Title: Reducing Costly Information Acquisition in Auctions Kate Larson, University of Waterloo
1Reducing Costly Information Acquisition in
AuctionsKate Larson, University of Waterloo
- Presented by David Thompson,
- University of British Columbia
- July 10, 2006
2Overview
- Deliberative Agents
- Auctions and Deliberative Bidders
- Optimal Search
- Larsons Auction
- Results
3Deliberative Agents
- Can deliberate (to gain information) as well as
bidding like a normal agent
4Deliberative Agents Properties
- R Resources dedicated to deliberation on each
possible problem - cost function mapping resource allocations to
cost in utility - A Algorithms provide solutions to problems
- PP Performance profiles describe how
allocating resources to an algorithm affect the
quality of solution it returns
5Deliberative Agents Anytime Algorithms
- All algorithms are assumed to have the anytime
property (similar to local search) - Can be stopped at anytime (or work with any
amount of resources) - Always return a solution
- Increasing time/resources always produces a
weakly better solution
6Auctions and Deliberative Bidders
- Agents pay deliberation costs
- Strategy space is expanded to include
deliberation actions (equilibria in this space
deliberation equilibria) - Agents may want to deliberate about each others
valuations (strategic deliberation)
7Auctions Desirable Properties
- Deliberation-proof agents have no incentive to
strategically deliberate - Non-misleading agents have no incentive to act
inconsistently with their valuation - Preference-formation independence auction
doesnt depend on cost functions, algorithms or
performance profiles - This combination is impossible (result from a
previous paper), drop preference-formation
independence
8Optimal Search
- An abstract problem from Operations Research
- n boxes, each with contents of different values
- fi(v), distribution over value of box i
- costi, cost of opening box i
- Agent gets to keep 1 box (after exploring)
9Optimal Search Solution
- Assign each box a cutoff value Ki, where agent is
indifferent to opening box i - Selection Rule open box with highest cut-off
value - Stopping Rule stop when the maximum observed
reward is greater than cutoff of all unopened
boxes
10Larsons Auction
- Using knowledge of agents algorithms and
performance profiles, calculate cutoffs for each
agent and order them - At stage t, the first t bidders participate in a
2nd price auction with a reserve price - Reserve prices are set to produce a
non-misleading Bayes-Nash equilibrium (acting as
a proxy for bidders t1..n)
11Larsons Auction Properties
- Non-misleading by reserve-price design
- Deliberation-proof
- Agents have no incentive to deliberate before
they can bid - Earlier agents have already demonstrated
unexpectedly low valuations (by not buying) - On expectation, later agents wont affect the
outcome (the auction will close)
12Experimental Results Efficiency(Uniform Costs)
13Experimental Results Efficiency(Informative
Costs)
14Experimental Results Cost of Deliberation vs.
2nd Price Auction (Uniform Costs)
15Experimental Results Cost of Deliberation vs.
2nd Price Auction (Informative Costs)
16Thank You.