the price of uncertainty maria-florina balcan georgia tech avrim blum carnegie mellon yishay mansour...

22
The Price of The Price of Uncertainty Uncertainty Maria-Florina Balcan Georgia Tech Avrim Blum Carnegie Mellon Yishay Mansour Tel- Aviv/Google ACM-EC 2009

Post on 21-Dec-2015

213 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: The Price of Uncertainty Maria-Florina Balcan Georgia Tech Avrim Blum Carnegie Mellon Yishay Mansour Tel-Aviv/Google ACM-EC 2009

The Price of The Price of UncertaintyUncertainty

Maria-Florina Balcan Georgia Tech

Avrim Blum Carnegie Mellon

Yishay Mansour Tel-Aviv/Google

ACM-EC 2009

Page 2: The Price of Uncertainty Maria-Florina Balcan Georgia Tech Avrim Blum Carnegie Mellon Yishay Mansour Tel-Aviv/Google ACM-EC 2009

Games are an abstractionGames are an abstractionCosts/payoffs are a function of players

actions. No external influences– If all repeatedly play same way, get exact

same payoff

Assume all players incentives are modeled

High-level question: What happens to natural dynamics if these don’t quite hold?

Page 3: The Price of Uncertainty Maria-Florina Balcan Georgia Tech Avrim Blum Carnegie Mellon Yishay Mansour Tel-Aviv/Google ACM-EC 2009

We’d like to think that if we get people into a good equilibrium, and players are selfish, reasonably myopic, etc, then behavior will stay there.

High-level question: What happens to natural dynamics if these don’t quite hold?

Games are an abstractionGames are an abstraction

Page 4: The Price of Uncertainty Maria-Florina Balcan Georgia Tech Avrim Blum Carnegie Mellon Yishay Mansour Tel-Aviv/Google ACM-EC 2009

We’d like to think that if we get people into a good equilibrium, and players are selfish, reasonably myopic, etc, then behavior will stay there.

But what if there are small fluctuations in underlying cost/payoff structure?

Games are an abstractionGames are an abstraction

Page 5: The Price of Uncertainty Maria-Florina Balcan Georgia Tech Avrim Blum Carnegie Mellon Yishay Mansour Tel-Aviv/Google ACM-EC 2009

We’d like to think that if we get people into a good equilibrium, and players are selfish, reasonably myopic, etc, then behavior will stay there.

But what if there are small fluctuations in underlying cost/payoff structure?

Or a few players acting unpredictably?

Can small fluctuations or a few unpredictable players cause natural dynamics to spin out of control and get system into a high-cost state?

Games are an abstractionGames are an abstraction

Page 6: The Price of Uncertainty Maria-Florina Balcan Georgia Tech Avrim Blum Carnegie Mellon Yishay Mansour Tel-Aviv/Google ACM-EC 2009

High-level questionHigh-level question

A few ways this could happen:– Small changes cause good equilibria to disappear,

only bad ones left. (economy?)– Bad behavior by a few players causes pain for all

(nukes)– Neither of above, but instead through more subtle

interaction with dynamics…

Focus here on this last issue.

Can small fluctuations or a few unpredictable players cause natural dynamics to spin out of control and get system into a high-cost state?

Page 7: The Price of Uncertainty Maria-Florina Balcan Georgia Tech Avrim Blum Carnegie Mellon Yishay Mansour Tel-Aviv/Google ACM-EC 2009

Focus of this workFocus of this workGames with the following properties:• Potential games, best/better-response

dynamics.• Low Price of Stability. (good equilibria exist)

– Game has good equilibria even after perturbation.

• Small gap between potential and social cost.– Single perturbation can’t make

dynamics do bad things.

• No individual player can influence cost of others by too much.

