Multi-Agent Systems

7. Auction and Voting Protocols

Florin Leon
“Gheorghe Asachi” Technical University of Iași, Romania
Faculty of Automatic Control and Computer Engineering

http://florinleon.byethost24.com/lect_mas.html

2024
Auction and Voting Protocols
1. Auction Protocols
1.1. Auction Types
1.2. Bidding Strategies
1.3. Revenue Equivalence
1.4. Combinatorial Auctions
2. Voting Protocols
2.1. Voting Systems
2.2. Condorcet Consistency
2.3. Susceptibility to Manipulation
2.4. Fundamental Limitations of Voting Systems

2
Auction and Voting Protocols
1. Auction Protocols
1.1. Auction Types
1.2. Bidding Strategies
1.3. Revenue Equivalence
1.4. Combinatorial Auctions
2. Voting Protocols
2.1. Voting Systems
2.2. Condorcet Consistency
2.3. Susceptibility to Manipulation
2.4. Fundamental Limitations of Voting Systems

3
Auctions
 An auction involves the transfer of a specific item from a seller
(auctioneer) to a buyer (bidder) for a specific price or bid
 The purpose of the auction is for the auctioneer to find the best
bidder
 In most situations, the auctioneer wants to maximize the price
of the auctioned item, while the bidders want to reduce the
price
 In MAS, auctions are used for applications that involve the
allocation of goods, tasks or resources

4
Auctions in Real World
 Google Ads
 Auctions are run to decide which ads get shown in which position for which
searches
 77% of Google’s revenue (238 billion USD) in 2023 came from these auctions
 Frequency spectrum licenses
 After the mobile phone industry had taken off, governments decided to sell
spectrum licenses through auctions
 UK: auctions for 3G (2000), in 150 rounds, with the participation of 13
companies. The 5 licenses sold for 4-6 billion GBP, with the government
earning 23 billion GBP (2.5% of GDP). Afterwards, this market sector shrank
rapidly and many of the winners regretted their participation
 US: AWS-3 spectrum (2015) – 45 billion USD, C-Band (2021) – 81 billion USD
 Romania: in 2022 – 433 million EUR
Cloud Resource Allocation
 Buyers: users (individuals or businesses) that seek cloud resources
like virtual machines, storage, or computing power
 Sellers: cloud service providers (e.g., AWS, Google Cloud) that offer
these resources for allocation
 Bids: offers that specify resource needs (e.g., number of virtual
machines, CPU cores, GB of storage), other constraints like time
duration, quality of service, or geographical preferences, and the
price they are willing to pay
 Products: computing power (virtual CPUs or GPUs), storage (SSD or
HDD space), network bandwidth, virtual machines (specific
configurations of memory, CPU, and storage)

6
Task Allocation in Production Units
 Coordinator agent: oversees the operation of multiple plants, and
maintains a finite list of urgent tasks that can be executed within a
period of time
 Bidders: each plant in the production unit is represented by an agent
that tracks all pending maintenance tasks for the plant. These agents
compete to get their tasks a place in the coordinator agent’s list
 Tasks are evaluated based, e.g., on urgency (likelihood of machine
failure), impact of delays on production efficiency, and resource
requirements for completing the task

7
Auction and Voting Protocols
1. Auction Protocols
1.1. Auction Types
1.2. Bidding Strategies
1.3. Revenue Equivalence
1.4. Combinatorial Auctions
2. Voting Protocols
2.1. Voting Systems
2.2. Condorcet Consistency
2.3. Susceptibility to Manipulation
2.4. Fundamental Limitations of Voting Systems

8
Classification
 Auctions can be classified by:
 How buyers estimate the value of the auctioned item
 The methods used to submit bids and determine the final price

9
Types of Valuations
 Common (objective) value
 The value of the item is the same for all bidders, but each has only an
approximate estimate of it
 Each bidder estimates the item’s value independently, often based on
limited information
 E.g., an oil-drilling tract, where the amount of oil determines the revenue,
while each company relies on its own expert's estimate of the oil reserve
 Private (subjective) value
 Bidders assign different values to the item based on personal preference or
sentimental reasons
 Each bidder knows his own valuation but not the others' valuations
 The seller also lacks specific knowledge of individual valuations
 E.g., items with sentimental or historical significance, like a gown worn by
Princess Diana or a necklace from Jacqueline Kennedy Onassis
10
Types of Bidding
 Open cry
 Bidders announce their bids in public, and others are able to hear
the bids made
 Ascending bids
 Descending bids
 Sealed bid
 Won with the first price
 Won with the second price

11
English (Ascending) Auction
 Open cry, first price
 The auctioneer starts at a low price and announces increasingly high
prices for a product, waiting to receive bids at each price before
increasing it
 When bidding stops, the product is awarded to the last bidder who
offered the highest price
 Around 95% of Internet auctions are of this kind
 The classic use is the sale of antiques and artwork

12
Example

13
Dutch (Descending) Auctions
 Open cry, first price
 The auctioneer starts at a very high price and announces lower and
lower prices until one of the bidders makes an offer, and takes the
item
 Also called a “descending clock” auction (some auctions use a clock to
display the prices)
 Ties are broken by restarting the descent from a slightly higher price
than where the tie occurred
 The winner pays the price at which he “stops the clock”

