self-enforcing strategic demand reduction
DESCRIPTION
Self-Enforcing Strategic Demand Reduction. Paul S. A. Reitsma 1 , Peter Stone 2 , J á nos A. Csirik 3 , Michael L. Littman 4 1 Brown University 2 U. Texas at Austin 3 AT&T Research 4 Stowe Research. Overview. Complex, high-stakes auctions - PowerPoint PPT PresentationTRANSCRIPT
Self-Enforcing Strategic Demand Reduction
Paul S. A. Reitsma1, Peter Stone2, János A. Csirik3, Michael L. Littman4
1Brown University 2U. Texas at Austin 3AT&T Research 4Stowe Research
00.511.522.533.544.55
Bill
ions
of d
olla
rs
Normal PRSDR
Bidding Strategy
Total Profit
Overview
• Complex, high-stakes auctions
• Complex, realistic simulations
• Highly effective strategy
• Robust, stable, simple
• Theoretical issues
Auctions Important
• Tiny toys to giant resources
• Commercial interest
• Theoretical interest– testbed for ideas
• Agents appearing in auctions
FCC Auction #35
• 422 licenses (spectrum blocks)
• 195 markets (major US cities)
• 80 bidders
• 101 rounds
• Dec 12 – Jan 26 2001
FCC Rules
• Theory: more information more efficient– all bids known– current winners known
• Bids: only 1 to 9 bid increments– 10% - 20% of current price
• Eligibility requirements• i.e., complex scenario
Auction Simulator
• FAucS
• Faithful to published rules
• Client-server architecture
• Runs auctions with agents and/or humans
FAucS Agents
• 5 important bidders– modeled individually– input from actual bidder team
• Other 75 served to raise prices– model as 5 secondary bidders– same role price floor 75%
Agent Goals
• Utility is profit
• Separate values per market– based on Merril Lynch data, real bidder input,
real auction analysis– per-agent
• Desire 0-2 licenses per market
• Assume no inter-market dependencies
Uncertain Knowledge
• Estimate other agents’ goals, budget– budget: within 20%– license valuations: within 20%
• per-license, per-agent
– desired licenses / market: 25% chance wrong• even one error can double perceived total desires
General Agent Strategy
• Each round:1. Get prices from server
2. Compute remaining budget, eligibility
3. Compute market values, costs
4. Choose desired licenses within constraints
5. Submit bids to server
Bidding Strategies
• Self-Only– knapsack approach effective
• Strategic Bidding (consider others)– threats– budget stretching– Strategic Demand Reduction (SDR)– explicit communication not allowed…
Randomized SDR
• Determine allocations dynamically– bid for desired licenses– tie-breaking creates allocation– respect allocation; no competition– ignore secondary bidders– don’t waste profit– great expected results
0
200
400
600
800
1000
1200
1400
Mill
ions
of d
olla
rs
1 2 3 4 5
Per-Agent Profit
Normal PRSDR
Luck
• Great expected results
• Random luck
• Unlucky winning little of desires– low satisfaction
• Incentive to defect– lowers expected profits
Fairing
• Unlucky bidder takes licenses until satisfaction near average
• Also bias compensation
• Equitable distribution– Yet, incentive to defect again!
Crime and Punishment
• Temptation to take too much– big profit gain– destroys fairness, destabilizes strategy
• Punish cheater to remove all profit gain– removes incentive– stabilizes strategy– Punishing RSDR
Detection
• Should take licenses only if:1. Low satisfaction rating
2. Punishing a cheater– i.e., focused
• Cheater takes when satisfied
• Cheater takes indiscriminately
• Flawless detection
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
Pro
fit
vs. b
asel
ine
Missed Caught
Detection Results
Enforcement EffectsDefector PRSDR
Enforcement Effects
• Large win for uncaught cheater
• All extra profit lost when cheater caught– strong disincentive
• Slight enforcement cost– raises expected profit by dissuading cheating– less aggressive punishment scheme possible– people willing to pay to punish cheaters
Alternative Scenarios
• Change price floor
• PRSDR preserves profit nearly optimally– larger profit margin larger absolute and
relative profit from PRSDR
• Large numbers of defectors– drop back to all-Knapsack without loss
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
Fra
ctio
n w
aste
d
85% 75% 50%
Price Floor
Wasted Profit
Knapsack PRSDR
Algorithm Overview
1. Bid on desired licenses
2. Tie-breaking creates allocation
3. No competition
4. Fairing balance
5. Auto-punish defectors
6. Punishment removes defection incentive
Improved Auction Design
• Information sources:1. via low prices
2. from auctioneer
• Traditionally, more info greater efficiency• However, more info more strategies
– PRSDR hard to thwart
– less efficiency?
– tradeoffs in auction design
Game Theory
• Analyze as 3-option Prisoner’s Dilemma:1. Cooperate (RSDR)
2. Hedge (PRSDR)
3. Defect (Knapsack)
• Pure Nash equilibrium
• Suggestive, not conclusive, for auction
3 3 0
3 3 2
5 1 1
Real-World Application
• Relies on few assumptions:1. Bidders desire maximum profit
2. Bidders know of PRSDR, benefits
3. Bidders willing to try, risk-free
4. Information available
Conclusions
• Effective• Realistic
– related real strategies– safe to try
• Stable– self-enforcing
• Robust• Fair
Questions?