OPT

PoS £ OPT

cost

Page 8: The Price of Uncertainty Maria-Florina Balcan Georgia Tech Avrim Blum Carnegie Mellon Yishay Mansour Tel-Aviv/Google ACM-EC 2009

ModelModel• Players follow best (or better) response

dynamics.• Costs of resources can fluctuate between

moves: cit 2 [ci/(1+), ci(1+)]

(alternatively, one or more Byzantine players who move between time steps)

• Play begins in a low-cost state.• How bad can things get?

Price-of-Uncertainty() of game = maximum ratio of eventual social cost to initial cost.

Page 9: The Price of Uncertainty Maria-Florina Balcan Georgia Tech Avrim Blum Carnegie Mellon Yishay Mansour Tel-Aviv/Google ACM-EC 2009

ModelModel• Players follow best (or better) response

dynamics.• Costs of resources can fluctuate between

moves: cit 2 [ci/(1+), ci(1+)]

Price-of-Uncertainty() of game = maximum ratio of eventual social cost to initial cost.

Page 10: The Price of Uncertainty Maria-Florina Balcan Georgia Tech Avrim Blum Carnegie Mellon Yishay Mansour Tel-Aviv/Google ACM-EC 2009

ModelModel• Players follow best (or better) response dynamics.• Costs of resources can fluctuate between moves:

cit 2 [ci/(1+), ci(1+)]

Price-of-Uncertainty() of game = maximum ratio of eventual social cost to initial cost.

One way to look at this:• Define graph: one node for each state. Edge u ! v if

perturbation can cause BR to move from u to v.• What do the reachable sets

look like?

Page 11: The Price of Uncertainty Maria-Florina Balcan Georgia Tech Avrim Blum Carnegie Mellon Yishay Mansour Tel-Aviv/Google ACM-EC 2009

Set-cover gamesSet-cover games• n players (points)

• m sets, with costs c1,…,cm.

• Each player chooses some set that covers it.• Players split cost with all others choosing same

set. c1 c2 c3 cm

Page 12: The Price of Uncertainty Maria-Florina Balcan Georgia Tech Avrim Blum Carnegie Mellon Yishay Mansour Tel-Aviv/Google ACM-EC 2009

Set-cover gamesSet-cover games• Price of anarchy = n

• Price of stability = O(log(n)). • Potential 2 [cost, cost¢log(n)]

n- 1

Page 13: The Price of Uncertainty Maria-Florina Balcan Georgia Tech Avrim Blum Carnegie Mellon Yishay Mansour Tel-Aviv/Google ACM-EC 2009

Main resultsMain resultsSet-cover games:Set-cover games:• If = O(1/nm) then PoU = O(log n).• However, for any constant > 0, PoU = (n).• Also, a single Byzantine player can take state

from a PNE of cost O(OPT) to one of cost (n¢OPT).

Upper bounds hold even for better-response

Lower bounds hold even for best-response

Page 14: The Price of Uncertainty Maria-Florina Balcan Georgia Tech Avrim Blum Carnegie Mellon Yishay Mansour Tel-Aviv/Google ACM-EC 2009

Main resultsMain resultsGeneral fair-cost-sharing games:General fair-cost-sharing games:

Matroid congestion games:Matroid congestion games:• Here, “better-response” is easy to lure astray.

Can make state exponentially worse by using even an exponentially-small .

• So, focus on best-response.

s t

Page 15: The Price of Uncertainty Maria-Florina Balcan Georgia Tech Avrim Blum Carnegie Mellon Yishay Mansour Tel-Aviv/Google ACM-EC 2009

Main resultsMain resultsGeneral fair-cost-sharing games:General fair-cost-sharing games:• If many players for each (si,ti) pair (ni = (m)),

then PoU = O(1) even for constant >0.• Open for general number of players.

Matroid congestion games: Matroid congestion games: (strategy sets are bases of matroid. E.g., set-cover where must choose k sets)

• If = O(1/nm) then PoU = O(log n) for fair cost-sharing.

• In general, if = O(1/nm) then PoU = O(GAP).• Better-response not enough.

