1 of 53

Degree of Truthfulness

for Risk-Avoiding Agents

Speaker: Eden Hartman (Bar-Ilan University)

Joint With:

Erel Segal-Halevi

(Ariel University)

Biaoshuai Tao

(Shanghai Jiao Tong University)

2 of 53

Social Choice

Q: How to aggregate individual preferences into one collective choice?

Mechanism

A collective choice

Examples: Voting, Fair Division, Matching….

3 of 53

A Truthful Mechanism

Telling the truth is a dominant strategy for all agents

4 of 53

A Truthful Mechanism

 

Telling the truth is a dominant strategy for all agents

 

 

 

 

 

 

 

5 of 53

A Truthful Mechanism

 

Telling the truth is a dominant strategy for all agents

 

 

 

 

6 of 53

A Truthful Mechanism

Telling the truth is a dominant strategy for all agents

Joint report 1

Joint report 2

Joint report 3

Joint report 4

Joint report 5

Joint report 6

Tell the Truth

Manipulation 1

Manipulation 2

Manipulation 3

Manipulation 4

Always at-least as good

Sometimes strictly better

7 of 53

Truthfulness is the dream

Telling the truth is a dominant strategy for all agents

  • Predictable (in theory)
  • Save time and effort

8 of 53

Truthfulness is the dream

Many impossibility results

9 of 53

Truthfulness is the dream

Many impossibility results

Indivisible Goods Allocations:

  • With more than 5 goods, EF1 is incompatible with truthfulness

[Amanatidis, Birmpas, Christodoulou, and Markakis, 2017]

Cake Cutting:

  • Proportionality and EF are incompatible with truthfulness [Tao, 2022]

Voting:

  • With three or more candidates, the only truthful rules are dictatorships

[Gibbard, 1973; Satterthwaite, 1975]

Two-Sided Matching:

  • No stable matching mechanism is truthful for all agents [Roth, 1982]

Examples:

10 of 53

Truthfulness is the dream

Many impossibility results

Goal: Mechanisms that are close to truthfulness

11 of 53

Goal: Mechanisms that are close to truthfulness

Computational Complexity

Regret-Free Truth-Telling

Incentive Ratio

Probability of Having Profitable Manipulation

Risk-Avoiding Truthfulness

(RAT)

Maximin Strategy-Proofness

Not-Obviously Manipulable

Queries and Bits

And many more…

12 of 53

 

Computational Complexity

13 of 53

 

Computational Complexity

14 of 53

 

Computational Complexity

Getting information

Informational

15 of 53

The amount of information on others

necessary to enable good manipulations

Degree of Truthfulness

16 of 53

The amount of information on others

necessary to enable good manipulations

Degree of Truthfulness

17 of 53

The number of “known-agents”

necessary to enable good manipulations

Degree of Truthfulness

18 of 53

The number of “known-agents”

necessary to enable good manipulations

Degree of Truthfulness

19 of 53

The number of “known-agents”

necessary to enable good manipulations

Degree of Truthfulness

Truthful

Not Truthful

20 of 53

The number of “known-agents”

necessary to enable good manipulations

Degree of Truthfulness

Truthful

Not Truthful

n

21 of 53

The number of “known-agents”

necessary to enable good manipulations

Degree of Truthfulness

Truthful

Not Truthful

n

0

1

2

3

n-1

n-2

22 of 53

The number of “known-agents”

necessary to enable good manipulations

Degree of Truthfulness

Truthful

Not Truthful

n

0

1

2

3

n-1

n-2

  • Applicable to all social choice settings
  • Does not depend on the specific problem or input
  • Applicable to cardinal and ordinal valuations

23 of 53

The number of “known-agents”

necessary to enable good manipulations

Degree of Truthfulness

Sometimes Profitable, Never Harmful

24 of 53

The number of “known-agents”

necessary to enable good manipulations

Degree of Truthfulness

Risk-Avoiding Truthfulness (RAT)

For agents who avoid risk from not knowing others’ preferences

[Bu, Song, and Tao, 2023]

Sometimes Profitable, Never Harmful

25 of 53

The number of “known-agents”

necessary to enable good manipulations

Degree of Truthfulness

Risk-Avoiding Truthfulness (RAT)

For agents who avoid risk from not knowing others’ preferences

[Bu, Song, and Tao, 2023]

Sometimes Profitable, Never Harmful

26 of 53

The number of “known-agents”

necessary to enable good manipulations

Degree of Truthfulness

Risk-Avoiding Truthfulness (RAT)

For agents who avoid risk from not knowing others’ preferences

[Bu, Song, and Tao, 2023]

Sometimes Profitable, Never Harmful

27 of 53

The number of “known-agents”

necessary to enable good manipulations

Degree of Truthfulness

Sometimes Profitable, Never Harmful

Risk-Avoiding Truthfulness (RAT)

For agents who avoid risk from not knowing others’ preferences

[Bu, Song, and Tao, 2023]

28 of 53

The number of “known-agents”

necessary to enable good manipulations

RAT

Degree

Risk-Avoiding Truthfulness (RAT)

For agents who avoid risk from not knowing others’ preferences

[Bu, Song, and Tao, 2023]

Sometimes Profitable, Never Harmful

29 of 53

The number of “known-agents”

necessary to enable good manipulations

RAT

Degree

Worst case for the mechanism designer

Does there exist a situation

for which this number is sufficient?

