Jan 16, 2006 Auctions in SCM 1 Auctions in SCM •Auction descriptions •Decision-theoretic approach •Collusive and no-collusive game-theoretic approach
Mar 31, 2015
Jan 16, 2006 Auctions in SCM 1
Auctions in SCM
•Auction descriptions•Decision-theoretic approach•Collusive and no-collusive game-theoretic approach
Jan 16, 2006 Auctions in SCM 2
Trading Agent Competition: Supply Chain Management game
• 6 software agents compete to run profitable PC assembly business for 220 days– Bidding for components from 8 suppliers– Bidding for orders from 100’s of customers
(simultaneously)– Managing production & delivery planning
Jan 16, 2006 Auctions in SCM 3
Auction Descriptions
•Cheat-sheets
•Agent valuations
•Analogs
Jan 16, 2006 Auctions in SCM 4
Supplier auction Periodic-clear, multi-unit, bizzaro-
price auction
• Bid structure: component type, quantity, date and reserve price
• Information: 20-day summary, offer price, date & quantity
• Clearing rules: ...• The good: Offers of components of type t,
satisfying reserve price, satisfying one or both of date and quantity
Jan 16, 2006 Auctions in SCM 5
Supplier pricing
• i = days in advance• d = date• i*Cac
d = naïve estimate of supply• Cavl' = supply - demand (not counting RFQs from
agents with less reputation) up to date d+i• delta = 0.5
Jan 16, 2006 Auctions in SCM 6
Reputation
• Reputation is ordered : offered ratio.
• Includes a prior of 2000 ordered & offered + 100 per day.
• Renormalized on the range [0,0.9] for IMD & Pintel, [0,0.45] for all others, but limited to 1.
• For each RFQ, offered = max (smallest offer, ordered, 0.2*requested )
Jan 16, 2006 Auctions in SCM 7
Customer auctionSimultaneous/sequential, reverse,
single-unit, 1st-price auction• The good (fully disclosed): Right to sell 1-20 computers
of specified type for bid price, or be penalized 5-15% of the reserve price every day late to a maximum of 5 days (after which the order is cancelled)
• Bid structure: price, date and quantity• Clearing rules: Lowest price that satisfies date, quantity
& reserve price• Information: Whether or not you win, max and min
winning bid over auctions for that type of PC, 20 day summary
Jan 16, 2006 Auctions in SCM 8
20-day summary
• Suppliers– Total ordered/shipped for each class (eg,
CPU)– Mean production capacity for each class– Mean price for each component
• Customers– Total requested/ordered for each SKU– Mean price for each SKU
Jan 16, 2006 Auctions in SCM 9
Valuation and exposure definitions
• Super-additive valuation: The value of a set of goods is greater than the sum of the values of those goods
• Sub-additive valuation: The value of a set of goods is less than the sum of the values of those goods
• Exposure: The risk of winning some sub-optimal set of goods
Jan 16, 2006 Auctions in SCM 10
Exposure in SCM
• In isolation, every good has negative value:– Components cost money to store– Customers charge for missed shipments
• Super-additive: Only "matched sets" (components and orders) can turn profit
• Sub-additive: Too many matched sets will overwhelm production capacity and cause loss
Jan 16, 2006 Auctions in SCM 11
Analogs
• Compare supplier auctions and 2nd price– Payment independent of reserve price– Reserve price is a bound on payment– Probability of winning increases monotonically
with reserve price– Therefore, Dominant Strategy Truthful? (At
least for the reserve price)
Jan 16, 2006 Auctions in SCM 12
Decision theoretic approach to customers
• Customer side seems to come down to conditional distribution modeling– P ( winning | bid price , state)– State includes auction parameters and known
facts about the world (eg, recent prices, 20-day reports)
• Then bid to maximize valuation– Naïve P()=1 approach– Expectation approach
Jan 16, 2006 Auctions in SCM 13
Decision theoretic approach to suppliers
• Reserve price is DS Truthful?