Also positive results for Also positive results for -nice games, bounds for -nice games, bounds for consensus, job schedulingconsensus, job scheduling

Page 16: The Price of Uncertainty Maria-Florina Balcan Georgia Tech Avrim Blum Carnegie Mellon Yishay Mansour Tel-Aviv/Google ACM-EC 2009

Set-Cover games Set-Cover games (lower bound)(lower bound)

Two kinds of players:• n of Class I:

• n-1 of Class II:

Plus one Byzantine player… Implies bound for case that costs can fluctuate

by factor of 2. More complicated ex for !1/n1/2.

n

n- n- n- n- n-

n/2 n/3 n/4 … n/n

Page 17: The Price of Uncertainty Maria-Florina Balcan Georgia Tech Avrim Blum Carnegie Mellon Yishay Mansour Tel-Aviv/Google ACM-EC 2009

Set-Cover games Set-Cover games (upper bound)(upper bound)

For upper bound, think of players in sets as a stack of chips.

• View ith position in stack j as having cost cj/i. Load chips with value equal to initial cost.

• When player moves from j to k, move top chip. Cost of position goes up by at most (1+)2.

cj ck

• At most mn different positions. So, following the path of any chip and removing loops, cost of final set is at most (1+)2nm times its value.

So, if = O(1/nm) then PoU = O(log n).

Page 18: The Price of Uncertainty Maria-Florina Balcan Georgia Tech Avrim Blum Carnegie Mellon Yishay Mansour Tel-Aviv/Google ACM-EC 2009

Matroid gamesMatroid gamesIn matroid games, can think of each player

as controlling a set of chips.

• Nice property of best response in matroids:– Can always order the move so that each

individual chip is doing better-response.• Apply previous argument.• Fails for better-response.

– Here, can get player to do kind of binary counting, bad even for exponentially-small .

Page 19: The Price of Uncertainty Maria-Florina Balcan Georgia Tech Avrim Blum Carnegie Mellon Yishay Mansour Tel-Aviv/Google ACM-EC 2009

Fair cost sharing in general Fair cost sharing in general graphsgraphs

If many players of each type, can also show best-response dynamics can’t do badly.

Outline of argument:• Hard to analyze cost of state directly, instead

track upper bound c*(St) = cost(S0 [ … [ St).

– c* changes at most m times.

• Many players of each type ) average cost of each is low compared to c* ) each change to c* is small.

Page 20: The Price of Uncertainty Maria-Florina Balcan Georgia Tech Avrim Blum Carnegie Mellon Yishay Mansour Tel-Aviv/Google ACM-EC 2009

PoU versus Price of AnarchyPoU versus Price of AnarchyMain focus: games with both good and bad

equilibria.• Can small fluctuations or single Byzantine

player cause behavior to move from good to bad?

Can also have cases where state can get worse even than worst equilibrium.

• Market-sharing: 2 vs log(n) even for best-response.

Page 21: The Price of Uncertainty Maria-Florina Balcan Georgia Tech Avrim Blum Carnegie Mellon Yishay Mansour Tel-Aviv/Google ACM-EC 2009

PoU versus Price of AnarchyPoU versus Price of Anarchynice games [AAEMS-EC08] : incentives grow

stronger as cost gets above times optimal (typically = PoA).

• Here, at least can show state won’t get above O() times optimal, even with substantial perturbation or many Byzantine players (random order).

Page 22: The Price of Uncertainty Maria-Florina Balcan Georgia Tech Avrim Blum Carnegie Mellon Yishay Mansour Tel-Aviv/Google ACM-EC 2009

Summary and open problemsSummary and open problemsLooking at: when can small perturbations or a few

bad players lead natural dynamics astray?• When is it safe to turn your back?

• Upper/lower bounds for a number of classes of games.

Open problems:• Close gap for set-cover [1/(mn), 1/n1/2].• General case of fair cost-sharing games?• Cyclic order best resp