An Adaptive Proportional Value-per-Click Agent for Bidding in Ad Auctions
Agent Negotiation via Auctions
description
Transcript of Agent Negotiation via Auctions
Outline
• Market/Negotiation overview• Computational Market Systems
– Blue Skies/Mobile economy– University of Michigan Digital Library
(UMDL) Information Economy– Auction Manager middleware
• Buyer/Seller’s choice bundles• Market policies
• Case Study: Bandwidth Exchanges• Future Directions
Why markets?
• Ronald Coase: The Nature of the Firm (1937)– Alternative modes organizing transactions:
• Markets: decentralized, price signals• Firms: hierarchies
– Why do we have any firms? Why don’t we have just one mega-firm?
Why markets?
• Ronald Coase: The Nature of the Firm (1937)– Alternative modes organizing transactions:
• Markets: decentralized, price signals• Firms: hierarchies
– Why do we have any firms? Why don’t we have just one mega-firm? TRANSACTION COSTS
– Other factors related to transaction costs: price discovery costs, information asymmetries, uncertainty, limits of 3rd party enforcement
One View of Commerce Fundamentals
Step 1What do I want?
Where is it?
Step 2How much is it?
What should I bid?
Step 3What should I pay?How should I pay?
Infrastructure
Discover Negotiate Exchange
More Details
Business Models/Social & Legal Environment
Discover:• Advertisements• Junk mail/coupons• Catalogs• Browse/Shop• Consumer Reports
Negotiate:• Price tag• Barter• Auction• Stock Market
Exchange:•Payment type: $$/check/credit card•Delivery options•Follow up care
Interconnection Medium
Physical => Internet
• Spatial restrictions of current physical markets often no longer apply.– Participants no longer have to be spatially co-located.
• Lower transaction costs lead to new marketplaces: E-trade, eBay, Priceline.com, DemandLine.com.
– Computational power/real-time communication lead to:
• lower information manipulation costs, lower transaction costs.
• automated search and negotiation tools.
• Standardized commodities & customized products => mass customization:
• information/digital products: personalized newspapers, online flexible subscription models.
• non-digital products: bundling of travel packages.
Negotiation configurations
Buyer
Buyer
Buyer
Seller
Seller
Seller
Buyer
Buyer
Buyer
Seller
Seller
Seller
Auction
Some configuration issues
• Scale up– # of agents, # of messages– vs. auction bottlenecks
• Distributed auction approach:Ygge, Power Load Management, ICMAS 96
• Security– Buyer/seller trust
• Andersson, Sandholm, Leveled Commitment Contracts, AAAI 1998
– Trusting the auctioneer• Franklin, Reiter, The Design and Implementation
of a Secure Auction Server, IEEE Trans. on Info Theory, 96
Negotiation Mechanisms: Auctions• What is an auction?
– Set of rules for determining price and/or allocation– Enforces a protocol
• McAfee & McMillian, Auctions and Bidding, JEL 87
• Auction framework provides structured, yet flexible market infrastructure which promotes automated negotiation:– Mediated vs. Unmediated
• Buyers do not have to separately find & contact every seller
– Price vs. Barter• Price minimizes communication between agents
– Formal vs. Informal• Standardized offers simplify communication between agents
Why Mediation?
• Manages communication, information
• Encapsulates negotiation rules• Source of constraint, structure• Enforcement• Not an agent: No discretion!
Mediatoragent
agent
agent agent
agent
Parametrized Auction Specification
English Procurement CDABidding RulesBidding Rules
Participation many:1 1:many many:manyBid Format single-unit single-unit stepwise demandOther Rules beat-quote none none
Information RevelationInformation RevelationInformation RevelationPrice Quotes ask price none bid-ask pricesQuote Schedule activity none activityOther Info none none transaction history
Clearing PolicyClearing PolicyPricing k = 1 k = 0 earlier bidClear Schedule inactivity fixed time activityClosing inactivity fixed time none
Internet Bandwidth• Smart markets in network bandwidth:
Varian, MacKie-Mason– Message packet includes a willingness-to-pay/bid– Network interface admits packets in descending
order of their bids, until congestion bound reached
– All packets priced at congestion cost -- the amount that highest denied service packet bid (Vickrey Auction)
• There is no reason for Messages not to honestly bid their willingness-to-pay: incentive compatible, efficient
Second Price (Vickrey) Auction
Case 0: You value the item at $3 and you bid $3
$3
$3.25
B. Item is sold for $3 to bidder whovalues it more than you do
$3
$2.5
$2.75
A. You get item the at $2.75
But...
