Rationality as a Paradigm for Internet Computing

Slide 1 of 27Noam Nisan

Rationality as a Paradigm for

Internet Computing

Noam NisanHebrew University, Jerusalem

Slide 2 of 27Noam Nisan

Contents

• The Internet and the new face of computing• Analyzing computing systems in equilibrium • Designing computational mechanisms• A defining problem: Combinatorial auctions

Slide 3 of 27Noam Nisan

What is Computing?

20th Century(second half)

21st century(first decade)

von Neumann Machine The Internet

Slide 4 of 27Noam Nisan

The Internet

• Huge dynamic heterogeneous distributed system – “normal distributed CS”

• Not centrally owned – different parts owned by different people, firms, or organizations with differing goals – “CS+economics+game-theory”’

Slide 5 of 27Noam Nisan

TCP Retransmission Rule• Transmission Control Protocol

Used for most Internet communication Breaks messages into packets, and assembles the packets back into messages Handles packet delay/loss

• TCP Retransmission Rule When a packet is lost, decrease transmission rate (by a factor of

2) Rational: Network is congested – fix it by reducing demand

down to capacity• “Improved” Rule

When a packet is lost, start sending each packet twice Rational: Packets are lost – fix it by increasing the probability

that at least one copy of each packet arrives• Why not?

Slide 6 of 27Noam Nisan

Internet Resource Sharing• The vision

everyone connected to the Internet should have access to all resources that are connected to the Internet

• Examples: CPU-time Files I/O devices Data Knowledge Humans

• Why share?

Slide 7 of 27Noam Nisan

Electronic Commerce

• How will computers talk business?• Using communication, security software, agents, …

• Using standards: XML, .NET, J2EE, … and other TLAs

• What will they say to each other?• “Book X costs Y”

• “Bid X for Y units of stock Z”

• “Here’s a complicated offer to you guys: @#$%^ ”

Slide 8 of 27Noam Nisan

Internet Computing Protocols

• Should take into account Computational issues:

CPU time, communication, robustness, memory, languages, … Incentive issues:

Selfishness, strategies, payments, coalitions, risk, …

• Should combine the points of view of Computer Science and of economics

• Should apply game theory in a computational context

• Rational behavior is more easily assumed from computers than from humans The strategy is in the software

Slide 9 of 27Noam Nisan

At All Protocol Levels …

• eCommerce: eStores, auctions, exchanges, supply chains

• Online Services: games, web-hosting, ASPs• Information Resources: music, databases

• Computational resources: CPU, disk space, proxies, caching,

• Network Infrastructure: routing, admission control, QoS

Low level (traditional CS domain)

High level (traditional business domain)

Slide 10 of 27Noam Nisan

The Price of Anarchy Papadimitriou

• Take a “normal” CS protocol that works well if everyone does what they should….

• Say “Oh my god – the participating computers may do whatever they want…”

• Analyze what happens when “they do whatever they want”

• Radical departure from CS: “want” utility rationality game-theory equilibrium

• Aim to prove that things are still not too bad• Or else: argue against using on the Internet

Slide 11 of 27Noam Nisan

Minimizing Packet Delay Braess’s Paradox

1

1

x

x

0

delay proportionalto load

constant delay

• Many “small”packets – total quantity = 1

• Each knows the delay situation

• Each chooses how to get to destination

Slide 12 of 27Noam Nisan

Minimizing Packet Delay Braess’s Paradox

1

1

x

x

0

• Many “small”packets – total quantity = 1

• Each knows the delay situation

• Each chooses how to get to destination

Optimal routing (delay = 1.5)

1/2

0.50.5

1

1

11

0

Selfish routing (delay = 2.0)

Slide 13 of 27Noam Nisan

The Price of Anarchy is Low Roughgarden&Tardos

Theorem: for all network topologies, for all sets of routing requests, for all delay functions on the links:

1. If all delays are linear functions, then the previous example is as bad as it gets – the price of anarchy is at most a factor of 4/3 in delay