14
Dutch Auction
 High volume, since auction proceeds swiftly
 Often used to sell perishable products:
 Flowers in the Netherlands (e.g., Aalsmeer)
 Fish in Spain
 Tobacco in Canada

15
Example
 Aalsmeer auction trades 43 million flowers every day

16
Example: Price Clock

17
Sealed Bid Auctions
 The auction is private and the bidders cannot see the bids made by
others
 In many cases, only the winning bid is announced
 Sealed bid auctions do not require an auctioneer, only a supervisor to
open the bids and announce the winner
 First price sealed bid
 The highest bid wins, and the winner pays that highest price, i.e., what he
bid
 Public procurement, contracting infractructure work, privatizing state
companies
 Second price sealed bid
 Vickrey auction, an example of mechanism design

18
All-Pay Auctions
 A common value, sealed bid first price auction in which every bidder,
win or lose, pays to the auctioneer the amount of his bid
 An auction where the losers also pay may seem strange, but in fact it
models many situations involving competitions
 Hundreds of competitors spend 4 years preparing for the next
Olympic games, but only a few get medals and recognition
 In political contests, all candidates spend a lot of time and money, but
the losers do not get any refunds

19
Auction and Voting Protocols
1. Auction Protocols
1.1. Auction Types
1.2. Bidding Strategies
1.3. Revenue Equivalence
1.4. Combinatorial Auctions
2. Voting Protocols
2.1. Voting Systems
2.2. Condorcet Consistency
2.3. Susceptibility to Manipulation
2.4. Fundamental Limitations of Voting Systems

20
English Auction
 Determine your maximum willingness to pay – valuation V – which
guides all bidding decisions: only continue bidding if the current bid
remains below this valuation
 If the last bid r < V, raise it by the minimum increment allowed
 If r ≥ V, stop bidding to avoid overpaying
 If no one outbids, you get the item at a price close to the final r
 Your profit is V – r, the difference between your valuation and the
final price

 In an English auction, the winner pays near the valuation of the

second-highest bidder, with the final price influenced by the auction’s
minimum bid increment
21
First Price Sealed Bid
 As the winner must pay the bid amount, bidding your full valuation V
would result in zero profit
 Lowering your bid slightly increases your chance for a positive profit
 By bidding less than V, you risk losing the item to a slightly higher
bid but gain the possibility of winning with a higher profit margin
 The strategy is to balance this trade-off: the ideal bid is achieved
when the reduction from V, called shading, maximizes profit without
excessively reducing your chances of winning
 An exemple using calculus is presented later

22
Dutch Auction
 The bidding strategy is similar to that for the first price sealed bid
(FPSB) auction
 In a Dutch auction, wait until the auctioneer’s price drops close to or
below your valuation V before bidding
 Bidding immediately at V yields zero profit, so a delayed bid can be
more profitable
 Waiting for a lower price increases potential profit but also risks a
rival claiming the item before you bid
 Like in a FPSB auction, the optimal strategy is to shade your bid by
waiting, with the ideal moment found by weighing profit gain against
the increased risk of loss

23
Vickrey (Second Price) Sealed Bid Auction
 In this modified sealed bid auction, the highest bidder wins the item
but pays the price of the second highest bid
 Vickrey showed that in this format, each bidder’s best strategy is to
bid his true valuation
 This approach is informally referred to as Vickrey’s truth serum because it
incentivizes honest bidding
 However, there is a cost to this
 While this format elicits true valuations from buyers, it typically reduces the
seller’s profit
 This effect is similar to bid shading in first price auctions, where the bid
reduction limits seller earnings

24
Example: Bidding for a Tea Set
 Your valuation V = 3000: the maximum amount you are willing to pay
 Your bid b: you can submit any real value
 Rival’s highest bid r: the outcome depends only on the largest competing bid
r, as it determines whether your bid wins

 1. Bidding above 3000: bidding more than V is irrational since winning at a

price above your valuation yields a negative profit, resulting in a loss rather
than a gain
 2. Bidding below 3000: bidding less than V risks losing the tea set to a rival
whose bid falls between your bid b and 3000
 3. Bidding 3000: bidding exactly your true valuation maximizes your profit
potential without risking overpaying or losing unnecessarily

25
Bidding Above 3000
 1.1. r < 3000
 You win the tea set and pay r with a
profit of 3000 − r. This outcome
matches the profit you’d get if you
had bid exactly 3000
 1.2. 3000 < r < b
 You win the tea set but pay more
than it is worth to you, with a
negative profit. Bidding 3000 would
have been better, as it would have
saved you from this loss
 1.3. r > b
 You do not win the tea set. This
outcome remains unchanged even if
you had bid your true valuation of
3000.

