Professional Documents
Culture Documents
Course Introduction
Pradeep K. Murukannaiah
Catholijn M. Jonker
Myrthe L. Tielman
Luciano Siebert
• Course organization
• Course contents
2
Course Organization
3
Teaching Team: Lecturers
4
Teaching Team: Course Assistants
• Enrico Liscio (Head TA)
• Florentin Arsene
• Nathan Ordonez Cardenas
• Radek Kargul
• Simran Karnani
• Imko Marijnissen
• Bram Renting
• Carolina Jorge
• Ruben Verhagen
5
Communication Channels
• Stackoverflow (TU Delft instance)
– All content-related questions and discussions
– We count on your contributions!
• Email collabai-cs-ewi@tudelft.nl
– For questions concerning individuals
– Feedback on the course
• Brightspace
– All important announcements
– Discussion forums, e.g., for team formation in the first week
6
Course Organization
7
Hybrid Teaching
• Lectures
– Enroll in the queue for attending lectures on campus
• The queue for a lecture closes 4 days in advance
– Use the fixed Zoom link for attending online
• Labs
– Separate queues for on-campus and online support
– Detailed instructions on Brightspace.
8
Collaborative AI:
What, Why, and How?
9
Activity: Centralized or Collaborative?
11
Artificial Intelligence (AI)
AI is the study and design of Intelligent agents
Intelligent agents:
• Are autonomous
12
Centralized AI Systems
Centralized AI System Single-Agent System
Centralized computation
and control
Source: https://pixabay.com/photos/robot-softbank-pepper-tablet-white-1695653/
Distributed computation
yet centralized control!
Source: https://www.flickr.com/photos/ibm_media/33838065805
13
Collaborative AI Systems
14
Collaborative AI
• Collaborative AI is
the study and design of multi-agent systems
– Example: https://youtu.be/GdUkQObeKbY
15
Collaborative AI: Why? (1)
• Data is decentralized
• Computation is asynchronous
16
Collaborative AI: Why? (2)
17
Collaborative AI:
A Sociotechnical Architecture
18
Collaborative AI: How?
You tell me …
Participate in an activity on
Brightspace Quizzes and Activities
https://brightspace.tudelft.nl/d2l/le/content/399757/v
iewContent/2324105/View
19
What makes engineering a collaborative
AI system challenging?
https://www.menti.com/rjzqvw8ssa
20
Collaborative AI Creates New
Opportunities
Example Use Case: Activity Recommender
User: Alex
Goal: Lead a healthy life
• The two agents can ask the doctor to intervene if serious advice is to
be given (hybrid intelligence)
21
Collaborative AI: Who Cares?
22
Collaborative AI: The Future of AI?
• DARPA Robotics Challenge (2015)
– Promoting innovation in human-supervised robotic technology for disaster-
response
23
Course Content
24
Learning Objectives
After completing this course, you should be able to:
25
Course Contents and Schedule
27
Practical Assignment 1: Negotiating Agent
• Programming language:
Python 3
• Platform: MATRX
• Efficient and
generalizable agents score
high!
29
Reading Material and Resources
• Multiagent Systems, 2nd Edition.
Edited by Gerhard Weiss
– Chapters 1, 2, 4, 6, 8, and 9
– E-book accessible via TUD login
• Lecture slides
30
Summary
• Collaborative AI: What, why, and how?
– Compare centralized and collaborative AI paradigms
• Course organization
31
Negotiation
• Why negotiate ?
• Negotiation Analysis
• Do it yourself !
Why Negotiate?
• Positive situation:
• Win-win opportunities
• Short term relations
• Long term relations
• Conflict:
• Power play (or war) will not lead to a solution
• All parties need a solution
• All parties need to at least pretend to be willing
3
Definitions (1)
Mediated? Open?
5
Bilateral Negotiation
Consumer Provider
Agreement
Agreement 2
Agreement 1
Agreement 3
Service
Consumer
Tuesday, March 8, 2022 7
Multilateral Negotiation
• Negotiation among more than two parties
• All parties mutually agree on the final decision/offer
• A group of service providers would like to give a service together and
they need to mutually agree on some issues.
• A group of friends are planning to go on holiday together and they
need to find a joint agreement on some issues such as location,
duration etc.
Alic
Agreement
Bob e
Mary
… to Win-Win Negotiation
Position - the first demand of a party, often vocal and self-centered. Requests for cash ($) compensation is often a position.
Interest - the deeper motivations underlying a party’s position, and which offers greater room to explore a range of options
from which to compile a solution that satisfies the most number of parties.
9
Source: http://www.snsi.org/negotiation.php
Positional Bargaining
Negotiation: bidding
11
Negotiator Guidelines
12
Negotiator Guidelines
How?
• Ensure that all parties are identified and represented
• Develop understanding through active listening
• Remember that perceptions can be real, since they
may drive decision-making
13
Source: http://www.snsi.org/negotiation.php
Negotiator Guidelines
14
Negotiator Guidelines
15
Negotiator Guidelines
How?
• Adopt consensual negotiation style, not a positional
• Identify peoples underlying concerns, fears,
motivations, deadlines, aspirations & values
• Create room to explore a wide range of options
16
Source: http://www.snsi.org/negotiation.php
Negotiator Guidelines
17
Negotiator Guidelines
How?
• Use creativity and engage in joint problem solving
• Apply brainstorming techniques
• Assess the uncertainty and risks arising from the
proposed solution
18
Source: http://www.snsi.org/negotiation.php
Negotiator Guidelines
How?
• Ensure that parties implement commitments
• Incorporate a grievance mechanism to address
concerns over implementation
• Evaluate resulting relationships and outcomes
19
Source: http://www.snsi.org/negotiation.php
Negotiator Guidelines
20
Phases of Negotiation
Closing: Bidding:
Contract Strategy determination
Relationships Bid evaluation
Reflection Next bid determination
Bidding analysis
21
Definitions
Definitions (2)
Definitions (3)
Prepare!
Prepare (2)
Reciprocate:
Share Information if Other Reciprocates!