30 of 53

The number of “known-agents”

necessary to enable good manipulations

RAT

Degree

Single Good Auctions

Indivisible Goods Allocations

Cake Cutting

Two-Sided Matching

Single-Winner Voting

31 of 53

Truthful

Not Truthful

n

0

1

2

3

n-1

n-2

Single Good Auctions

Second Price

32 of 53

Truthful

Not Truthful

n

0

1

2

3

n-1

n-2

Single Good Auctions

First Price

Second Price

33 of 53

Truthful

Not Truthful

n

0

1

2

3

n-1

n-2

Single Good Auctions

First Price

Second Price

First Price

with Discount

Value = 100

Known-Agents

Bid 105

  • Bid 110

Discount = 20$

34 of 53

Truthful

Not Truthful

n

0

1

2

3

n-1

n-2

Single Good Auctions

First Price

Second Price

First Price

with Discount

Avg. Between

First & Second Prices

35 of 53

Truthful

Not Truthful

n

0

1

2

3

n-1

n-2

Single Good Auctions

First Price

Second Price

First Price

with Discount

Avg. Between

First & Second Prices

Value = 100

Known-Agents

Unknown-Agents

n-2

Bid 90?

Bid 95

Max = 80

36 of 53

Truthful

Not Truthful

n

0

1

2

3

n-1

n-2

Single Good Auctions

First Price

Second Price

First Price

with Discount

Avg. Between

First & Second Prices

For any weighted average!

37 of 53

Single-Good Auctions

38 of 53

Single-Good Auctions

Combinatorial auctions?

39 of 53

With more than 5 goods, EF1 is incompatible with truthfulness

[Amanatidis, Birmpas, Christodoulou, and Markakis, 2017]

Indivisible Goods Allocations

Degree of max Nash-welfare mechanism?

Non-additive utilities?

40 of 53

Proportionality and EF is incompatible with truthfulness [Tao, 2022]

Cake Cutting

41 of 53

Proportionality and EF is incompatible with truthfulness [Tao, 2022]

Cake Cutting

Even-Paz?

n/2?

Max Nash welfare?

42 of 53

With three or more candidates, the only truthful rules are dictatorships

[Gibbard, 1973; Satterthwaite, 1975]

Single-Winner Voting

Degree of non-positional voting rules?

 

 

 

 

 

 

Degree of multi-winner voting rules?

 

43 of 53

No stable matching mechanism is truthful for all agents [Roth, 1982]

Two-Sided Matching

44 of 53

No stable matching mechanism is truthful for all agents [Roth, 1982]

Highest degree for stable matching mechanisms?

Two-Sided Matching

Stable matching mechanism with higher degree?

45 of 53

  1. Other mechanisms & Other applications:
    • Mechanisms: Max Nash-welfare, Even-Paz, Copeland, ….
    • Applications: Combinatorial auctions, Multiwinner voting, Budget aggregation, Facility location, …
  2. Using other definition for “good” manipulations:
    • Always profitable [Brams, Jones & Klamler, 2006]
    • Profitable in best or worst cases [Troyan & T Morrill , 2020]

Open Questions

Please contact us for any question and any answer!

eden.r.hartman@gmail.com erelsgl@gmail.com bstao@sjtu.edu.cn

46 of 53

  1. Other mechanisms & Other applications:
    • Mechanisms: Max Nash-welfare, Even-Paz, Copeland, ….
    • Applications: Combinatorial auctions, Multiwinner voting, Budget aggregation, Facility location, …
  2. Using other definition for “good” manipulations:
    • Always profitable [Brams, Jones & Klamler, 2006]
    • Profitable in best or worst cases [Troyan & T Morrill , 2020]
  3. Different types of information about the known-agents
  4. Tradeoffs: what is the highest degree for…

Open Questions

Please contact us for any question and any answer!

eden.r.hartman@gmail.com erelsgl@gmail.com bstao@sjtu.edu.cn

47 of 53

Goal: Mechanisms that are less likely to be manipulated

Computational Complexity

Regret-Free Truth-Telling

Incentive Ratio

Probability of Having Profitable Manipulation

Risk-Avoiding Truthfulness

(RAT)

Maximin Strategy-Proofness

Not-Obviously Manipulable

Queries and Bits

And many more…

48 of 53

 

The next step after information

Incomparable between different domains

Requires knowledge of the distribution over preference profiles

Only for numerical valuations

49 of 53

 

Our approach can also be combined with other types of manipulations

50 of 53

 

51 of 53

 

 

52 of 53

Truthful

Truth-telling is a dominant strategy

RAT (Risk-Avoiding Truthful)

No strategy dominates truth-telling

Relation of the RAT-degree to previous definition:

  • RAT-degree = n ⇔ Truthful (best for designer)
  • RAT-degree = 0 ⇔ Not even RAT (worst for designer)
  • Higher RAT-degree: more info is needed to manipulate safely (better for designer)

53 of 53