26
Bidding Below 3000
 2.1. r < b
 You win the tea set and pay r, with
the same profit as if you had bid
exactly 3000. Bidding lower does not
provide any additional benefit
 2.2. b < r < 3000
 Your rival wins the tea set. If you had
bid 3000, you would have won
instead, paid r, and achieved a profit
of 3000 − r. Bidding lower results in
a lost profit opportunity
 2.3. r > 3000
 You do not win the tea set. Even if
you had bid 3000, you still wouldn’t
win, so there is no advantage to
bidding lower

27
Dominant Strategy
 Truthful bidding (bidding your true valuation) is never worse and
often better than bidding above or below
 Regardless of rivals’ bids, bidding your true valuation maximizes
your potential profit
 In second price sealed bid auctions, bidding your true valuation is the
dominant strategy

28
Optimal Strategy in All-Pay Auctions
 In all-pay auctions, your bid is a sunk cost if you lose, which
motivates participants to bid aggressively in hopes of getting the
prize
 However, excessive bidding by all participants often leads to total
bids exceeding the prize’s value, benefiting the auctioneer
 Pure aggressive bidding leads to unsustainable competition, while
abstaining allows a single bidder to get the prize easily
 For an auction with n bidders, the equilibrium strategy for each
bidder is to participate in the auction with probability 1 / n
 Each bidder ensures that his expected cost of bidding does not exceed
the expected benefit of winning the auction

29
Example
 At Princeton there is a tradition of giving the professor a polite round
of applause at the end of a semester
 Once Prof. Avinash Dixit offered 20 USD to the student who kept
applauding continuously the longest
 Although most students dropped out between 5 and 20 minutes,
three went on for 4 hours

30
The Winner’s Curse
 The winner’s curse states that the bidder who wins an auction may
pay more than the item is worth
 Examples: bidding for a jar of coins or a lease for the oil or gas-
drilling rights on a tract of land or sea
 Each bidder has an approximate estimate of the value
 From these multiple estimates, probably the average is close to the
true, unknown value
 The winning bid is obviously (much) higher than the average

31
Seller Strategies
 A concern of many sellers is that a product might end up with a
bidder for less than expected value
 The solution is to set a reserve price for auctioned items
 Sellers reserve the right to withdraw the item from sale if they do not
receive a bid higher than the reserve price
 The type of auction best for the seller depends on the bidders’ risk
attitude and their beliefs about the value of the auctioned item

32
Risk-Averse Bidders
 Risk-averse bidders generally want to win
 The seller does better with the first price auction in the presence of
risk aversion only in the sealed bid case
 In English auction, the bidders’ attitudes toward risk are irrelevant to
the outcome

33
The Amsterdam Auction
 Since medieval times, property in the Netherlands has been sold
using the Amsterdam auction
 It starts with an English auction
 When down to the final two bidders, it starts a Dutch auction stage
 Dutch auction starts from twice the final price of the English auction

 This profits the seller who may get more than the second bid value
Bidder Collusion in Vickrey Auctions
 In a Vickrey second price sealed bid auction, a small group of bidders
can exploit the auction by colluding: they submit one high bid and
another low second-highest bid, and win the item at a low price
 Setting a reserve price can partly mitigate the risk of collusion by
ensuring a minimum acceptable price
 However, reserve prices alone may not fully counteract the effects of
collusion in auctions with few bidders

35
Collusion in First Price Sealed Bid Auctions
 In first-price sealed-bid auctions, collusion is harder to maintain because it
resembles a multi-person prisoners' dilemma
 Each colluding bidder has an incentive to cheat by submitting a high bid
independently, hoping to win the item and bypass profit-sharing with the
group
 The sealed bid structure makes it difficult to detect or punish a collusive
member who bids outside of the agreed terms
 Collusion breaches become visible only after the auction concludes, by which
time it is too late to respond or penalize the cheater
 While collusion is difficult in isolated auctions, it can be more feasible if the
same bidders participate in a series of similar auctions
 Repeated interactions may create conditions for a repeated game, where
cooperation among bidders may be easier to sustain

36
Fraudulent Seller Practices:
Shilling and Bid Inflation
 Shilling involves the seller placing false bids to drive up the auction
price
 In English auctions, this can be achieved by using a shill: an agent who
pretends to be a legitimate bidder, but raises the final bid amount artificially
 Online, shilling becomes easier as sellers can register multiple identities to
bid on their own items
 Although online platforms enforce rules and monitoring systems to
counteract this behavior, it remains a risk
 Sellers can also inflate the second-highest bid in a Vickrey auction to
raise the final selling price, taking advantage of the fact that this bid
level is not public

37
Auction and Voting Protocols
1. Auction Protocols
1.1. Auction Types
1.2. Bidding Strategies
1.3. Revenue Equivalence
1.4. Combinatorial Auctions
2. Voting Protocols
2.1. Voting Systems
2.2. Condorcet Consistency
2.3. Susceptibility to Manipulation
2.4. Fundamental Limitations of Voting Systems

38
Revenue Equivalence Principle
 Risk-neutral bidders focus solely on the expected value of outcomes,
disregarding uncertainty
 Bidders determine an object’s value based solely on their independent
assessments, uninfluenced by others’ estimates

 The revenue equivalence principle says that under these conditions –

risk neutrality and independent valuations – the seller can expect the
same average revenue (over a large number of trials) from any of the
four primary types of auction: English, Dutch, and first- and second-
price sealed bid

39
Example
 Alice wants to sell her house, which has no value to her without a
buyer
 Potential buyers, Bob and Carol, have unknown valuations
 If Alice knew their valuations, she would offer the house at one unit
below the higher valuation for an immediate sale
 Alice assumes that each valuation (for Bob and Carol) is equally likely
between 0 and 36 million and independent of the other’s valuation
 In any auction format the agent with the higher valuation will win.
Therefore, each bidder’s chance of winning is determined only by his
valuation, regardless of auction type

