Title: Repeated Auction Games and Learning Dynamics in Electronic Logistics Marketplaces: Regulation through Information
1Repeated Auction Games and Learning Dynamics in
Electronic Logistics Marketplaces Regulation
through Information
Hani S. Mahmassani University of
Maryland Potentials of Complexity Science for
Business, Governments and the Media Collegium
Budapest, August 3-5, 2006
2Repeated Auction Games and Learning Dynamics in
Electronic Logistics Marketplaces Regulation
through Information
Hani S. Mahmassani University of
Maryland Potentials of Complexity Science for
Business, Governments and the Media Collegium
Budapest, August 3-5, 2006
3Once upon a time, there was a physical world
4(No Transcript)
5Motivation
- Developments in Information and Communication
Technologies are - Transforming Supply Chain Operations
- Enhancing transportation service levels and
optimizing its performance - Introducing new ways of meeting supply (capacity)
and demand (shippers) - Freight Matching
- Transportation Auctions
- Supply Chain Integration Tools
6Repeated Auction Games and Learning Dynamics in
Electronic Logistics Marketplaces Regulation
through Information
7Presentation Outline
- Motivation
- Characteristics of Transportation Auctions
- Problem Definition
- Methodology
- Results
- Ongoing and Future Research
8Collaborative research withMiguel Figliozzi
(former PhD Student at Maryland, now Asst.
Professor at University of Sydney) Patrick
Jaillet (MIT)
9 Range of Shipper/Carrier Procurement Structures
Vertical Integration
Long Term Contracts 3PL Services
Private Exchanges
Spot Market Brokers/Public Exchange
Private Fleet
Core Carriers
Any Carrier/Shipper
Long Term Relationships
Number of Participants
Control, Collaboration, Reliability
Savings from Better Routing Economies of
Scale/Scope
Savings from Collaboration Customized Services
10- Illustrative Benefits of
- Market-based Procurement
- Wealth Generation
11Comparing Transportation Market Environments
- Common Characteristics
- Stochastic arrival of shipments
- Hard Time Windows
- Simulate market under
- Different arrival rates (low to high)
- Different Time Windows Widths (short to long)
- Truth revealing second price auction
- Vertical Integration
- Assignment to fleets
- To own fleet
- One shipment at a time
- In real time
- In order of arrival
- Spot Market
- Assignment to fleets
- To best bidder
- One shipment at a time
- In real time
- In order of arrival
- (zero probability of bidding on two shipments)
Performance Total Wealth Generated, Shipments
Served, System Empty Distance
12Vertical Integration vs. Spot Market
4 shippers and 4 Carriers
Wealth Generated Change ( increase)
13Vertical Integration vs. Spot Market
(4 shippers and 4 Carriers Shipment Served)
14Vertical Integration vs. Spot Market
(4 shippers and 4 Carriers Empty Distance
Change)
15Dynamic Pricing in a Sequential Auction
Marketplace...
- The market generates a sequence of auctions
- Prices are generated as
- The outcome of carrier bids
- Predefined set of rules (auction rules)
- A Carriers behavior is expressed through his/her
bids
16Auction Marketplace is a useful laboratory to
gain insight into
- Carrier behavior
- Learning and Adaptation
- Effectiveness of Competitive strategies
- Impact of Information Availability on System
Performance
17Presentation Outline
- Motivation
- Characteristics of Transportation Auctions
- Problem Definition
- Methodology
- Results
- Ongoing and Future Research
18What are the characteristics of transportation
auctions?
- The traded entity is a service
- Transportation services are perishable,
non-storable commodities - Demand and supply are geographically dispersed
but they exchange real time information online
19What are the characteristics of transportation
auctions?