• Does this mean you should really bid your true willingness-to-pay? Let’s suppose:– You value the item at $2, but you bid $3– You value the item at $3, but you bid $2
• What happens?
$3
$2
Second Price (Vickrey) Auction
Case 1: You value the item at $2, but you bid $3
$3
$2
$2.5
You might get item at $2.5, more than it’s worth to you
Second Price (Vickrey) Auction
Case 2: You value the item at $3, but you bid $2
$3
$2
$2.5
You might lose item to competitor for $2,when you could have had it for $2.5
Mechanism Design
• Market Allocation Mechanism: a communication process whereby dispersed knowledge is coordinated and used to determine a collective resource allocation.
• An allocation mechanism is defined by:– the interaction protocol (e.g. set of allowable
messages and protocol)• what kinds of bids/offers can agents make• what kinds of information about price quotes, other
bidders, etc. can agent’s request.– the rules that define the allocation outcome
• when does the final allocation get decided• what is the price/quantity allocation based on the
current set of bids
Mechanism Properties
• Incentive Compatible– No agent has anything to gain by departing from the
mechanism interaction rules.
• Pareto Optimality– No other allocation can make an agent better off
without making at least one other agent worse off.
• Privacy Preserving– Don’t need to have agents send entire set of
preferences and budgets to central planner.
• Individual Rationality– Agents benefit by participating
• Information Viability– Messages can’t be huge.
Combinatorial Auctions
• Combinatorial auctions– FCC plan to use a combinatorial auction
for their 3G spectrum auction in 2002– Combinatorial auctions are useful when
there are synergies involved buying different items.• For FCC: spectrum/BW packages, say across
different geographical regions, MHZ, QoS, etc.
• Ex: Suppose you have 3 goods: Northeast spectrum (NE), Middle Atlantic spectrum (MA), Southeast spectrum (SE)
– Possible bundles are NE, MA, SE, [NE, MA], [NE,SE], [MA, SE], [NE,MA,SE]
A Few Risks/Concerns• Possibility for strategic maneuvers
– Inherent “threshold” or “free rider” problems• Company A values NE license at $50, Company B values MA license at
$50, Company C values {NE, MA} package at $90. Company A and B can jointly outbid Company C, however one may end up paying more than its fair share. E.g., Company A bids $50, while Company B bids $40. Company B is a “free rider”.
– Defaulting/withdrawing bids• A single bid default in package bidding can affect the award of many
other licenses and be used strategically.• In spectrum auction: FCC Commission recommends stringent default
penalties.
• Added auction bidding complexity– E.g., FCC restricted set of possible spectrum packages
• Reduces complexity of both bidding and determining auction winners.• Also creates some (debated) concern over whether the limited choice
favors firms with certain kinds of business plans.
Related Research Questions• Auction Model
– How robust are these auctions against collusion/strategic manipulation among bidders?
– What information do bidders need to provide to the auction?
• How can we either minimize the information requirements and/or simplify them?
• Are there ways to simplify the number of combinations offered, based on networking knowledge and/or buyer/seller’s choice type constructs?
– How can computational efficiency be increased?
• Buyer/seller behavior and strategies– What kinds of strategies are required for buyers/sellers to
participate successfully? How complicated are these strategies? Can this be done using agents?
– How much information do the participants need to gather?
• Should there be auction stopping and pacing rules?