40
The Selling Price
winner = 0
loser = 0
counts = 1000000
max_val = 36

for _ in range(counts):
v1 = random.random() * max_val
v2 = random.random() * max_val
winner += max(v1, v2)
loser += min(v1, v2)

winner /= counts
loser /= counts

print(f"W = {winner:.0f} L = {loser:.0f}") # => W = 24 L = 12 (2/3 max, 1/3 max)

41
The Selling Price
 In any auction format, the winning bidder expects to pay, on average,
half of his own valuation for the house
 Given rational bidding and a reserve price of zero, Alice’s average
revenue is 12 million
 English auction: Bob and Carol bid up to their respective valuations,
stopping once the price hits the lower of their two valuations. The
winner pays the loser’s valuation, which is on average half the
winner’s valuation
 Vickrey auction: the winner pays the second-highest bid (the same as
the loser’s valuation in an English auction), yielding the same
average revenue

42
Dutch / First Price Sealed Bid
 Let V be the valuation, B the bid, R the profit, 𝔼[R] the expected
profit and P(win) the probability to win
 𝔼 𝑅 = 𝑅 ∙ 𝑃(𝑤𝑖𝑛)
 𝐵~𝑉 ⇒ 𝐵 = 𝛼 ∙ 𝑉, where 𝛼 ∈ [0, 1]
 𝑅 = 𝑉 − 𝐵 = (1 − 𝛼) ∙ 𝑉
 𝑃 𝑤𝑖𝑛 𝐵 ~𝐵 ⇒ 𝑃 𝑤𝑖𝑛 𝐵 = 𝛽 ∙ 𝐵
 ⇒ 𝔼 𝑅 = 1 − 𝛼 ∙ 𝑉 ∙ 𝛽 ∙ 𝛼 ∙ 𝑉 = 𝑉 2 ∙ (1 − 𝛼) ∙ 𝛼 ∙ 𝛽

 𝔼[R] should be maximized

 V and β are fixed ⇒ (1 − 𝛼) ∙ 𝛼 should be maximized
 By derivation: 𝛼 − 𝛼 2 ′ = 0 ⇒ 1 − 2𝛼 = 0 ⇒ 𝛼 = 1 2
 ⇒𝐵=𝑉 2
43
Auction and Voting Protocols
1. Auction Protocols
1.1. Auction Types
1.2. Bidding Strategies
1.3. Revenue Equivalence
1.4. Combinatorial Auctions
2. Voting Protocols
2.1. Voting Systems
2.2. Condorcet Consistency
2.3. Susceptibility to Manipulation
2.4. Fundamental Limitations of Voting Systems

44
Combinatorial auctions
 Bidders bid on combinations of items, not individual items
 A set of items can be worth more than the sum of the individual item
values
 Applications:
 Lots of land (adjoining lots have a higher value than separate lots)
 Allocation of radio frequency spectrum (the same frequency in adjacent
localities)
 Allocation of television advertising
 The goal is to maximize profit for the seller
 Determining the winner is an NP-hard problem

45
Linear Programming Formulation of
Winner Determination Problem
 The winner determination problem in combinatorial auctions can be
reduced to a linear programming problem and solved in polynomial time
 Maximize:

𝑥𝑏 ∙ 𝑣𝑎𝑙𝑢𝑒(𝑏)
𝑏∈𝐵
 Subject to:

𝑥𝑏 ≤ 1 , ∀𝑗 ∈ 𝑀
𝑏|𝑗∈𝑖𝑡𝑒𝑚𝑠 𝑏

𝑥𝑏 ∈ 0, 1 , ∀𝑏 ∈ 𝐵

 xb is a bit that denotes whether bid b is a winning bid

 Maximize the sum of the bid values given that each item can be in, at most,
one winning bid

46
Example
Bids:

1) {A}: 1
2) {B}: 3
3) {C,D}: 5
4) {B,E}: 6
5) {A,E}: 7
6) {A,C}: 8

Transformed into:

A) x1 + x5 + x6 <= 1
B) x2 + x4 <= 1
C) x3 + x6 <= 1
D) x3 <= 1
E) x4 + x5 <= 1

47
1) {A}: 1
2) {B}: 3
3) {C,D}: 5
4) {B,E}: 6
5) {A,E}: 7
6) {A,C}: 8

Total: 15
Auction and Voting Protocols
1. Auction Protocols
1.1. Auction Types
1.2. Bidding Strategies
1.3. Revenue Equivalence
1.4. Combinatorial Auctions
2. Voting Protocols
2.1. Voting Systems
2.2. Condorcet Consistency
2.3. Susceptibility to Manipulation
2.4. Fundamental Limitations of Voting Systems

50
Voting and Social Choice Theory
 Multi-agent environments consist of self-interested agents with diverse
or conflicting preferences, which makes consensus challenging but
essential
 Clear aggregation techniques help to balance individual agent interests,
and enable decisions that represent the collective preferences of the
system
 Voting is a general mechanism for combining agent preferences
 Originally from economics and political science, social choice theory
study the methods that allow multi-agent systems to function as a
“society” of autonomous agents with distinct goals
 Voting theory focuses specifically on electoral systems, while social
choice theory addresses broader questions of collective decision-making,
including non-electoral contexts and welfare implications
51
Auctions vs. Voting
 The purpose of auctions is to maximize profit only for the
parties involved, i.e., the buyer and the seller
 The purpose of voting is to maximize the social good for all
agents involved or affected