26
How To: Joint Exploration
Negotiation Atmosphere
27
Negotiator Guidelines
Negotiation Styles
Mastenbroek’s
advice
Activity Scale
Interests
lenient hard
Power
bending, dominant
subdued
Climate
jovial hostile
personal formal
Flexibility
exploring repetitive
avoiding
30
How To: Evaluate
give take
March 8, 2022 31
Definitions
Utility
• n issues: • 2 issue example
b = (v1, …, vn) G: g1, g2
v1 D1, …, vn Dn H: h1, h2
D = D1 x … x Dn • Agent A
• For every bid b D
EA,1(g1) = 1
Utility of A: UA(b)
EA,1(g2) = 0.5
Utility of B: UB(b)
EA,2(h1) = 0.3
• UX : D [0, 1]
EA,2(h2) = 1
• UX(b) = j wX,jEX,j(b)
wA,1 = 0.7
• wX,j : weight issue j wA,2 = 0.3
• EX,j(b) : evaluation of value of issue j UA((g1, h1)) = 0.7*1 + 0.3*0.3 = 0.7
in bid b + 0.09 = 0.79
How To: Evaluate
Utility (2)
• 2 issue example
• 2 issue example
G: g1, g2
G: g1, g2
H: h1, h2
H: h1, h2
• Agent B
• Agent A
EB,1(g1) = 1
EA,1(g1) = 1
EB,1(g2) = 0.6
EA,1(g2) = 0.5
EB,2(h1) = 1
EA,2(h1) = 0.3
EB,2(h2) = 0.4
EA,2(h2) = 1
wB,1 = 0.4
wA,1 = 0.7
wB,2 = 0.6
wA,2 = 0.3
UB((g1, h1)) = 1
UA((g1, h1)) = 0.79
UB((g2, h1)) = 0.84
UA((g2, h1)) = 0.44
UB((g1, h2)) = 0.64
UA((g1, h2)) = 1
UB((g2, h2)) = 0.48
UA((g2, h2)) = 0.65
How To: Evaluate
Result space
• Result of a bid:
(g1,h2) • R : D ([0, 1], [0, 1])
• R(b) = (UA(b), UB(b))
1
• 2 issue example
(g1,h1) G: g1, g2
(g2,h2)
Utility A
H: h1, h2
• Bid UA UB
(g2,h1) (g1,h1) 0.79 1
(g2,h1) 0.44 0.84
(g1,h2) 1 0.64
(g2,h2) 0.65 0.48
0 Utility B 1
How To: Evaluate
March 8, 2022 36
How To: Evaluate
Comparing results
March 8, 2022 37
How To: Evaluate
Nash product
max(Ubuyer Useller)
0 Seller
38
Definitions
Definitions
• Let UA and UB be the utility functions of party A and
party B. Both functions are defined on the domain D of
all possible bids that the agents can make, mapping D to
the range [0, 1]. Thus,
UA : D [0, 1] and UB : D [0, 1].
• The Pareto Optimal Frontier is the set of bids, such
that there is no other bid that is better for at least one
party, without making things worse for the other parties.
Thus, if two agents A and B are involved,
POF = { b D | b’ D : bb’ ( UA(b’) < UA(b)
UB(b’) < UB(b) ) }
39
Definitions (2)
*It is a bit more technical than this if the domain is not continuous
How To: Evaluate
Result space
• Result of a bid:
(g1,h2) • R : D ([0, 1], [0, 1])
• R(b) = (UA(b), UB(b))
1
• 2 issue example
(g1,h1) G: g1, g2
(g2,h2)
Utility A
H: h1, h2
• Bid UA UB
(g2,h1) (g1,h1) 0.79 1
(g2,h1) 0.44 0.84
(g1,h2) 1 0.64
(g2,h2) 0.65 0.48
0 Utility B 1
Relation to Artificial Intelligence
Artificial Intelligence:
• Intelligent decision making
• Engineering Heuristic Approaches
for Machines
• Engineering Cognitive Skills used
in Negotiation
Russell & Norvig: 2 – 6, 10, 13 - 22
The Future:
• Enhanced negotiations by using the Negotiation Support Systems
42
Some classics on negotiation from a
human perspective
• Harvard Business Review Staff, (2003). Negotiation (Harvard
Business Essentials Series). Harvard Business School Press.
• Lewicki, R.J., Saunders, D.M., and Minton, J.W., (1999),
Negotiation: readings, exercises, and cases. Boston, McGraw-
Hill/Irwin.Raiffa, H. (1982). The art and science of
negotiation, how to resolve conflicts and get the best out of
bargaining. Cambridge, Mass., Belknap Press of Harvard
University Press.
• Shell, G.R., (2000), Bargaining for advantage: negotiation
strategies for reasonable people. Penguin Books.
March 8, 2022 43
Classics on the way to automated
negotiation
44
State of the art and classics for
automated negotiation
• Baarslag T, Fujita K, Gerding EH, Hindriks K, Ito T, Jennings NR, Jonker CM,
Kraus S, Lin R, Robu V, Williams CR, (2013). Evaluating Practical Negotiating
Agents: Results and Analysis of the 2011 International Competition, Artificial
Intelligence, 198, pages:73 - 103, issn: 0004-3702, doi:
10.1016/j.artint.2012.09.004.
http://www.sciencedirect.com/science/article/pii/S0004370212001105?v=s5
• Jonker CM, Hindriks K, Wiggers P, Broekens JD, (2012). Negotiating Agents,
AI Magazine,
33. http://www.aaai.org/ojs/index.php/aimagazine/article/view/2421
• Luo, Xudong, Miao, C, Jennings, Nick, He, Minghua, Shen, Z and Zhang, M
(2012) KEMNAD: A Knowledge Engineering Methodology for Negotiating
Agent Development. Computational Intelligence, 28, (1), 51-105.
45
Supporting humans
• Michele Joy Gelfand and Ya'akov Gal. The brave new negotiating world:
Challenges and opportunities. In B. Goldman & D. Shapiro (Eds.) The psychology
of negotiations in the21st Century Workplace. SIOP Frontiers Book, Routledge,
2013.http://www.routledge.com/books/details/9780415871150/
• Ya'akov Gal, Sarit Kraus, Michele Gelfand, Hilal Khashan, Elizabeth Salmon.
Negotiating with People across Cultures using an Adaptive Agent. ACM
Transactions on Intelligent Systems and Technology 3(1), Article 8, 2011.
http://www.eecs.harvard.edu/~gal/Papers/culture10.pdf
• http://ii.tudelft.nl/negotiation/index.php/Pocket_Negotiator
• Jonker C.M., Aydogan R, Baarslag T, Broekens J, Detweiler C, Hindriks K.V.,
Huldtgren A, Pasman W, (2017). An Introduction to the Pocket Negotiator: A
General Purpose Negotiation Support System. In: Multi-Agent Systems and
Agreement Technologies, Multi-Agent Systems and Agreement Technologies –
Proceedings of the 14th European Conference, (EUMAS 2016), LNAI 10207,
Springer International Publishing, pp:13—27.
• The InterNeg Research Centre: http://interneg.concordia.caa
46
Formalizing Negotiations
* 1
Outline
* 2
Recap
* 3
How To: Evaluate
Nash product
max(Ubuyer ⋅
Useller)
0 Seller
4
Social Welfare (Fairness)
5
When no equal utility solution exists
C
How to set the probabilities?
What will the negotiation be about if you
B
do this? What is a fair outcome then?
