Auction Design for Atypical Situations Overview General review of common auctions Auction design for agents with hard valuation problems Auction design for goods in unlimited supply Auction Design Many different protocols Major auction types: Ascending Price (English) Descending Price (Dutch) First Price, Sealed Bid Second Price, Sealed Bid (Vickrey) Essentially mixing and matching certain properties, but some combinations work better than others Auction Evaluation Revenue for the sellers Profit for the bidders Avoidance of “winner’s curse” helps both Winner in most auction protocols is the participant who made the biggest overvaluation mistake Auctions that decouple an agent’s bid from the actual price paid encourage higher bids Ascending Price (English) Most commonly known protocol Auctioneer starts with an opening bid and successively raises the price as participants are willing Allows for dynamic adjustment of bidders’ valuations by giving information about other bidders Buyers bid at most their utility, which may be adjusted on the fly Descending Price (Dutch) Biddings starts at an extremely high price and descends until someone claims the item High degree of Winner’s Curse Intuitively, raises seller’s revenue in cases when the high bidder wants an item badly First Price, Sealed Bid Buyers submit bids once No knowledge of one another’s bids Auctioneer opens bids and sells the item to the highest bidder at the price he submitted Encourages buyers to bid conservatively (shade down from utility) to maximize profit versus probability of winning Second Price, Sealed Bid (Vickrey) Designed to alleviate Winner’s Curse in the First Price, Sealed Bid protocol Same sealed bid format, but item is awarded to the highest bidder at the second highest price Buyers can bid their utility to increase probability of winning, but are guaranteed a price closer to market consensus Optimal Auction Design for Agents with Hard Valuation Problems David C Parkes Motivation Standard auction theory assumes that either agents know own their utility for an item (private value) or that there is a common utility that is unknown to the agents (common value) As transactions are increasingly turned over to software agents, the cost of obtaining this utility value may be significant Certain auction designs can simplify this problem Paper Goals Compare the performance of agents with hard valuation problems within three auction designs Posted Price, sequential Second Price, sealed bid Ascending Price What the paper isn’t about: Actual methods for refining beliefs about values Examples of Hard Valuation Agents bidding for components on behalf of a manufacturer Agents bidding for collectibles or other rarities Problem Formulation Agents with hard valuation problems operate in three phases: Metadeliberation: Decide how much effort to spend refining the valuation of the item Valuation: The refinement process – solve an optimization problem, interact with a human expert, etc Bidding Agent Model Each agent has an unknown true value for an item Each agent maintains an upper and lower bound, in which the true value is assumed to be somewhere, uniformly distributed, in between Expected true value is then the average of the upper and lower bounds The deliberation process refines the upper and lower bounds Agent Model Upper Bound D New Upper Bound Expected True Value New Expected True Value aD New Lower Bound Lower Bound Deliberation Incurs cost C Parameters Cost of Deliberation (C) Computational Effectiveness of Deliberation (1a) Metadeliberation Solve the tradeoff between reducing uncertainty and incurring the cost of deliberation Deliberation is only worthwhile when: It changes the agent’s bid The new bid has a greater expected utility than the old bid Metadeliberation strategies vary by auction type Metadeliberation Strategies Vickrey Auction Need distributional information about the bids of other agents Paper assumes such information is somehow obtained by the agents (eg learning) Metadeliberation strategy: Deliberate until utility of placing a bid now is greater than the estimated utility of placing a bid after another deliberation step Bid utility is a nonlinear function of expected value Higher bid decreases profit but raises probability of winning Metadeliberation Strategies Vickrey Auction Length of time that an agent spends deliberating depends on: The number of agents in the auction The agent’s current upper and lower bounds on the value of the item The computational effectiveness of deliberation (1a) The cost of deliberation (C) Metadeliberation Strategies Posted Price Sequential No uncertainty about the actions of other agents Need only to worry about the cost of the good Deliberate only when the ask price is within the bounds ofv some threshold function g*(a, C, D) Reject Price g*D v Deliberate Accept Price Metadeliberation Strategy Ascending Price Third action available in addition to bid and deliberate: wait Agents that wait benefit from the deliberation of others Optimal Strategy: v Leave auction Wait, or Deliberate if auction will close, with probability 1/(Na -1) v Bid Evaluation Metrics Efficiency: True Value for the Good for the Winning Agent / Maximum True Value over All Agents Revenue: Price Paid for the Good / Maximum True Value over All Agents Utility of Participation: (Surplus to Winning Agent – Total Deliberation Cost for All Agents) / Number of Agents Evaluation Arms Variance of: Number of Agents Computational Effectiveness (1-a) Cost of Deliberation (C) Agent “experience” – Adjust C and a for fractions of the agent population Varying the Number of Agents + Ascending Price X Sealed Bid O Posted Price Sequential Varying the Bidding Increment + Ascending Price Varying the Computational Effectiveness of Deliberation + Ascending Price X Sealed Bid O Posted Price Sequential Varying Agent Experience + Ascending Price X Sealed Bid O Posted Price Sequential Competitive Auctions and Digital Goods Andrew Goldberg Jason D Hartline Andrew Wright Motivation Looking for an optimal way to sell goods in unlimited supply Downloadable music Pay per view movies Examines auctions as an alternative