Computational Market Systems
• In a computational economy, markets coordinate the activities of individual agents each acting in their own self-interest.– Individual user preferences regarding goods
and services, as well as their quality and cost, are summarized and communicated via price.
• A computational market system is the:– Set of interaction protocols– Infrastructure services– System policiesthat implement a computational economy.
Distributed ResourceAllocation Problem
E-Commerce
Resource Allocation: Blue-Skies EconomyShould a mirror site be established on the LAN?
Internet Router 0 Site 1
Site 2
Carrier(0,1)
Carrier(I,0)
Carrier(0,2)
NetworkResources
Transport(I,1)
Transport(I,2)
Delivery(I,1)
Delivery(I,2)
Blue-Skies
Consumer@site 1
Consumer@site 2
Internet LAN
Steps in Designing an Economy• Define Goods
– Goods define the problem search space– Homogeneity vs. preserving important differences
• Define Producers– Maximize profits under given technology
• Carrier: Quadratic cost technology– E.g., x = y^2 + y + 1 (111 units input -> 10 units output)
• Transport: arbitrageur y = min(x1, x2)
• Define Consumers– Maximize utility subject to budget constraint– Pre-existing utility or model via econ framework
• Endowments material balance constraints• Utility func params found via competitive equil conditions
– Price = marginal cost– Marginal utility ratio = price ratio: MU1/MU2 = p1/p2
• Solve optimization problem via markets
UMDL Overview
• Provide library services in distributed network environment– Information agents buy and sell information
services
• Requirements: support dynamic, open, large-scale system– Distributed agent architecture, commerce
frameworkCollectionInterfaceAgents
MediatorsUser
InterfaceAgents
Information Sources Information Consumers
Information Goods and Services
• Bundling: subscription, per-issue, per-article• Timeliness: pre-publication, immediately, delayed• Terms: redistribute, read-only• To Whom: individual, library, group
Problem: Goods and services can vary across manydimensions in ways not determined at design time.
Magazine Dimensions:
Approach: Flexible supporting infrastructurebased on Ontologies and Auctions.
UMDL Commerce Infrastructure
Service Classifier Auction ManagerAuctionAuctionAuction
QuerySeller
User
1. How do Idescribe whatI want to sell?
1. How do Idescribe whatI want to buy?
2. Wheredo I go tosell it?
2. Wheredo I go tobuy it?
3. Match me witha buyer at a price
3. Match me witha seller at a price
4. Transactfor service
Market Management Services
• Market Matching– Automate finding service markets for agents
• Notification of new markets of interest to agents• Arbitrage between related markets for liquidity and to
keep prices in line
– Complex goods: selectable bundles
• Market Policy– Market creation and selection issues– Implemented via rules or incentives
• Data Collection and Information Dissemination– Data can be used to measure system welfare,
assess auction charging policy– Post market information for agents
Market Policies
• Support and uphold established business practices– Libraries, publishing, financial, business
• Endogeneous market creation/adaption policies market structure
• Account for system externalities such as market creation costs– Infrastructure costs– Agent decision complexity costs
Market Creation Policy
• Agents decide– Internalize costs to system via auction fees
• Auction Manager recommends– Supply defaults based on market policy,
service characteristics, current market configuration
– Testbed for evaluating different policies for market creation
• General question: are there system policies that result in better economic performance?
Computational Market System Summary• Demonstrate use of economic analytic
tools for system design• Design and implementation of generic
negotiation framework in UMDL• Identify and implement several market
management services– Market middleware: Auction Manager– Managing the scope of markets: policies,
matching, arbitrage
• Formal representation for describing service selection options/rules.
Internet Auctions and Agents
• How do we design economic mechanisms and agents to operate in an Internet economy?– What happens when humans and computational
agents participate in the same mechanism?• Can we design auctions that level the playing field
between humans and agents? (Do we want to?)– Design of auction to be “transparent”– Provide mediators to reduce information gathering
requirements– ??
– Policy rules for deploying/adapting mechanism• What kinds of tradeoffs between computational
efficiency, economic efficiency, profit maximization, fairness are necessary