0
* 7
Domain model
• For the negotiations we do we need
• A set of issues X, and for each issue x ∈ X a set of
values for each of those issue: V(x)
• Another word for issues is attributes.
• For each stakeholder we need a preference profile for
this domain
• If repeated negotiations in a domain, then apply ML to
build and extend the domain and cluster preferences
to a set preferences that occur most.
• Learn from other negotiating agents by sharing
anonymized negotiation logs.
* 8
Domain Modelling
• Domain modelling is difficult for humans:
• Constructive process: humans become aware of them when
working with the material/problem. For negotiation this
implies that they might realise that something is an issue they
would like to negotiate about when the other party makes a
remark about it.
• Lack of experience in structuring problems
• Lack of formal techniques (math/logic, data structures)
• Can the stakeholder understand your model?
• An exam question might be: construct and explain a
domain model for the following domain. <followed by
an informal description of the domain>
* 9
Preferences
* 11
Preference elicitation
* 12
The Pocket Negotiator
http://ii.ewi.tudelft.nl:8080/PocketNegotiator/index.jsp
• Domain specification
• Preference elicitation
• Underlying interests
• Weights of issues
• Evaluation functions of issues
• Bidding support
• What to bid
• When to stop
* 13
IAGO platform – Group Jon Gratch
* 14
Additional literature
• Preference elicitation
Chen, L., & Pu, P. (2004). Survey of preference elicitation
methods (No. EPFL-REPORT-52659).
* 15
Automated Negotiation
Catholijn M. Jonker
1
Contents of these slides
• Recap: Formalizing Negotiations
o Domain models
o Preferences and preference elicitation
o Analysing results
• Protocols
o bilateral
o Multi-lateral
• Negotiation strategies
o BOA framework
o An overview of important strategies
• Analysis of Negotiation Dynamics
• Using the BOA framework
• Prepare for Tutorial and Test
2
Why negotiation is difficult for
humans?
• Bounded rationality
• Outcome space might be too big.
• Emotions
• May have negative effect on acting rationally
• Some people are too shy to ask what they want.
Saturday, September 12, 2020 3
Automated Negotiation
Consumer Producer
Agent Agent
6
Normalization (2/3)
• Formally: Given issue j with range Dj, and valuation
function vj
• Let m=max({vj(x)| x ∈ Dj}),
• We define the normalized evaluation function
ej : Dj → [0,1] by: ej(x)=vj(x)/m
• Formally:
• Let D = D1 x D2 x …x Dn be the cartesian product of the
ranges of all issues. D is also the bid space.
• Let wj be the weight of issue j, for all j: 1 ≤ j ≤ n.
• ∑1 ≤ j ≤ n (wj = 1)
• Let ej be the normalized evaluation function of issue j.
• The normalized utility function u: D → [0,1] is defined
by: u(b) = ∑1 ≤ j ≤ n (wj * ej(bj) ), where bj is the
projection of b on issue j.
8
NEGOTIATION PROTOCOLS
9
Negotiation Protocol
10
Alternating Offers Protocol (Rubinstein 1982)
Mediator generates an offer and asks negotiating agents for their votes
either to accept or to reject this offer.
Mediator
Accept
Accept ...
Accept
15
Mediated Single Text Negotiation
Hill-Climber Agent (Klein et al., 2003)
• Accept a bid if its utility is higher than the utility of the
most recently accepted bid
Note: If the utility of initial bid is quite high for one of the
agents, that agent may not accept other bids even though
those bids might be better for the majority.
http://ebusiness.mit.edu/research/Briefs/4Klein_Negotiation_Brief_Final.pdf
Mediated Single Text Negotiation:
Annealer Agent (Klein et al., 2003)
• Calculates the probability of acceptance for the current bid:
NEGOTIATION STRATEGIES
18
Negotiation Strategy
• Determines
• which action the agent will take
• how the agent will generate its offers
• how the agent decide whether the opponent’s counter-offer is acceptable
O B A
19
Bidding strategies
B
Random Walker
21
Time-dependent Concession Strategy
[Faratin, Sierra & Jennings, 1998]
• Each agent has a deadline and the agent's behavior changes with
respect to the time.
• An offer which is not acceptable at the beginning, may become
acceptable over time (conceding while approaching the deadline).
• A function determines how much the agent will concede
• Remaining negotiation time
• Parameter related to concession speed (β)
• Conceder Tactic:
• β >1 concedes fast and goes to its reservation value quickly.
• Boulware Tactic:
• β <1 hardly concedes until the deadline
22
Trade-Off Strategy (1)
• Not only considers its own utility but also take its
opponent’s utility into account.
23
Trade-Off Strategy (2)
[Faratin, Sierra, Jennings, AIJ 2002]
O
Opponent Modelling (1) Why?
• Exploit the opponent
• Maximize chance of reaching an agreement
• Requiring outcome with acceptable utility for opponent, i.e. resolving
the conflict of interest.
• Increase the efficiency of a negotiated agreement
• Searching through the outcome space for outcomes that are
mutually beneficial
• Reaching better/optimal agreements
27
Opponent Modeling
Outcome Space
Contrac Pareto
tA frontier
My Utility
Contract
B
31
A Simple Example:
Frequency Analysis Heuristic
• A heuristic adopted by some of the agents in ANAC
competition such as HardHeaded Agent
• Based on how often the value of an issue changed and
the frequency of appearance of values in offers
• Learning issue weights: importance of issues
• Heuristic: If the value is often changed, then the issue gets a
low weight.
• Learning issue value weights: evaluations of the issue values
• Heuristic: A preferred value will appeared more often in agent’s
offers than a less preferred value.
32
Frequency Analysis Heuristic (2)
Estimation of issue weights
33
Frequency Analysis Heuristic (3)
Estimation of evaluation values for issues
34
Acceptance Conditions
A
Introduction
Why and when should we accept?
Example t = 0.7
0.1
0.2
0.3
0.4
0.5
0.6
Acceptance Conditions
Combining Acceptance Conditions
• Acceptance Strategies
Saturday, September 12, 2020 43
Hindriks, Jonker, Tykhonov, 2011
ANALYSIS OF NEGOTIATION
DYNAMICS
Analysis of negotiation strategies
Nash product
max(Ubuyer ⋅ Useller)
0 Seller
1
46
Negotiation traces
Negotiation step
by A
Utility of Agent A
Current Bid
of Agent A
Current Bid
of Agent B
Utility of Agent B
Negotiation traces tA: trace of A
tB: trace of B
t = tA + tB is the complete nego trace
t
A
Utility of Agent A
t
B
Utility of Agent B
Utility, negotiation steps, and traces
concession
Utility of Agent B
unfortunate
Utility of Agent B
Fortunate step denoted by (S+, O+), s is an
unfortunate step iff:
Utility of Agent A
fortunate
∆S(s)>0, and ∆O(s)>0.