to fixed pricing, which requires expensive and probably inaccurate market research Scary implication: Charge more for media created by entities with small, rabid followings? The Optimal Threshold Function Given a set of bids, determine the single price that maximizes revenue 3 3 3 2 2 Sell… • 3 at 3 (9) • 5 at 2 (10) • 6 at 1 (6) 1 Paper Goals Examine classes of single round, sealed bid auctions for products with no marginal cost of reproduction Need to solve tradeoff between selling a lot at a low price versus a few at a high price Need to ensure that participants bid their utilities Would like an auction mechanism that compares well to optimal fixed pricing Terminology Truthful auctions Encourage participants to bid their utility More formally: Bidder’s profit (bid – price if wins, or 0 otherwise) is maximized when bid is the same as utility for any fixed values for the bids of other participants Example: Vickrey Counterexample: First price sealed bid Why is this important? Revenue is maximized in truthful auctions Truthful Auction Example Imagine participating in an auction for a jar of 100 pennies that you (and only you) have counted In a first price sealed bid auction, you cannot hope to profit by bidding 100 In a Vickrey auction, by bidding 100 you will at worst break even, but most likely pay less Terminology Competitive auctions Produce revenue within a constant factor of optimal fixed pricing Must vary the number of items sold based on the bids received Why is this important? Matching optimal fixed pricing is the best possible result Being within a constant factor is a reasonable approximation Evaluated Auction Designs All auctions in this paper are single round, sealed bid Auction mechanisms studied: Deterministic Deterministic Optimal Threshold Randomized Single Price Dual Price Weighted Pairing Bid Independence Agent’s bid determines whether or not he wins the auction, but not the price Typically multi-price, but not always Example: Vickrey Why is this important? Bid independence allows for one to bid her utility and still hope for a profit Deterministic auctions must be bid independent in order to be truthful Truthful Deterministic Auctions Consider the deterministic optimal threshold auction To determine if bid bi wins Remove bi from the set of bids (ensure bid independence/truthfulness) Determine the threshold price at which maximal revenue is attained in the remaining set of bids If bi >= this threshold, accept bi at the threshold price Note that this removal of bi is the only thing that differentiates this auction from optimal fixed pricing Deterministic Optimal Threshold Auction Example 5 3 3 3 2 2 Step 1: Remove the bid to be evaluated: 3 1 Step 2: Compute the optimal threshold on the remaining bids: 5 3 3 2 2 1 Sell 1 at 5: 5 3 at 3: 9 5 at 2: 10 6 at 1: 6 Optimal threshold is 2 Step 3: Compare the removed bid to the optimal threshold 3 Step 4: Accept bid 3 at price 2 >2 Truthful Deterministic Auctions Theoretical Results Truthful Deterministic Bid-Independent Auctions are not Competitive in the worst case Removing bi causes bigger problems than one would expect Consider an input set where the high bid h occurs r times, and there are (h – 1) (r – 1) other bids at 1 Proof Sketch Example: h = 5, r = 3 5 5 5 1 1 1 1 1 Sell… • 11 at price 1 (11 revenue) • 3 at price 5 (15 revenue) 1 1 1 Proof Sketch, Continued Deterministic Optimal Threshold Auction on this input: 5 5 5 1 1 1 1 1 1 1 1 To determine if b1 wins, remove it and compute the threshold on the rest of the input 5 5 5 1 1 1 1 1 1 1 1 Sell 2 units at 5, or 10 units at 1 Threshold falls to 1, so end up selling to the high bids at the low price More Results This result generalizes to all Deterministic Auctions due to the theorem that all Truthful Deterministic Auctions are Bid Independent Proof intuition: Can’t profit by bidding utility if you’ll pay it upon winning Paper goes on to present empirical evidence that on realistic data, the worst case for these auctions doesn’t come up often Random Sampling Auctions Motivation Auction mechanisms need not only be resistant to bad inputs, but also to attack By using some nondeterminism in deciding who wins and at what price, we can avoid dominance by worst case inputs Random Sampling Auctions Choose a random sample from the set of bids Run the optimal threshold function on the sample and use the result on the bids not in the sample Single price Nondeterministic Dual price variant: Choose roughly half of the bids for the sample Calculate thresholds on both sets Use the threshold from one set on the other, and vice versa Avoids having to throw out bids from the sample Random Sampling Example 5 3 3 3 2 2 1 Step 1: Choose subset at random: 3 2 2 Step 2: Compute optimal threshold: 3 2 2 Sell 1 at 3: 3 Sell 3 at 2: 6 Optimal threshold is 2 Step 3: Apply optimal threshold to those not in the sample: 5 3 3 1 Step 4: Accept 5 3 3 at price 2 Random Sampling Auctions Theoretical Results Random sampling auctions are competitive General flavor of the proof: Revenue generated is within a constant factor of optimal fixed pricing with arbitrarily high (but not 1.0) probability The higher the probability, the lower the constant The chosen subset is a good representation of the whole most of the time The revenue lost from losing the sampled bids is a constant factor of the whole The dual price version performs even better Weighted Pairing Auctions To determine if a particular bidder i wins with bid bi: Choose another bid b with probability proportional to its value –higher bids get picked more often Compare b with bi. If bi >= b, bidder i wins and pays b Multi-price Nondeterministic High bidders likely to win and pay high prices, but some low bidders sneak in as well Weighted Pairing Example 5 3 3 3 2 2 Step 1: Choose a bid 3 Step 2: Choose another bid with probability proportional to its value 2 Step 3: Compare 3 Step 4: Bid 3 wins and pays price 2 1 > 2 Weighted Pairing Auctions Theoretical Results Weighted Pairing auctions are not quite competitive Within a logarithmic factor of fixed pricing, so not bad either Perform well on inputs that random sampling does not