- Group Effect value of traded item (shipment) may
be strongly dependent upon the acquisition of
other items (e.g. nearby shipments) - Network Effect value of a shipment is related to
the current spatial and temporal deployment of
the fleet. - Uncertainty
- demand/supply over time and space
- prices
20Sources of Complexity
- Multiple interacting agents with multiple
conflicting objectives - Modeling agents bounded rationality and learning
- Demand spatial and temporal stochasticity
- Uncertainties about a shipment value and cost
- Fleet management complexities (routing, time
windows, penalties, etc.) - New class of problems created by different market
designs and levels of information
21Presentation Outline
- Motivation
- Characteristics of Transportation Auctions
- Problem Definition
- Research Methodology
- Results
- Ongoing and Future Research
22Problem Context
- Sequential Auction Market Environment
- Stochastic arrival of non-identical shipments
- Sequential auction of arriving shipments
- Bidding is done one shipment at a time, in order
of arrival - Carriers objective is to maximize expected
profits while managing the fleet to satisfy
service quality constraints (time windows) - The principal operating costs are proportional
to the shipment haul-length and the empty
distance required - Carrier costs and capacity affected by history of
bidding and shipment assignment decisions
23Problem Categorization
- Two Layers of Allocations
- Auction Shippers ? Bidders
- Pricing of resources
- Profit Maximization problem
- Strategic Problem
- Fleet Management Shipments ? Trucks
- Allocation of own resources
- Cost Minimization problem
- Non-strategic problem
- The joint bidding/fleet management problem is
highly complex
24Game Formulation
- A sequential auction as a dynamic game of
imperfect information - dynamic carriers face each other at different
stages - imperfect information carriers are uncertain
about competitors private information (what
affects competitors shipment cost) - Stages are identified with shipment arrival
epochs - A carrier has full knowledge about his fleet
status (vehicles and shipments) and technology - A carrier has uncertainty about competitors
fleet status, technology, or bidding function
25Finding a Bidding Policy
- In auctions, profits are highly dependent on the
quality of the bidding policy. - A bidding policy is a function produces a bid
value using information about - the state of the carrier,
- the characteristics of the shipment for auction,
- the marginal cost of serving the shipment,
- auction type, and beliefs about the competitors
and environment
26Finding a Bidding Policy
- Problem complexity generally precludes finding a
policy that optimizes the entire
auction/assignment problem. - Each auction provides opportunity for carriers to
learn about - The environment
- Other players strategies
- Learning potential is dependent on information
disclosed after each auction
27Information Levels
- The information revealed after each auction can
influence the nature and rate of the process by
which carriers learn about the game and their
competitors behavior. - Information includes
- Actions (bids) placed.
- Number of players (carriers) participating
- Links (name) between carriers and bids
- Individual characteristics of carriers (e.g.
fleet size) - Payoffs received
- Knowledge about who knows what, information
asymmetries, or shared knowledge about previous
items.
28Information Levels
- Define TWO information levels
- maximum information environment, all the above
information is revealed. - minimum information environment where no
information is revealed. - These two extremes approximate two realistic
situations - Maximum information would correspond to a real
time internet auction where all auction
information is accessed by participants. - Minimum information would correspond to a shipper
telephoning carriers for a quote, and calling
back only the selected carrier.
29Learning in Different Environments
- Minimum information setting
- Genetic Algorithms
- Reinforcement Learning
- Maximum information setting
- Fictitious Play
- Machine Learning
- Rationalizable and Machine Learning
- Rule Learning
- Rules of Thumb Learning (e.g. Tit for Tat)
30Carriers Decisions
- Strategic decisions investment of resources, for
the purpose of learning about or influencing
competitors, to improve profits by manipulating
future auction outcomes. - Identifying decisions are characterized by
attempts to identify or discover a competitors
behavior. - Signals are decisions that aim to establish a
reputation or status for the carrier. - Operating decisions decisions that are not
strategic but aim to improve a carriers profit
level (e.g. rerouting of the fleet after a
successful bid)
31Bounded Rationality
- Carriers can analyze with different degrees of
sophistication (bounded rationality) the history
of play and estimate the possible future
consequences of current actions. - Research in the area of learning in games is
actively seeking to explain how agents acquire,
process, evaluate or search for information.
32Bounded Rationality
- Cognitive and computational limitations can be
evidenced in - Identification the carrier has limited ability
to discover competitors behavioral types, which
may require complex econometric techniques - Signaling limited ability to read or send
signals that convey a reputation - Memory limited ability to record and keep past
outcome information or memory to simulate all
future possible paths in the decision tree - Optimization even if carriers could identify
competitors behavior, their ability to formulate
and solve stochastic optimization problems is
likely limited.