Utility of Agent B
Selfish step denoted by (S+, O≤), s is a selfish
step iff:
Utility of Agent A
selfish ∆S(s) > 0, and ∆O(s) ≤ 0.
Utility of Agent B
Silent step denoted by (S=, O=), s is a silent
step iff:
Utility of Agent A ∆S(s) = 0, and ∆O(s) = 0.
Utility of Agent B
Nice step denoted by (S=, O+), s is a nice
step iff:
Utility of Agent A ∆S(s) = 0, and ∆O(s) > 0.
Utility of Agent B
Classification of negotiation steps
Step
Classes
of Agent A
selfish fortunate
silent
nice
Utility of Agent A
unfortunate concession
Current Bid
of Agent A
Utility of Agent B
Classification of negotiation steps
Step
Classes
of Agent A
selfish fortunate
silent
Utility of Agent A
nice
Steps of
unfortunate concession Agent B
Current Bid
nice
concession fortunate
of Agent A
silent
Utility of Agent B
Sensitivity to Opponent Preferences
A rational negotiator would try to make
fortunate, nice, or concession steps.
• Overall utility:
• ABMP 0.72,
• Trade-Off 0.74, and
• Random Walker 0.69.
• Trade-Off:
• Outperforms ABMP on the SON domain with complete
information and on the AMPOvsCity domain;
• Underperforms wrt ABMP on the second hand car domain due
to wrong weights and unpredictable issues;
• ABMP:
• Strong on the second hand car domain;
• Underperforms on the SON domain.
Conclusions
• Want to negotiate efficiently? Know your partner!
O B A
Applying the BOA Framework
Combining components
Accept as late as
Accept early possible
AC next discount
AC next discount Accept very good
Frequency model (1.0; 0.0; 1.0, 0.0)
(1.0;Accept above
0.0; 1.0, 0.0)a bids only
No opponent model Bayesian learning fixed threshold Accept at the last
Accept the best Never accept
moment
Bayesian learning
No opponent offer so far
Frequency model
model
O’s A’s
73
Required reading
74
References (1)
• Baarslag, T., Aydogan, R., Hindriks, K.V., Jonker, C.M., Fujita, K., & Ito, T.,
(2015). The Automated Negotiating Agents Competition, 2010-2015, AI Magazine,
36, pp:115-118.
• Tim Baarslag, Alexander Dirkzwager, Koen Hindriks, and Catholijn Jonker. The
significance of bidding, accepting and opponent modeling in automated
negotiation. In 21st European Conference on Artificial Intelligence, 2014.
• Baarslag T, Fujita K, Gerding EH, Hindriks K, Ito T, Jennings NR, Jonker CM, Kraus
S, Lin R, Robu V, Williams CR, (2013). Evaluating Practical Negotiating Agents:
Results and Analysis of the 2011 International Competition, Artificial Intelligence,
198, pages:73 - 103, issn: 0004-3702, doi: 10.1016/j.artint.2012.09.004.
http://www.sciencedirect.com/science/article/pii/S0004370212001105?v=s5
• Jonker CM, Hindriks K, Wiggers P, Broekens JD, (2012). Negotiating Agents, AI
Magazine, 33. http://www.aaai.org/ojs/index.php/aimagazine/article/view/2421
• Luo, Xudong, Miao, C, Jennings, Nick, He, Minghua, Shen, Z and Zhang, M (2012)
KEMNAD: A Knowledge Engineering Methodology for Negotiating Agent
Development. Computational Intelligence, 28, (1), 51-105.
75
References (2)
Klein, M., et al., Protocols for Negotiating Complex Contracts. IEEE
Intelligent Systems, 2003. 18(6): p. 32 - 38.
http://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber=1249167&filter%3DAND%28p_IS_Num
ber%3A27968%29
but easily accessible:
http://ebusiness.mit.edu/research/Briefs/4Klein_Negotiation_Brief_Final.pdf
Rosenschein, J.S., and G. Zlotkin. Rules of Encounter. MIT Press,
Cambridge, MA, 1994. 16.
Rubinstein, A., Perfect equilibrium in a bargaining model. Econometrica,
50:97–109, 1982. http://arielrubinstein.tau.ac.il/papers/11.pdf
Raiffa, H., The art and science of negotiation. 1982, Cambridge, Mass.:
Belknap Press of Harvard University Press. x, 373.
Baarslag T. 2014. What to Bid and When to Stop. Delft University of
Technology. http://mmi.tudelft.nl/sites/default/files/thesis.pdf
Hindriks KV, Jonker CM, Tykhonov D, (2011). Let's dans! An analytic
framework of negotiation dynamics and strategies, Web Intelligence and
Agent Systems, 9, pages:319-335 (see reading material on blackboard for
the paper) 76
Computational Coalition Formation
Pradeep K. Murukannaiah
1
Collaborative AI: Learning Objectives
• Compare centralized and collaborative AI paradigms
2
Learning Objectives for Today
3
Reading Material and Resources
• Multiagent Systems, 2nd Edition.
Gerhard Weiss
– Chapter 8
• Sections 1, 2, 5.1, 5.2, and 5.3
• A research paper by
Rahwan et al., JAIR, 2009
– Sections 1--4
– Brightspace: Resources and tools
• Lecture slides
4
What is an Agent?
Hello…
I am an intelligent entity….
I am situated in an environment…
I can perceive the environment
I have goals…
I have a list of available actions…
I act autonomously to satisfy my goals
and…
most importantly…
I am social !!!
5
Example: the Prisoner’s Dilemma
6
What should a “Rational” Agent do in the
Prisoner’s Dilemma?
https://www.menti.com/2ommwhq7ew
7
Prisoners’ Dilemma: Matrix
Representation
a2 quiet confess
a1
(-1,-1) (-4, 0)
quiet
(0, -4) (-3, -3)
confess
8
Prisoners’ Dilemma: the Rational
Outcome
• a1’s reasoning: a
a1 2 Quite Confess
9
Cooperative vs. Non-Cooperative Games
• Game theory studies interactions between agents in
situations known as games
10
Coalition Game Theory
11
Organizations in Multiagent Systems
Hierarchies
Compounds
Markets
MatrixControl
Organizations
Teams Coalitions
Holarchies
Congregations
Federations
Societies Data
12
Organizations in Multiagent Systems
Hierarchies
Compounds
Markets
Matrix
Organizations
Teams Coalitions
Holarchies
Congregations
Federations
Societies seller buyer
13
Organizations in Multiagent Systems
Hierarchies
Compounds
Markets
Matrix
Organizations
Teams Coalitions
Holarchies
Congregations
Federations
Societies
14
Agent Coalitions
Main characteristics
Coalitions in general are goal-directed and short-lived
No coordination among members of different coalitions
The organizational structure within each coalition is flat
15
Why Agents form Coalitions?