52
Examples
 Collaborative robot navigation
 In robotic teams (e.g., drones or autonomous underwater vehicles),
individual agents may have varying sensor data about their environment
 The challenge is to collectively decide on the most efficient path that avoids
obstacles and achieves the objectives of the team
 By voting based on their observations, the agents aggregate their data to
make a unified navigation decision
 Automated meeting scheduling
 Each user is represented by an agent, with different preference dimensions,
e.g., day of the week, time of day, and invitees
 These dimensions induce rankings of meeting options
 The rankings are aggregated using a voting rule to compute the optimal
meeting choice automatically, which balance preferences and constraints

53
Examples
 Decision-making in smart cities
 In smart cities, residents vote on urban planning proposals or resource
allocations, such as park locations
 Each citizen is represented by an agent with unique interests
 Voting can aggregate preferences and resolve conflicts
 Sensor networks
 Sensor networks, composed of autonomous and heterogeneous sensors,
monitor and identify events in an environment
 The challenge lies in collaboratively determining the ground truth from
distributed, noisy, or partial observations
 Each sensor acts as an agent, and votes on its findings, and the aggregated
votes help the network to infer accurate, reliable event identification

54
Auction and Voting Protocols
1. Auction Protocols
1.1. Auction Types
1.2. Bidding Strategies
1.3. Revenue Equivalence
1.4. Combinatorial Auctions
2. Voting Protocols
2.1. Voting Systems
2.2. Condorcet Consistency
2.3. Susceptibility to Manipulation
2.4. Fundamental Limitations of Voting Systems

55
Binary Methods
 Binary methods ask voters to choose between two alternatives at a
time
 In elections where there are only two candidates, the votes may be
aggregated using the majority rule: the one with the most votes wins
 We use the term candidate in a general sense – it can represent a
joint plan, an allocation of resources, etc.

56
Binary Methods
 When dealing with several candidates, pairwise voting can be used
 Such a method, in which each alternative is presented in turn against
the others, is called a Condorcet method
 The candidate who wins all one-to-one votes against the other
candidates is called the Condorcet winner
 Other pairwise voting procedures produce scores
 The Copeland index measures the number of wins and losses a
candidate achieved in all one-to-one voting (presented later)

57
Binary Methods
 Another method of paired voting is the amendment procedure, used
by the US congress to vote on new laws
 This procedure is used to determine a winner from three alternatives
 In the first round, one of two alternatives is chosen: the proposed
new law vs. the law + amendments
 In the second round, the winner is chosen between the winner of the
first round and the third alternative, the status quo

58
Pluralistic Methods
 They allow voters to consider three or more alternatives
simultaneously
 Positional methods use the scores of alternatives on the ballot to
assign points

 The best-known positional method is the plurality rule: each voter

gives 1 vote (point) to the preferred candidate and 0 points to the rest
of the alternatives
 The candidate with the most votes (points) wins
 A plurality winner does not necessarily have the majority of votes

59
Example: Plurality Voting
 4 people vote for A
 3 people vote for B
 2 people vote for C
 A wins with the most votes, although it has no majority

60
Duverger’s Law
 Single-member districts with plurality voting (“first past the post”)
tend to lead to a two-party system
 Proportional representation systems tend to encourage multi-party
systems

 Plurality voting disadvantages smaller or third-party candidates who

are less likely to achieve the largest share of votes
 Voters may want to avoid wasting their vote and vote strategically
for a candidate of the two most viable options rather than their first
choice, if that candidate has little chance of winning

61
Pluralistic Methods
 Borda count: voters must rank all alternatives and mark these ranks
on the ballot papers
 Scores are calculated based on these ranks
 For example, for 3 candidates, rank 1 has 3 points, rank 2 has 2
points, and rank 3 has 1 point
 It is also possible that rank 1 has 2 points, rank 2 has 1 point, and
rank 3 has 0 points
 The points for each candidate are added up and the candidate with
the most points wins

62
Example
 Consider 4 candidates
 Scoring: 1st place = 3 points, 2nd place = 2 points, 3rd place = 1 point,
4th place = 0 points
 Voter preferences:
 Voter 1: A > B > C > D (prefers A to B, B to C, C to D)
 Voter 2: B > C > D > A
 Voter 3: C > D > B > A
 Voter 4: D > A > B > C
 Total points:
 A – 3+0+0+2=5, B – 2+3+1+1=7, C – 1+2+3+0=6, D – 0+1+2+3=6
 The winner is B

63
Pluralistic Methods
 In approval voting, voters approve the alternatives they prefer
 All upvotes get 1 point
 The alternative with the most approvals wins
 It is used, for example, to elect representatives from faculty councils
or senate members in Romanian universities
 Unlike positional methods, no distinction is made between
alternatives based on the position (rank) obtained
 A minimum threshold of approvals may be imposed

64
Example
 Voter approvals:
 Voter 1: A and B
 Voter 2: B and C
 Voter 3: C and D
 Voter 4: B
 Total approvals:
 A – 1, B – 3, C – 2, D – 1
 The winner is B

65
Mixed Methods
 Combines plurality and binary voting
 For example, majority runoff is a two-stage method used to reduce a
large number of alternatives to a binary decision
 In the first round voters choose their preferred options
 If an option has the majority of votes, it is declared the winner
 Otherwise, in the second round, the first and second places from the
first round compete, and the winner is determined using majority
rule