• Large quantities bids can be risky to reputation
• Effect of local price fluctuations is exaggerated
Jan 16, 2006 Auctions in SCM 14
Non-Collusive Approaches
•Disrupting markets
•Disrupting agents
•Risk-attitudes
Jan 16, 2006 Auctions in SCM 15
Disrupting markets
• Increasing demand in supplier markets– Limited scope: doesn’t affect customer
reserve price (or late penalty) or storage costs
• High- or low-balling customer auctions (exploiting other agents exposure risks)
Jan 16, 2006 Auctions in SCM 16
Disrupting agents algorithms
• Crashing unreliable agents (making the impossible happen)
• Preventing convergence (adding noise to the available information)
• Exploiting simplistic models (oscillating strategies)
• What about between-game learning (human and machine)?
Jan 16, 2006 Auctions in SCM 17
Reconsidering reputation
• Agents can compute their own reputations exactly and can deliberately “manage” them
• Reputation has monotonic, but “relative” value (best can’t improve, worst can’t get worse)
• An opportunity for active learning?
Jan 16, 2006 Auctions in SCM 18
Externalities & Risk-attitudes
• Outcome space can be profitability or it can be the final (relative) rankings– Agent has negative externalities against
another’s benefit
• Utility from ranking– Implies actual value of money is a sum of step
functions– Observation of rankings will be very noisy
Jan 16, 2006 Auctions in SCM 19
Collusive Approaches
•FCC Spectrum Auction–Applications to TAC
•Set-your-hair-on-fire Collusion
Jan 16, 2006 Auctions in SCM 20
FCC spectrum auctionSimultaneous, single-good, English
auctions• The good: Exclusive rights to broadcast on
a given frequency range in a given US city
• Bid structure: “Real-value” jump bids are allowed
• Information rules: Bidders are not anonymous
• Clearing rules: Auctions all clear at once
Jan 16, 2006 Auctions in SCM 21
How to collude in the FCC auction
• Agents decide on a distribution "outside" of the mechanism
• Defection is punished by threat or retaliation bids on multiple goods held by the defector
• Communication through the identity of the bidder and possibly the timing or value of the bid (nothing else needed)
Jan 16, 2006 Auctions in SCM 22
Cost/benefit of collusion
• Cost: “Freedom” to bid on any auction you wish
• Benefit: “Protection” from the full costs of market competition
Jan 16, 2006 Auctions in SCM 23
Bad news: prisoner’s dilemma
Jan 16, 2006 Auctions in SCM 24
Outcome of collusion
• Robust: cartel was a small subset of actual bidders (6 out of 153)
• Profitable: member of the cartel paid significantly less ($2.50/person vs. $4.34/person) for more (476 out of 1479)
Jan 16, 2006 Auctions in SCM 25
Differences from SCM
• Auctions are sealed bid
• Winner and winning bid aren't announced
• Auctions close periodically
Jan 16, 2006 Auctions in SCM 26
Communication options for SCM
• Outside channels: "In poor taste"
• Supplier auctions: Difficult and expensive
• Customer auctions: The min-bid for a type of PC– Cost to send– Very limited bandwidth– Shared bandwidth
Jan 16, 2006 Auctions in SCM 27
Enforceability in SCM
• Defector-detection is difficult because of anonymity and sealed-bid– Probabilistic inference?
• Is enforcement necessary?– Not if the cartel all represent one entity– Multiple agents can advance and cartels
would benefit from advancing collectively
Jan 16, 2006 Auctions in SCM 28
Set-your-hair-on-fire collusion
• Ignoring cost, an agent can disrupt supplier and customer markets indefinitely
• Agents that can anticipate (or request) disruptions have a significant advantage
Jan 16, 2006 Auctions in SCM 29
What if Collusion is illegal?
• Set-their-hair-on-fire collusion!– Our martyr agent tries to help another team
instead
Jan 16, 2006 Auctions in SCM 30
Thanks!