Electronic-commerce
Cooperation among buyers to
obtain quantity discounts, and
sellers to maintain pricing.
Disaster Management
UN report said: “Efforts by the
United Nations in Haiti have
lacked sufficient coordination”
17
Coalition Formation: Applications
Distributed sensor networks:
Coalitions of sensors can work together to track
targets of interest [Dang et al. 2006]
Information gathering:
Several information servers can form coalitions to
answer queries [Klusch and Shehory, 1996].
18
Coalition Formation Games
Partition Function
Does a coalition Yes Game (PFG)
influence other
co-existing
coalitions? No Characteristic
Function Game
Cooperative (CFG)
Game
Transferable
Yes Utility (TU)
Can a player
Game
transfer part of its
utility to
another? No Non-Transferable
Utility (NTU) Game
19
Example: Buying Ice-Cream
• n children, each has some money:
• Supermarkets sells many ice-cream tubs, in different sizes:
– Type 1 contains 500g, costs $7
– Type 2 contains 750g, costs $9
– Type 3 contains 1kg, costs $11
• Children have utility for ice-cream, and don‘t care about money
• The payoff of a group is the maximum amount of ice-cream the members
of the group can buy by pooling their money
20
Example: Writing Papers
• n researchers working at n different universities can form
groups to write papers
• the composition of a group determines the quality of the
paper they produce
• each author receives a payoff
from his own university:
– promotion
– teaching load reduction
21
Example: Growing Fruits
22
What types of coalition formation games
are the “writing papers” and “growing
fruits” examples?
https://www.menti.com/q2its8o42j
23
Coalition Formation Games
Partition Function
Does a coalition Yes Game (PFG)
influence other
co-existing
coalitions? No Characteristic
Function Game Focus
Cooperative (CFG) of this
Game
lecture
Transferable
Yes Utility (TU)
Can a player
Game
transfer part of its
utility to
another? No Non-Transferable
Utility (NTU) Game
24
Transferable Utility Games Formalized
• A transferable utility game is a pair (A, v), where:
– A ={a1, ..., an} is the set of players (or agents)
– v: 2A → R is the characteristic function
• for each C ⊆ A, v(C) is the value of C, i.e., the payoff that
members of C can attain by working together
– Usually, it is assumed that
• v(Ø) = 0
• v(C) ≥ 0 for any C ⊆ A
• v(C) ≤ v(D) for any C, D such that C ⊆ D
• The biggest possible coalition (the one containing
all agents) is called the grand coalition
25
Ice-Cream Game: Characteristic Function
C: $6, M: $4, P: $3
26
Transferable Utility Games: Outcome
An outcome of a TU game G =(A, v) is a pair (CS, x), where:
27
Coalition Formation Process
optimal?
Coalition structure
generation
Value
Value Value
Payoff
distribution
Value
28
Coalition Structure Generation in CFGs
Given 3 agents, the set of agents is:
{a1,a2,a3}
Exercise
What is the optimal
coalition structure ?
Answer
{ {1}, {2}, {3,4} }
30
value L6 value L5 value L4 value L3 value L value L
2 1
Exercise
What is the optimal coalition
structure ?
31
Complexity of Coalition Structure Generation
• Bell number
• ω(nn/2)
• O(nn)
32
Coalition Structure Generation Algorithms
33
Dynamic Programming Algorithm
Main observation: To find the optimal partition of a set of agents, it is sufficient to:
• Try the possible ways to split the set into two sets, and
• For every half, find the optimal partition of that half.
34
f Best split Evaluations performed before setting f coalitionsize
{1} V({1})=30 {1} 301
{2} V({2})=40 {2} 40
{3} V({3})=25 {3} 25
{4} V({4})=45 {4} 45
V({1,2,3,4})=140 f({1})+f({2,3,4})=150 4
{1,2,3,4} {1,2} {3,4} 150
f({2})+f({1,3,4})=150 f({3})+f({1,2,4})=145
f({4})+f({1,2,3})=145 f({1,2})+f({3,4})=150
f({1,3})+f({2,4})=145 f({1,4})+f({2,3})=145 35
Improved Dynamic Programming Algorithm (IDP)
36
The Coalition Structure Graph
{a1},{a2},{a3},{a4}
optima
l
{a1},{a2},{a3,a4} {a3},{a4},{a1,a2} {a1},{a3},{a2,a4} {a2},{a4},{a1,a3} {a1},{a4},{a2,a3} {a2},{a3},{a1,a4}
{a1,a2,a3,a4}
37
Coalition Structure Generation Algorithms
38
Basic Idea of IP
Search space
Upper bound = 300
Upper bound = 250
Sub-space
Lower bound = 100
Upper bound = 600
39
Basic Idea of IP
S
Upper bound [4]
= 550 {{a1, a2, a3, a4}}
Sub-space 500 Upper bound = 450
{{abound
1,a2}, ={a200
3,a4}} Sub-space
S [2,2] Lower
{{a1,a3}, {a2,a4}} {{a=1},250{a2}, {a3,a4}}
Lower bound
40
How the Bounds are Computed
value L4 value L3 value L2 value L1
425 {1, 2, 3, 4} 200 {1, 2, 3} 175 {1, 2} 125 1
150 {1, 2, 4} 150 {1, 3} 50 2
Max4 = 425
300 {1, 3, 4} 100 {1, 4} 75 3
Avg4 = 425
150 {2, 3, 4} 150 {2, 3} 150 4
200 {2, 4}
Max1 = 150 Max3 = 300
125 {3, 4}
Avg1 = 100 Avg3 = 200
Max2 = 200
Avg2 = 150 { {a1}, {a1, a2, a3} }
{ {a1}, {a1, a2, a4} }
S [4] {{a1, a2, a3, a4}}
{ {a1}, {a1, a3, a4} }
Avg=425 UB=425 { {a1}, {a2, a3, a4} }
[1,1,1,1,2,2] [1,1,1,1,1,3]
Avg=500 UB=650 Avg=320 UB=400
[1,1,1,1,1,1,2]
Avg=360 UB=390
? ?
[2,3,3] c h
[2,2,4]
r
[1,3,4]
Avg=500a UB=700
[1,2,5] [1,1,6]
Se
Avg=350 UB=400 Avg=450 UB=575 Avg=390 UB=480 Avg=460 UB=510
?
[2,2,2,2]
Avg=520 UB=600
[1,2,2,3]
Avg=450
[1,1,2,2,2]
UB=520
[1,1,3,3]
Avg=440 UB=480
[1,1,1,2,3]
?