66
Example
 First round votes:
 A: 30 votes
 B: 35 votes
 C: 25 votes
 D: 10 votes
 No candidate gets a majority
 Second round votes:
 A: 45 votes
 B: 55 votes
 The winner is B

67
Mixed Methods
 Another mixed method is using successive rounds
 In each round, the least voted option is eliminated
 In the last round, there are two options; the method becomes
binary and the majority rule is applied

68
Example
 First round votes:
 A: 25 votes, B: 30 votes, C: 20 votes, D: 15 votes
 D is eliminated for having the least votes
 Second round votes:
 A: 40 votes, B: 35 votes, C: 25 votes
 C is eliminated for having the least votes.
 Final round votes (binary majority rule):
 A: 55 votes, B: 45 votes
 The winner is A

69
Mixed Methods
 In the single transferable vote (or instant runoff), each voter
indicates the order of his preferences
 In the first round, if none of the candidates get a majority of one, the
worst performing candidate is eliminated
 His votes on the ballots where he received first place are
redistributed to second place
 Similar reallocations continue in subsequent rounds until a majority
is achieved

70
Example
 Voter preferences:
 Voter 1: B > A > C
 Voter 2: A > C > B
 Voter 3: A > B > C
 Voter 4: C > B > A
 Voter 5: C > B > A
 Round 1 (first preferences):
 A - 2 votes, B - 1 vote, C - 2 votes
 No candidate has a majority (3 votes)
 B is removed

71
Example
 Voter preferences:
 Voter 1: B > A > C → A > C (gives his vote to A)
 Voter 2: A > C > B → A > C
 Voter 3: A > B > C → A > C
 Voter 4: C > B > A → C > A
 Voter 5: C > B > A → C > A
 Round 2:
 A - 3 votes, C - 2 votes
 The winner is A

72
Summary
 Binary methods
 Majority rule
 Pairwise voting: Copeland, amendment procedure
 Pluralistic methods
 Plurality rule
 Borda count
 Approval voting
 Mixed methods
 Majority runoff
 Successive rounds voting
 Single transferable vote

73
Copeland Method
 Candidates are ranked by the number of wins minus the
number of losses
 Consider an election with 5 candidates, who receive the
following number of votes:

74
Copeland Method
 The result of the 10 possible pairs of comparisons

75
Copeland Method
 There is no Condorcet winner
 Candidate A has the highest score (wins minus losses) and is
the Copeland winner

76
Schulze Method
 Selects a single winner using preference votes
 Can also create an ordered list of winners
 Currently, the most widely used Condorcet method, e.g.,
by organizations such as Wikimedia, Debian, Gentoo

77
Schulze Method
 Step 1: for each pair of uneliminated candidates X and Y, if there is a
direct path of uneliminated links from X to Y, then we write X → Y,
otherwise we write not X → Y
 X → Y means X defeats Y (directly or transitively)
 Step 2: for each pair of remaining candidates X and Y, if X → Y and
not Y → X, candidate Y is removed and all links starting or ending in
Y are removed
 Step 3: the weakest link is removed; if multiple links are tied, all are
removed
 The procedure ends when all links have been removed
 The winner is the last remaining candidate

78
Schulze Method
 Consider 4 candidates and 30 voters with the following
preferences:

79
Schulze Method
 The pairwise defeat matrix

80
Schulze Method

 A → B, A → C, A → D (transitively: A → C → D)
 B → A (transitively), B → C, B → D (transitively)
 C → A (transitively), C → B (transitively), C → D
 D → A, D → B, D → C (transitively)

81
Schulze Method

 We have no X, Y pairs in the X → Y and not Y → X situation, so

no candidate is eliminated
 The weakest link (A → B = 16) is eliminated

82
Schulze Method

 A → B, A → C, A → D
 B → A, B → C, B → D
 C → A, C → B, C → D
 D → A, D → B, D → C

83
Schulze Method

 We have no X, Y pairs in the X → Y and not Y → X situation, so

no candidate is eliminated
 The weakest link (A → C = 17) is eliminated

84
Schulze Method

 not A → B, not A → C, not A → D

 B → A, B → C, B → D
 C → A, C → B, C → D
 D → A, D → B, D → C

85
Schulze Method

 From B → A and not A → B, candidate A and all links starting

or ending with this candidate are removed

86
Schulze Method

 The weakest link (B → C = 19) is removed

87
Schulze Method

 not B → C, not B → D
 C → B, C → D
 D → B, not D → C

88
Schulze Method

 From C → B and not B → C, candidate B and its links are removed

 From C → D and not D → C, candidate D and its links are removed
 Therefore candidate C is the winner

89
Auction and Voting Protocols
1. Auction Protocols
1.1. Auction Types
1.2. Bidding Strategies
1.3. Revenue Equivalence
1.4. Combinatorial Auctions
2. Voting Protocols
2.1. Voting Systems
2.2. Condorcet Consistency
2.3. Susceptibility to Manipulation
2.4. Fundamental Limitations of Voting Systems

90
Condorcet Consistency
 A voting system can be considered to have Condorcet
consistency, or be Condorcet consistent, if it elects a
Condorcet winner

 Methods not guaranteed to meet the Condorcet criterion

 Plurality voting
 Borda count
 Condorcet methods
 Copeland method
 Schulze method