2. For general delay functions, doubling the edge capacities compensates for selfishness – the price of anarchy is at most a factor of 2 in infrastructure

Slide 14 of 27Noam Nisan

Algorithmic Mechanism Design Nisan&Ronen

• Design the protocols so that they will work well under selfish behavior of participants “work well” – the usual computational optimization goals “under selfish behavior” – the usual game-theoretic concepts

of equilibrium• Use notions and techniques from the economic field of

Mechanism Design “Inverse game-theory”

• Concentrate on “incentive compatibility” (truthfulness) Equilibrium is reached when all players report their private

information truthfully The revelation principle shows that this is without loss of

generality

Slide 15 of 27Noam Nisan

VCG-Mechanism in CS Vickrey-Clarke-Groves

Basic positive result in mechanism design Allow monetary transfers to/from participants Basic idea: internalize externalities Each player pays/gets the total loss/benefit in utility he

causes to all others All players see the same goal: optimizing the total sum of players’ utilities

Shared Cache

Caching XXX will save me

100$

Caching XXX will save me 10$

Caching XXX will cost me 80$

Pay 70 (=80-10)Clarke tax

Slide 16 of 27Noam Nisan

Beyond Classical Mechanism Design• New domain of problems

Parameter-complexity: e.g. structure of network Brave-new-world: disregard human conventions and biases

• New optimization goals Not just sum-of-utilities: e.g. make-span in scheduling

• New limitations Computational complexity Distributed implementation Interaction with usual mechanism design often problematic

• New biases regarding solution concepts Computer scientists don’t like Bayesian analysis: real-world

distributions are too different from those in our analysis – worst-case will happen

Computer scientists are happy with approximations: optimality is often too hard

Slide 17 of 27Noam Nisan

A Sampling of Some Recent Results• Selling “digital goods” (unlimited supply) Goldberg&Hartline&Wright

A randomized mechanism can approximate monopoly price revenue

• Scheduling jobs on “unrelated machines” Nisan&Ronen

No better than 2-approximation for the make-span is possible, but randomized mechanisms can do better

• Scheduling jobs on “related machines” Archer&Tardos

A polynomial time 3-approximation mechanism for the make-span• Cost-sharing for multicast transmissions FPS

VCG mechanism can be implemented in linear communication• Auctions using a few bits Blumrosen&Nisan

An auction with 1-bit from each player can achieve 98% efficiency

Slide 18 of 27Noam Nisan

Combinatorial Auctions• Most mechanism design problems involve resource

allocation• The central problem in classical mechanism design is an

auction: how to allocate a single indivisible good? Abstracts many resource allocation problems English auction, Dutch auction, first price sealed-bid auction, … Gold standard: Vickrey’s 2nd price auction

• The emerging central problem in algorithmic mechanism design is a combinatorial auction: how to allocate a collection of goods, with complex dependencies between them? Abstracts many complex resource allocation problems Involves a wide spectrum of computational and game-theoretic

issues

Slide 19 of 27Noam Nisan

Combinatorial Auction Problem Definition

• N indivisible non-identical items are sold concurrently • k bidders compete for subsets of these items• Each bidder j has a valuation for each set of items:

vj(S) = value that j assigns to acquiring the set S vj is monotonic non-decreasing (“free disposal”)

• Objective: Find a partition (S1…Sk) of {1..N} that maximizes the social welfare: j vj(Sj).

• Means: protocol between bidders and auctioneer • Difficulties: communication, computation, incentives

Slide 20 of 27Noam Nisan

Complements and Substitutes

• vj() may have complements: vj(ST) > vj(S)+vj(T) for some S and T. Extreme case: “single-minded bid” -- will only pay for a

complete package -- pay p for the set S but pay nothing for anything else

• vj() may have substitutes: vj(ST) < vj(S)+vj(T) for some disjoint S and T. Extreme case: “unit demand bid” -- will pay for at most a

single item – the price may depend on the item

Slide 21 of 27Noam Nisan

Routing as a Combinatorial Auction

Bidder A

Bidder B

Bidder C• Each bidder wants to buy some path to the destination

• Each link is an item

Destination

Slide 22 of 27Noam Nisan

The FCC Spectrum Auctions• The FCC auctions spectrum licenses for many

geographic regions and various frequency bands• These auctions have raised billions of dollars• The value of a license to a bidder depends on the

other licenses it holds• Currently licenses are sold in a simultaneous auction• USA Congress mandated that the next spectrum auction be made combinatorial.