[1,1,2,4]
Avg=550 UB=620
[1,1,1,1,4]
[1,1,1,5]
Avg=520 UB=540
? c h
[1,1,1,1,2,2] [1,1,1,1,1,3]
a r UB=650
Se
Avg=500 Avg=320 UB=400
[1,1,1,1,1,1,2]
Avg=360 UB=390
Property
Algorithm
IDP IP
Worst case
n
performance O(3 ) O(nn)
Return solutions
anytime False True
Time to return
optimal solution Slow Fast
44
Summary
45
http://www.ihmc.us/users/mjohnson/publications.html
Seven “deadly” myths of autonomy
1. “Autonomous systems” are autonomous
2. Autonomy is unidimensional
3. The conceptualization of “levels of
autonomy” is a useful scientific
grounding
4. Autonomy is a widget
5. Once “achieved,” “full autonomy”
obviates the need for human-machine
collaboration
6. As machines acquire more “autonomy,”
they work as simple multipliers of
human capability
7. “Full autonomy” is not only possible, but
always desirable
-
Bradshaw, J.M, Robert R. Hoffman, Matthew Johnson, and David D. Woods. The Seven Deadly Myths of
"Autonomous Systems.” IEEEAutonomy
Intelligent Systems, May/June 2013 (vol. 28 iss. 3), pp. 54-61.
-
Lessons learned about autonomy
1. Autonomy != Replacement
Dependent - Independent
Autonomy
-
Lessons learned about autonomy
1. Autonomy != Replacement
Dependent - Independent
- Inter
Autonomy Teamwork
- -
Dependent - Independent
- Inter
Autonomy Teamwork
- -
Dependent - Independent
- Inter
Autonomy Teamwork
- -
Dependent - Independent
- Inter
Autonomy Teamwork
- -
• “improve the teaming of unmanned systems with the manned force” -
Unmanned Systems Integrated Roadmap FY2011-2036
Dependent - Independent
- Inter
Autonomy Teamwork
- -
What is Coactive Design?
• Coactive Design is about designing
human-robot systems that support
interdependence.
Interdependence
Observability
Predictability
Directability
What is the robot going Mutual Predictability What does the human
to do next? need from me?
How can we get the robot Mutual Directability Can the human provide
to do what we need? help?
How to design for
interdependence?
• Coactive Design is a design method
that helps design support for
interdependence and understand the
impact of change.
Selection and
implementatio
n process
Evaluation of
change process
Identification
process
Selection and
implementatio
n process
Evaluation of
change process
IA Table Example for Collaborative Control
IA Table Example for Collaborative Control
IA Table Example for Collaborative Control
Interdependence Analysis (IA) Table
notification predictability
Identification
process
Selection and
implementatio
n process
Evaluation of
change process
The mechanisms that
support an interdependent relationship
are the creative medium of the designer.
Selection and
implementatio
n process
IA Table Example for Collaborative Control
notification predictability
Next we will get more
complex…
IA Table Example for DRC
Hose Task IA Table
Hose Task IA Table
1
TRUST
2
TRUST
What is trust, and why is it relevant for Collaborative AI
Dimensions of trust
Evaluating Trust
Trustworthy AI
Explainability
This week: Trust concepts from the human-human & human-AI perspective
Next week: Trust modelling from the AI-AI & human-AI perspective
3
WHY TRUST?
What is Trust:
“if A believes that B will act in A’s best interest, and accepts vulnerability to B’s actions, then A
trusts B 1”
1 Roger C Mayer, James H Davis, and F David Schoorman. 1995. An integrative model of organizational trust. Academy of management
review 20, 3 (1995), 709–734.
4
WHEN DO WE NEED TRUST?
All circumstances in which we are in any way
dependent on other’s actions
5
BACK TO AI?
6
RISK
Do we need to trust if there is no vulnerability?
8
TRUST AS A CONTRACT?
I trust X to do Y in situation Z
10
TRUST VS. TRUSTWORTHINESS
Warranted trust
Trust is based on information about actual trustworthiness
Appropriate trust
Trust is appropriate given trustworthiness
Trust calibration
The process of calibrating trust to match trustworthiness
If successful: leading to warranted appropriate trust.
11
WARRANTED/UNWARRANTED TRUST
Unwarranted trust/distrust :
Trust does not depend on trustworthiness
Can lead to inappropriate trust:
Distrust in a trustworthy system
Trust in an untrustworthy system
Warranted trust/distrust :
Trust depends on trustworthiness
Expected to lead to appropriate trust:
Trust in a trustworthy system
Distrust in an untrustworthy system
Inappropriate distrust
- Disuse
Missed signals
Missed opportunity
14
DIMENSIONS OF TRUST
Trustworthiness:
Someone is trustworthy wrt some contract if they will maintain this contract
However, when we say we don’t trust Facebook not to share our data with 3rd
parties:
Do we mean we think they aren’t capable of not doing this?
Or do we believe they are not willing to not do this?
16
ABI MODEL
Factors of perceived
trustworthiness:
Ability
Ability is that group of skills, competencies,
and characteristics that enable a party to
have influence within some specific domain.
Benevolence
Benevolence is the extent to which a trustee
is believed to want to do good to the trustor,
aside from an egocentric profit motive.
Integrity
The relationship between integrity and trust
involves the trustor's perception that the
trustee adheres to a set of principles that
the trustor finds acceptable
17
ABI Johan has a new smart watch. However, he values his
privacy and he doesn’t really trust the smart watch to
not share his personal data with third parties.
Therefore he doesn’t really wear the watch.
Propensity to trust:
Quality of the trustor, irrespective of the
trustee
How likely are you, as a person, to trust?
19
ABI MODEL
Risk taking in relationship
Behavior based on trust
But also on perceived risk!
20
EVALUATING TRUST
Behavior
Risk taking behaviour / Reliance
Perceived
Trust
Risk
Given ABI, to properly estimate trust from behaviour, you need to
Know perceived risk
Control for it!
21
EVALUATING TRUST
Trust Many scales exist!
Perceived Trustworthiness Trust in organisations
Perceived = Subjective Trust in economics
Trust in 1 other person
Subjective measures
Trust in automation
Questionnaires
22
QUESTIONNAIRES ON TRUST
Multidimensional?
ABI is not the only model!
23
QUESTIONNAIRES ON TRUST
(1) I believe that there could be negative consequences when using
Willingness, Benevolence (—)
(2) I feel I must be cautious when using (—)
(3) It is risky to interact with (—)
(4) I believe that (—) will act in my best interest
Gulati et al. 2019 (HCI trust scale) (5) I believe that (—) will do its best to help me if I need help
Perceived Risk (6) I believe that (—) is interested in understanding my needs and
Benevolence preferences
Competence (7) I think that (—) is competent and effective in (—)
(8) I think that (—) performs its role as (—) very well
Reciprocity
(9) I believe that (—) has all the functionalities I would expect from
(—)
(10) If I use (—), I think i would be able to depend on it completely
(11) I can always rely on (—) for (—)
(12) I can trust the information presented to me by (—)
24
TRUSTWORTHY AI?