33Presentation Outline
- Introduction
- Characteristics of Transportation Auctions
- Problem Definition
- Research Methodology
- Results
- Ongoing and Future Research
34Case Study Myopic Carrier Learning in Second
Price Auctions
- Study the impact of different learning techniques
on - Carriers
- Profits
- Market share
- Under different market settings
- Minimum Information
- Maximum Information
- Shippers
- Consumer Surplus
- Number Shipments Served
- Low Arrival Rate (uncongested)
- High Arrival Rate (congested)
35Research Methodology
- Define Auction Type Second Price Auction
- DEFINITION (reverse auction)
- Carrier with lowest bid wins item
- Winner gets paid second lowest bid
- Rest of bidders do not pay or receive anything
- PROPERTIES (one shot auction - Vickrey 1961)
- Equilibrium strategies are truth-revealing and
dominant strategies bid true marginal cost - They do not require gathering or analysis of
information about the competitors situation - Leads to complete economic efficiency, the bidder
with the lowest cost wins
36Problems with Second Price Auctions
- Rothkopf, Teisberg, and Kahn (1990, JPE)
- Auctioneer may cheat in the auction
- Vulnerability to bidder collusion
- Revelation of private information
- Sandholm (2000, IJEC)
- Complexity of bidding looking at future sequence
of arrivals introduce speculation about future
bids of other bidders - Untruthful bidding with risk averse bidders
- Marginal Cost Bidding (Krishna, 2002)
- Not necessarily Optimal in sequential Auctions
- No known equilibrium for sequential auctions
where bidders have multi-unit demand curves
37Finding an Optimal policy...
- Limited to myopic bidding
- Will not consider impact of bidding on
competitors future behavior - Will not consider impact of bidding on future
service costs - Limited to finding the best constant marginal
cost factor c such that - bid c x marginal cost
38Finding an Optimal policy...
- Compare Learning Strategies
- Reinforcement Learning
- Tit for Tat
- And
- Marginal Cost Bidding ( c 1)
39Reinforcement Learning
- An agent chooses an action with a probability
that is directly proportional to the profit that
such action has achieved in the past - Initially the agent starts with positive uniform
profits over all possible marginal cost factors - As each action (bid) is played, the agent updates
the profit level with the payoff obtained - Over time, profit levels will converge (if
facing a stationary environment) - This learning method can be utilized under
minimum or maximum information settings
go to referring slide
40Reinforcement Learning (ctd.)
- A a1, ... ,an set of available actions (bids)
- r(ai) average reward obtained using action ai
in the past (includes both won and lost bids) - P (ai) probability of playing action ai
- P (ai) r(ai) / Si r(ai) i?A
go to referring slide
41Tit for Tat
- Tit for Tat is more an adaptive rule of thumb
than a learning mechanism - It is a robust strategy in many strategic
situations - This carrier roughly imitates what his opponent
is doing - With two carriers, A and T, where carrier T plays
Tit for Tat, this rule of thumb can be defined
as - Carrier T computes the average bid value of
carrier A over the last T auctions, called â - Carrier T computes his own marginal cost average
over the last T auctions, called c - Carrier T obtains a â / c,
- Ts next bid will be equal to bid a x
marginal cost
42Behavioral Assumptions and Rules
- Carriers
- Non-cooperative carriers
- Preference over game outcomes with highest
expected return, risk neutral - Myopic strategies
- Shippers
- Shipper selects carrier with lowest bid
- Shipper does not cheat
43Other Market Settings
- 2 Carriers
- Geographic Area 1 1 square space
- Shipment Origin and Destination ? Uniformly
distributed over space - Earliest Pick Up Time arrival time
- Latest Pick Up Time arrival time Time window
length (2 units of time uniform0,2 ) - Fleet size 12 vehicles (constant) serving the
market - The reservation price of the buyer (shipper) is
distributed uniform 1.4,1.5 - ? 0.25 arrivals/unit time/truck (not congested)
- ? 1.00 arrivals/unit time/truck (congested)
- Results obtained with 10 iterations of 10,000
arrivals each
44Presentation Outline
- Introduction
- Characteristics of Transportation Auctions
- Problem Definition
- Methodology
- Results
- Ongoing and Future Research
45Optimality of Bidding Marginal Cost with Low
Arrival Rates (AR3)
- Carrier MC bids always marginal cost
- Carrier D bids marginal cost multiplied by c
- Highest profits when c 1
46Reinforcement Learning
- Carriers discover that c 1 provides the highest
profits among all possible factors
47Learning is not Free
- Comparing Shipper Surplus and Rejected Shipments
between carriers playing Reinforcement Learning
(RL) and carriers bidding Marginal Cost (MC)
48Learning is not competitive...