3.1-3.2GHz

3.2-3.3GHz

3.3-3.4GHz

3.1-3.2GHz

3.2-3.3GHz

3.1-3.2GHz

3.2-3.3GHz

3.1-3.2GHz

3.2-3.3GHz

3.1-3.2GHz

3.2-3.3GHz

3.3-3.4GHz

Slide 23 of 27Noam Nisan

Basic Mechanism Design Approach• Basic Solution

Each bidder sends vj() to auctioneer. Auctioneer finds the partition that maximizes j vj(Sj). Auctioneer allocates Sj to each bidder j Auctioneer charges VCG payments – ensures incentive

compatibility• Computational difficulties

Bidding: How to send vj()? Requires communication of numbers – impractical Allocation: How can the auctioneer find an optimal allocation?

The problem is computationally intractable (even to approximate well)

N2

Slide 24 of 27Noam Nisan

Bidding Languages• The auction must fix a “language” for representing

valuations. All bidders will use that language to express their valuations Language must be expressive: express all reasonable

valuations succinctly Language must be simple: computationally easy to manage

valuations (represent, determine allocation,…)• Proposed languages use: package bids, OR, XOR

(left-sock & right-sock : 5$) OR ( (Red-shirt : 10$) XOR (blue-shirt : 9$))

• Different bidding languages have different power• What should the FCC allow?

Slide 25 of 27Noam Nisan

Iterative Auctions

Definition: The demand of valuation v at item prices p1 … pn is the set S that

maximizes the benefit: v(S)-i S pi A Walrasian equilibrium is an allocation S1…Sm and item prices p1

… pn such that each Sj is the demand of vj at these prices

Fact: Any Walrasian equilibrium gives an optimal allocation

Algorithm: Demange&Gale&Sotomayor

initialize prices of all items to 0 repeat: if an item is demanded by more than one bidder, increase the

price a little; until a Walrasian equilibrium is reachedTheorem: This works if valuations are “gross substitutes” Kelso&Crawford

Theorem: In general, exponential communication (equivalently, an exponential number of prices) is needed Nisan&Segal

Slide 26 of 27Noam Nisan

Allocation Algorithms• The allocation problem is computationally intractable • Approaches for overcoming computational difficulty

Solve (or approximate) special tractable cases• Gross substitutes Kelso&Crawford• Sub-modular (2-approximation) Lehmann&Lehmann&Nisan• Linear order on items Rothkopf&Pekec&Harstad

Heuristics that obtain optimal allocations and run “reasonable fast”

• Practical for 100s of items CABOB -- Sandholm et al. Heuristics that run quickly and find “reasonably good” solutions

• A few % loss for 1000s of items Zurel&Nisan

• Use the usual tools of combinatorial optimization LP relaxation Branch-and-bound, cutting-planes Local search Dynamic programming

Slide 27 of 27Noam Nisan

Incentives vs. Allocation• Challenge: find a mechanism that obtains “reasonably

good” allocations and is computationally efficient.• Key problem: Algorithms that find sub-optimal allocations

do not yield incentive compatible mechanisms Attaching VCG payments to sub-optimal algorithms essentially never

yields incentive compatibility Nisan&Ronen The only known incentive compatible mechanisms are VCG; for

“complete spaces” with at least 3 possible outcomes only VCG mechanisms exist. Roberts, Green&Laffont

• Special case: single minded bidders – have a single valuation parameter and desire a single package A Computationally efficient incentive compatible mechanism exists

Lehmann&Ocallaghan&Shoham

• Open problem: Find any non-VCG mechanism for any multi-dimensional valuation space