Trust
Perceived trustworthiness
Trustworthiness?
Trustworthy AI
Trustworthiness as general concept
What does it mean to be benevolent, competent and have integrity?
25
IEEE STANDARD – ETHICALLY ALIGNED DESIGN
“Ultimately, A/IS (autonomous and intelligent systems) should deliver services that
can be trusted.
26
GENERAL PRINCIPLES AS IMPERATIVES
We offer high-level General Principles in Ethically Aligned Design that we consider to be imperatives for creating and
operating A/IS that further human values and ensure trustworthiness. In summary, our General Principles are:
1. Human Rights – A/IS shall be created and operated to respect, promote, and protect internationally recognized human
rights.
2. Well-being – A/IS creators shall adopt increased human well-being as a primary success criterion for development.
3. Data Agency – A/IS creators shall empower individuals with the ability to access and securely share their data, to
maintain people’s capacity to have control over their identity.
4. Effectiveness – A/IS creators and operators shall provide evidence of the effectiveness and fitness for purpose of A/IS.
5. Transparency – The basis of a particular A/IS decision should always be discoverable.
6. Accountability – A/IS shall be created and operated to provide an unambiguous rationale for all decisions made.
7. Awareness of Misuse – A/IS creators shall guard against all potential misuses and risks of A/IS in operation.
8. Competence – A/IS creators shall specify and operators shall adhere to the knowledge and skill required for safe and
effective operation. 27
TRUSTWORTHY AGENTS?
1. Human Rights
2. Well-being I would trust my collaborative agent to…
3. Data
4. Effectiveness
5. Transparency
6. Accountability
Some are easier than others, but we have a
7. Awareness of Misuse way to go to achieve them all!
8. Competence
28
QUESTIONS
29
EXPLAINABLE AI
31
EXPLAINABILITY & TRUST
One of explainable AI (XAI)’s main objectives is to promote trust
Because:
- We need appropriate trust in human-AI systems
- For warraappropriate trust, we need to understand trustworthiness
- To understand trustworthiness, we need to understand if AI systems will
stick to the trust contract
- So we need to understand the AI
32
XAI & TRUST
Use case:
Search & Rescue drone to scout dangerous buildings
- To understand trustworthiness, we need to understand if AI systems will stick to the trust contract
Will the drone provide truthful info about the lay-out of the building?
33
XAI
If the goal is to promote warranted trust, an explanation should give insight into:
Whether a system will fulfil a contract
E.g.
I am 90% reliable on classifying pictures of cats and dogs
I have been trained on pictures of high quality
I am not capable of doing anything else
34
XAI
Explanation: “the details or reasons that someone gives to make something clear or easy to understand”1
So the other needs to understand something!
So you need to understand the other
1 https://dictionary.cambridge.org/es/diccionario/ingles/explanation
2 Miller 2019
35
XAI
36
EXPLANATIONS
• Contrastive
• Selected
• Causality over probability
• Social
37
EXPLANATIONS
• Contrastive Why did the branch fall of the tree?
• Selected
• Causality over probability 1. Branches have a .05% chance of
breaking
• Social
2. Because there was a strong wind
38
EXPLANATIONS
• Contrastive Why did Elisabeth open the door?
Instead of leaving it closed
• Selected
Instead of turning on the airco
• Causality over probability
• Social
1. Because she was hot
2. Because the air-conditioner wasn’t
working
39
EXPLANATIONS
• Contrastive How did you boil that egg?
• Selected
1. I put it in boiling water for 5 minutes
• Causality over probability
• Social
2. I took the egg out of the fridge, put it on
the countertop, filled a pan with water,
put on the stove, waited until the water
was boiling, put the egg in carefully,
waited for 5 minutes, took the pan off
the heat, took out the egg with a spoon,
drained the water, and put the egg on a
plate.
40
EXPLANATIONS
• Contrastive Sorry I was late, I got stuck in traffic this
morning.
• Selected
• Causality over probability
Please put on your coat, it’s cold today.
• Social
Goal:
Create shared understanding
Creating trust:
- Persuasion that a decision was correct might sometimes be better than to achieve
full shared understanding.
Do keep the long-run in mind though!
42
EXPLANATORY QUESTIONS
43
EXPLAINING YOURSELF
Explanations are causal
XAI attempts to explain itself / it’s own reasoning
So:
XAI agents should be able to reason about their own causal model
44
CAUSALITY & COUNTERFACTUALS
Regularity theory:
- There is a cause between event type A and event type B if :
events of type A are always followed by events of type B
Probabilistic theories:
- Event type E causes an event type F iff:
The occurrence of an event type E increases the probability of F occurring
45
CONTRASTIVE
The agent giving the explanation should understand the counterfactual case
Explanations are asked for when unexpected or abnormality is detected.
The counterfactual case is the expected / normal
48
FINDING THE FOIL?
To understand what is ‘normal & expected’, you need to understand either:
The world
The user
49
SELECTING EXPLANATIONS
Causal chains
To give an explanation, one should first know some causes for what needs to be explained
Because his heart failed – Because he had poison in his body – Because he drank poison – Because
someone put poison in his drink – Because he had betrayed them
51
SELECTING EXPLANATIONS
Causal chains
To give an explanation, one should first know some causes for what needs to be explained
“This theory rests on two ideas. The first is that the effect or the explanandum, i.e. the event to be explained,
should be construed, not as an object’s having a certain property, but as a difference between objects with
regard to a certain property. The second idea is that selection and weighting of causes is determined by
explanatory relevance.” [Emphasis from the original source] — Hesslow[69, p. 24].
52
SELECTING EXPLANATIONS
Causal chains
To give an explanation, one should first know some causes for what needs to be explained
Because his heart failed – Because he had poison in his body – Because he drank poison – Because
someone put poison in his drink – Because he had betrayed them
53
SELECTING EXPLANATIONS
Causal chains
To give an explanation, one should first know some causes for what needs to be explained
54
SELECTING EXPLANATIONS
Causal chains
To give an explanation, one should first know some causes for what needs to be explained
Because his heart failed – Because he had poison in his body – Because he drank poison – Because
someone put poison in his drink – Because he had betrayed them
55
SELECTING EXPLANATIONS
Causal chains
To give an explanation, one should first know some causes for what needs to be explained
Responsibility
An event considered more responsible for an outcome is likely to be judged as a better explanation than
other causes
56
EVALUATING EXPLANATIONS
1. Coherence
2. Simplicity
3. Generality
4. Truth
5. Probability
57
SOCIAL EXPLANATIONS
Explanations are a part of conversation.