- Comparing Shipments Won and Profits when a
- RL carrier competes against a MC (c1) carrier
49Tit for Tat is a Robust Strategy
- Maximum Information Setting
- Tit for Tat competing with a RL and MC carriers
(profits) - Tit for Tat successfully imitates competitor
50Too much information could be a problem...
- If the leader becomes aware of his leadership,
it is a dominating strategy to rise prices - Graphs compare profits for MC carrier and Tit for
Tat carrier when the leader goes from c1 to c2
51Conclusions
- Even with minimum information, Learning is
possible (e.g. convergence towards mc bidding) - Learning is expensive for both carriers and
shippers - Carriers suffer hard against competitors that
have already found optimal policies - Shippers do pay for learning when all carriers
are learning - Higher prices
- Fewer shipments served
- Therefore, market setting should be such that
learning duration is minimized
52Conclusions
- Maximum information settings allow a large array
of possible new behaviors - Tit for Tat or imitation is possible, typical
market with a leader and a follower - From a carrier point of view, Tit for Tat is
robust - From shippers point of view, Tit for Tat is
good as long as the leader follows a
competitive bidding policy - Problem with too much information
- If the leader becomes aware of his leadership,
it is a dominating strategy to raise prices - The follower gladly follows suit since his
profits also rise
53- Incentive Compatibility Issues
54Incentive Compatibility
- Truth revealing second price auction market has a
wealth creation potential - Can a truth revealing market be sustained?
- Experiment compare the performance of a truth
revealing carrier and a NON-truth revealing
carrier - One carrier uses a bidding factor ?1
- The other carrier bids his marginal cost
55Ongoing and Future Research
- Develop more sophisticated strategies that do
take into account future consequences of current
bid on - Players own future costs
- Players own future revenues
- Develop strategies that use marketplace
information to identify competitors behavior and
manipulate future outcomes - Study how market performance is affected by
varying - number of carriers
- auction mechanisms
- Information disclosed
- Develop discrete choice models of competitor
behavior under different observation and
information conditions
56 57Game Formulation
- The Truck-Load Procurement Market (TLPM)
formulation differs from other auction
formulations in several respects - (a) items auctioned (shipments) are
multi-attributed - (b) costs are functions of carriers status and
vehicle routing technologies - (c) history and fleet management decisions affect
future cost probability distributions. - (d) capacity constraints are linked to private
information and shipment characteristics - (e) bidding strategies are dependent on public
and private history - (f) timing of auctions is important
- (e) it is an online sequential auction.
58Notation
carriers competing, each carrier
(The set of carriers)
the set of auction announcement epochs is
the set of arriving shipments is
arrives
represents the time when shipment
(set of shipments arriving after )
each carrier simultaneously bids a monetary amount
.
59Notation
public information generated after auction for
information publicly known before bidding for
shipment
private information for each carrier at time
carrier status before bidding (shipments fleet)
assignment function
cost function
conditional probability about opponents type
60Notation
bidding function
competitors bidding functions
auction assignment function
probabilities of winning shipment
auction expected payment function
expected profit for shipment
61Online Equilibrium
probability space of arrivals and shipment
characteristics
Bidding Functions Equilibrium
62Relaxing Rationality
- In a bounded rational model, a carrier faces two
basic types of uncertainties regarding the
competition - an uncertainty relative to the competitors
private information - an uncertainty relative to the competitors
bounded rationality type or bidding function - These uncertainties can be combined into a
price function - Stationary case
63Price Problem Formulation1st price Auction
64Price Problem Formulation2nd price Auction
return