next Time.

Two person game.

Row is router.




Next Time.

Two person game.

Row is router.




Next Time.

Two person game.

Row is router.




Next Time.

Documents

Today. - Peoplesatishr/cs270/sp13/slides/lec-5.pdf · Today. Quick Review: experts framework/multiplicative weights algorithm Finish: randomized multiplicative weights algorithm for