Trust Based Mechanism Design

1
Trust Based Mechanism Design

• By
• Rajdeep Dash,
• Gopal Ramchurn,
• Nicholas Jennings

2
Goal

• Fuse the fields of trust-modelling and mechanism
  design
• Mechanism design concerns itself with the
  allocation of resources.
• Trust measures how good an interaction partner is
  according to personal experience and its
  reputation

3
Deciding on a service provider
Garage X is better than Y
1
2
Jo
Anna
3
Anna
Garage Y is better than X!
Anna
Vanessa
4
Motivation
1
Which garage to choose to get the best price AND
my car repaired successfully?
2
3
Anna
5
Mechanism Design

• The design of systems of interacting agents
• Protocols (allocation payment schemes) to
  ensure that certain global properties are
  achieved at equilibrium (game-theoretic).
• Selfish, rational agents possessing private
  information i.e. each agent maximises its utility
  function given its type.

6
Task Allocation

• Agent wants a task to be completed (e.g.
  repairing a car).
• Problem
• Who to allocate task to ?
• How much to pay task performer ?
• Solution
• Vickrey Clarkes Groves (VCG) mechanism
• (e.g. second price (Vickrey) auction)

7
VCG Mechanism

• Class of mechanisms satisfying
• Efficiency outcome maximising utility for all
  agents
• Individually rationality incentivise
  participation
• Incentive compatibility incentivise truthful
  report
• under dominant strategy
• Works by
• Allocation choose allocation maximising utility
• Payment charge marginal utility contributed by
  agent

8
Applying VCG
Centre
(210,repair_car)
Allocate task to 1, 4 pays 1 50
Anna (4)
Ask(40,repair_car)
Ask(50,repair_car)
Ask(80,repair_car)
2
1
3
9
Cheapest not always best ..
10
Adding uncertainties

• Suppose that Anna has formulated a belief about
  the Probability of Success (POS), ,of other
  agents.

Agent i
1 40 0.525 70.25
2 80 1.0 130
3 50 0.85 128.5
Chose allocation that maximises expected utility,
11
Trust

• So far ones own experience, .
• Would be better to factor in others experiences,
  , as well.
• First time agent
• Greater pool of experience for seasoned user.
• Trust a way of aggregating ones direct
  experience with the reputation information
  provided by others.

12
Our model of trust
2

• Each time a task is performed, record POS
• Publicly transmit the POS to other agents
• Trust
• Use a simple averaging function over reports from
  all agents
• Give more weight to more trusted referrals, or
  agents with similar properties

Vanessa
Transmit POS
Anna
13
Incorporate Trust
Agent i
1 40 0.4 1.0 0.85 0.5 65
2 80 0.6 1.0 0.9 1.0 130
3 50 0.5 1.0 0.9 0.87 132.7
4 0.525 1.0 0.85 na na
VCG Fails!!
14
TBMD
(210,task1)
Centre
Trust model
Anna
3
1
2
4 transmits to centre -- value of task, observed
POS and trust model
1, 2, 3 transmit to centre -- cost of doing task
and observed POS
15

TBMD
(210,task1)
Centre
Trust model
Anna
3
2
2
1
3
16

TBMD
Centre
(210,task1)
Trust model
Anna
3
1
3
2
17
Trust Based Mechanism Design (TBMD)

• An individually rational and efficient ICDR
  mechanism that can also select those agents that
  are most successful at their task

18
Experimental Setup

• Set of buyers and sellers.
• Biased (but truthful!) report from seller.
• Buyers attach different levels of importance to
  sellers report.

19
Results
K
KTBM
KTBM0.25
KVCG
KTBM0.5
KFTM
20
Conclusions

• Task allocation with uncertainty
• Only incentive-compatible, individually-rational
  efficient reputation mechanism
• Affords flexibility of context based trust
  modeling with efficiency

21
End

• Any Questions?

rkd02r_at_ecs.soton.ac.uk
22
Our Case Study

• Task Allocation
• Some agents may perform some tasks
• for which they incur a cost Ci
• Others may ask for tasks to be performed
• for which they have a valuation Vi
• Eg. File-sharing on P2P networks, Selfish
  Multi-Sensor Nets
• BUT there exists some uncertainty!
• Agents may not perform tasks completely at times
  i.e. there exists a probability of success (POS)
  in executing a task

23
Trust-Based Mechanism

1. Find the optimal allocation given all trust
   reports
2. Find the optimal allocation with the trust report
   of each agent taken out in turn.