91
Plurality Voting and Condorcet Criterion
 Consider an elections where:
 30% prefer A > B > C
 30% prefer C > A > B
 40% prefer B > A > C
 B is the winner (40 vs. 30, 30)

 However, A is the Condorcet winner, since it defeats B (60 to 40)

and C (70 to 30)

92
Borda Count
 The Borda count does not satisfy the Condorcet criterion in the
following case
 Consider an election with 5 voters and 3 alternatives:
 3 voters have preferences A > B > C
 2 voters have preferences B > C > A
 Scores: first – 2 points, second – 1 point, third – 0
 B is the Borda winner with 3 points, against A with 2 and C with 1
 However, A is preferred by 3 out of 5 voters, and is the Condorcet
winner

93
Condorcet Paradox
 It is the most famous and important voting paradox
 It occurs when no Condorcet winner results from the pairwise voting
process
 Condorcet paradox says that there are situations in which, no matter
which outcome is chosen, a majority of voters will be unhappy with
that result

 Let there be 3 councilors – Left, Center and Right

 They have to vote on 3 social policies:
 One that increases benefits (Generous – G)
 One that decreases benefits (Decreased – D)
 One that maintains the current state (Average – A)

94
Example
 The voting results:

 No policy has a majority

 Group preferences are cyclic: G > A > D > G
 This preference cycle is an example of an intransitive ordering of
preferences
 Individual preferences are transitive: if one prefers A to B and B to C,
one also prefers A to C
 It leads to the Agenda Paradox (presented later)
95
Condorcet Paradox
 All councilors have transitive preferences, but the council as a whole
does not
 This is the Condorcet paradox: even if individual preferences are
transitive, there is no guarantee that the social preference induced by
the Condorcet method is also transitive
 Therefore, societies, institutions or other large groups of people
should not always be analyzed as if they were acting as an individual
 The paradox is more likely to arise when large groups of voters
consider a large number of alternatives

96
Reversal Paradox
 Using Borda count, if the alternative with the lowest rank, d, is
eliminated, the worst alternative becomes the best and vice versa
 Scoring 1 point for last

97
Borda Count
 The only cases where Borda count produces paradoxical results occur
when some alternatives are removed after the votes have been
collected
 Because such results can be prevented by using only ballots with the
full set of candidates, the Borda procedure is regarded as one of the
best methods of vote aggregation

98
Heuristic Recommendation
 Select a Condorcet winner if one exists, otherwise use
Borda count

99
Auction and Voting Protocols
1. Auction Protocols
1.1. Auction Types
1.2. Bidding Strategies
1.3. Revenue Equivalence
1.4. Combinatorial Auctions
2. Voting Protocols
2.1. Voting Systems
2.2. Condorcet Consistency
2.3. Susceptibility to Manipulation
2.4. Fundamental Limitations of Voting Systems

100
Strategic Manipulation of Votes
 Some agents may not vote according to their true preferences
 They prefer to lie, just as some people do not vote for their favorite
political candidate if they think he has no chance of winning
 For example, they can vote for their second favorite (tactical voting)
 Also, the organizers can manipulate the voting system by
adding or removing candidates (strategic control)

101
Susceptibility to Manipulation
 Voting methods in descending order of manipulability,
from most to least:
 Plurality voting
 Borda count
 Approval voting
 Amendment procedure
 Single transferable vote
 Majority rule

102
Plurality Voting
 Plurality voting is highly susceptible to tactical voting
 Voters may not choose their true preference but instead vote for a
less-preferred candidate they believe has a better chance of winning
to avoid their least-preferred candidate winning
 “Lesser of two evils” strategy
 Plurality voting can be manipulated by a group of voters

103
Example: Tactical Voting
 True preferences of voters:
 40%: A > B > C (winner A)
 35%: B > C > A
 25%: C > A > B
 The 25% voters who prefer C may vote for B (their second
choice) instead of C:
 A: 40%
 B: 60% (winner)
 C: 0%

104
Example: Spoiler
 A spoiler candidate is introduced to split the vote for a strong
opponent, to improve another candidate’s chances of winning
 Initially, two candidates: A (60% support) and B (40% support)
 Candidate C is introduced as a spoiler backed by B’s supporters to
attract votes from A’s base
 C positions himself with policies similar to A’s
 Resulting votes: A: 35%, B: 40%, C: 25% (winner B)

105
Borda Count
 True preferences:
 Voter 1: A > B > C > D
 Voter 2: B > C > A > D
 Borda points (scoring 0 for last):
 A - 4, B - 5, C - 3, D - 0 (winner B)
 Manipulation (burying):
 Voter 1: A > C > D > B
 New points:
 A - 4, B - 3, C - 3, D - 1 (winner A)

106
Borda Count: Theorems
 Nonmanipulability of Borda count with exactly 3 candidates
 A single voter cannot unilaterally change an election result from one
candidate to another candidate that he prefers

 Manipulability of Borda count with 4 or more candidates

 There exists an election where a single voter can unilaterally change
the result from one candidate to another candidate that he prefers

107
Approval Voting
 Voters may approve only their top choice or strategically
approve others to prevent undesirable outcomes

108
Pairwise Voting, Amendment Procedure
 When using a pairwise voting method, one can use the anticipation
of the results of the second round to determine an optimal strategy
for the first round
 It may be in the voter’s interest to appear to support a particular
alternative in the first round, even if it is not the desired one
 That way, the least favorite alternatives will not be able to win the
second round

109
The Agenda Paradox
 It assumes a binary voting procedure, but the final result depends on the
ordering of the alternatives
 A chairman can choose which pair of alternatives participate in the first
vote; knowing the intransitive social preference he can manipulate the
outcome
 If Left is chairman, he can force policy G to win by selecting policies A
and D in the first round ⇒ winner A; then A vs. G ⇒ winner G
 The order of the agenda is the real game; because the chairman creates
the agenda, his appointment is the primary source of strategic thinking

110
Single Transferable Vote
 Determining whether a single voter or a group of voters can
manipulate an STV election to achieve a preferred outcome is
NP-hard
 This applies both to constructive manipulation (causing a specific candidate
to win) and destructive manipulation (preventing a specific candidate from
winning)
 Determining whether adding or removing candidates can influence
the outcome of an STV election is NP-hard
 Strategic addition: introducing new candidates to split votes or change transfer
dynamics
 Strategic elimination: removing candidates who may get votes from others
 Despite NP-hardness, heuristics and approximation methods may
allow practical manipulation in certain cases

111
Majority Rule
 May’s theorem for manipulability
 Among all two-candidate voting systems that never result in a tie,
majority rule is the only one that treats all voters equally, treats
both candidates equally and is nonmanipulable

112
Auction and Voting Protocols
1. Auction Protocols
1.1. Auction Types
1.2. Bidding Strategies
1.3. Revenue Equivalence
1.4. Combinatorial Auctions
2. Voting Protocols
2.1. Voting Systems
2.2. Condorcet Consistency
2.3. Susceptibility to Manipulation
2.4. Fundamental Limitations of Voting Systems
2.4.1. Arrow’s Impossibility Theorem
2.4.2. Gibbard-Satterthwaite Theorem
113
Arrow’s Impossibility Theorem
 Arrow proposed 6 principles for voting systems, which seem
common sense
 He proved that they are inconsistent, i.e., no voting system can
satisfy them all

114
Conditions of Arrow’s Theorem
 The group ranking must be complete: it should produce a ranking for all
possible individual preferences
 It must be transitive: the final ranking should not produce inconsistent
results, such as A > B, B > C, C > A
 It must be Pareto efficient: given two alternatives A and B, if all voters
prefer A to B, then the aggregate ranking should place A above B
 The ranking must not be imposed by external factors: the result should
be based only on the preferences of individual voters
 It must not be dictatorial: no single voter should determine the group
ranking
 It must be independent of irrelevant alternatives (IIT): no change in the
set of candidates (addition or removal) should change the rankings of
the unaffected candidates
115
Arrow’s Impossibility Theorem
 Arrow’s impossibility theorem basically says that no infallible voting
system exists for three or more choices
 For two choices, majority rule is sufficient
 Simplifying slogans (not necessarily true): “No voting method is
fair”, “The only voting methods that is not flawed is a dictatorship”
 Often, the first 4 conditions are considered met, and the theorem
expresses the impossibility of satisfying the last 2 simultaneously.
Since dictatorship is unacceptable, IIT is impossible to satisfy
 Otherwise, the voting system can produce intransitive results

116
Black’s Condition
 To mitigate this result, some form of compromise is required when
choosing an aggregation method
 The transitivity condition places restrictions on the ordering of
individual voters’ preferences
 One such restriction is the single-peak preference requirement
(Black’s condition)

117
Single-Peak Preferences
 On the horizontal axis: voter preferences
 On the vertical axis: intensity of preferences
 Each voter must have only one most-preferred alternative, and
alternatives far from it must provide lower utilities
 If each voters’ preferences are single-peaked, then the voting
procedure produces a transitive social ordering

118
Gibbard-Satterthwaite Theorem
 In any voting system with three or more choices, if the system is designed to
always produce a single, deterministic winner and allows voters to rank
choices (i.e., express preferences), then one of the following must be true:
 The system is dictatorial: there exists a single voter whose preferences
determine the outcome, regardless of others’ preferences or
 The system limits the outcomes: some candidates can never win or
 The system is manipulable: there exist situations where a voter can achieve a
more preferred outcome by misrepresenting his true preferences
 In short, any reasonable voting system with three or more options is
manipulable (strategic voting is possible)
 While Arrow’s theorem focuses on the impossibility of designing a voting
system that satisfies a set of fairness conditions, the Gibbard-Satterthwaite
theorem shows the inevitability of strategic manipulation in voting systems

119
Conclusions
 Auctions enable the allocation of goods, tasks, or resources in
multi-agent systems
 Some auction types demand strategic bidding, which balances profit
maximization with the risks of losing resources or overpaying
 Voting aggregates diverse agent preferences, and ensures that
collective decision-making aligns with social good despite varying
goals and conflicting interests
 Voting systems face challenges such as strategic manipulation,
susceptibility to paradoxes, and trade-offs between fairness and
efficiency in representing agent preferences

120
Main References
 Avinash K. Dixit, Susan Skeath, David H. Reiley Jr. (2014).
Games of Strategy,‎ W. W. Norton & Company, 4th edition
 Jose M. Vidal (2012). Fundamentals of Multiagent Systems with
NetLogo Examples

121