So they typically follow the rules of conversation:
Quality
Say what you believe to be true
Quantity
Say only as much as is necessary
Relation
Say only what is relevant
Manner
Say it in a nice way
58
SOCIAL EXPLANATIONS
Theory of mind
Understand what the other already knows
Why did the man die?
When asking the coroner: it was poison
When, afterwards, asking the perpetrator: because he betrayed me
Social norms
Norms govern what we ‘should’ or ‘should not’ do
Prohibitions, Obligations, Permissions
Fact and foil?
Norms relate what is socially expected
“I will not shake your hand today, as I have a cold”.
59
QUESTIONS
60
Trust
“if A believes that B will act in A’s best interest, and accepts
vulnerability to B’s actions, then A trusts B”
Trust as a contract
Dimensions of trust
ABI Model
Trustworthy AI
XAI
Explanations
Causality & Contrastive
Evaluating explanation
61
COLLABORATIVE AI 8. - Trust
1
TRUST
2
TRUST IN MAS
Goal: Social control mechanism
3
REPRESENTING TRUST
To keep in mind when judging a representation:
Inspired by social concepts
But with a goal (e.g. security)
Expressiveness
Simplicity
4
REPRESENTING TRUST - BOOLEAN
No trust Trust
5
REPRESENTING TRUST - BOOLEAN
No trust Trust
Simple
Easy to base decisions on
Not very expressive
Not very fine-grained
6
REPRESENTING TRUST - NUMBERS
Integer scale
Interval
But also think about
-1 to 1
What does 0 mean?
-5 to 5
Is 0,5 good?
0 to 10 Is 0,6 better in an important way?
When do you act in what way?
7
REPRESENTING TRUST - NUMBERS
Integer scale Allows for direct comparison of higher-
lower trust
Interval
More fine-grained than Boolean
-1 to 1
What does trust of 0 mean? – hard to
-5 to 5 interpret for people
0 to 10 How do you get these numbers? – hard
to establish for people
8
REPRESENTING TRUST - LABELS
Trust is by nature a vague concept.
What really is the difference between 0,6 and 0,7?
9
REPRESENTING TRUST - LABELS
Easier for people to interpret than
Trust is by nature a vague concept. numbers
What really is the difference between 0,6 and 0,7? More fine-grained comparison than
Boolean
Labels might represent this better? Are distances between labels equal? –
more difficult to compare than numbers
{bad, neutral, good}
{very_bad, bad, neutral, good, very_good}
10
REPRESENTING – UNCERTAINTY & VAGUENESS?
When we talk about trust, we’re often not precise
How much do you trust your doctor on a scale from 0 to 10?
11
PROBABILITY DISTRIBUTIONS
Probability distribution over
sorted discrete sets
Represent uncertainty
12
FUZZY SETS
Set where the membership is ‘fuzzy’, e.g. as represented with a truth level/membership level
Your reputation is 30 out of 100. R=30
Set: not very trustworthy, membership range 0-1
Truth value of you not being very trustworthy: 0,7
Fuzzy sets don’t reason about uncertainty, they reason about vagueness!
Probability:
0,75 p of you being trustworthy
Fuzzy set
3/10 reputation: {not very trustworthy, membership value = 0.7}
13
FUZZY SETS
Often help to describe meaning of
language
Very much
Somewhat
A little
14
PROBABILITY VS FUZZY
John is very tall to huge, which means he is either 1,9m or 2m tall
{very tall, 0.5},{huge, 0.5}
Very tall = 1,9m & Huge = 2m
John is 1,95 tall. This is referred to as either very tall, or huge depending on who you ask
{very tall, 0.5},{huge, 0.5}
John is 1,95m, half of the people would describe this as very tall, the other half as huge
Probability: uncertainty
Fuzziness: vagueness in language
15
PROBABILITY VS FUZZY
Good at representing uncertainty Good at representing vagueness
Usable with both integers or labels Can capture language use
Easy to interpret
Difficult to understand for many
More complex than other options How do you establish truth values?
How do you establish uncertainty?
16
HOW WOULD YOU REPRESENT?
You are building a team manager robot which will work in a storage center. This manager robot is in charge of
delegating tasks to a diverse team of different robots (there are no humans involved in the team). It uses the concept
of trust to do this.
For every trust decision, the robot wants to use the simplests representation of trust which still expresses what it needs.
1. In one situation, the manager robot is meeting another robot, and it needs to decide whether or not to trust this other robot for this
task. Given the story above, what representation of trust should the manager robot use and why?
2. In another situation, the manager robot needs to decide between two different robots. It needs to choose which robot it trusts to
do the task more. Given the story above, what representation of trust should the manager robot use and why?
3. In a third situation, the robot needs to communicate to an external human regulatory operator about its team. It needs to
communicate about how much it trusts its different robot team mates. The robot isn’t always sure of how trustworthy the other
robots are, however. Given the story above, what representation of trust should the manager robot use and why?
4. In a fourth situation, the manager robot has a new job in a new storage center, which does include human teammates. The
previous manager (human), debriefs the robot manager about the robot team. They tell the robot manager that robotA is
somewhat untrustworthy, that robotB is very trustworthy, and that robotC is extremely untrustworthy. Give a representation of the
trustworthiness of these three robots using fuzzy sets. Describe your truth value and trust value, and why you chose those values
for that specific robot. You can determine the ranges yourself, but mention them and give a justification for your choice.
17
BDI - TRUST
Belief
Desire
Intention
Trust: “an agent i trusts another agent j in order to do an action α with respect to a
goal ϕ” 1
19
REPAGE
We can infer that:
BDI RepAge
Representing reputation as beliefs
Probability distribution S(buy(j), V BadProduct, 0.6, seller)
S(buy(j),BadProduct, 0.1, seller)
Reputation value: S(buy(j),OKProduct, 0.1, seller)
S(buy(j),GoodP roduct, 0.1, seller)
Rep( j, seller, [0.6,0.1,0.1,0.1,0.1])
j = the agent S(buy(j), V GoodProduct, 0.1, seller)
seller = role of the agent
Probability defined over the discrete set: S = what people/agents say
{VBadProduct,BadProduct,OKProduct,GoodProduct,VGoodProduct} So subjective!
21
TRUST PROCESSES IN MAS
MAS = Multi-Agent Systems
Trust decision
Trust as an action
I will rely on you to do X
22
TRUST PROCESS IN MAS
Image (Conte & Paolucci), “an evaluative
belief; it tells whether the target is good or
bad with respect to a given behaviour”
Direct experiences
Information about interactions between trustor and trustee
Communicated experiences
Information about interaction between trustee and other agent
Social information
What is the social position of the trustee?
Reputation:
“what a social entity says about a target regarding
his or her behaviour”
(often): Communicated image rather than experience!
23
TRUST PROCESS IN MAS
Self motivations
Stakes
Risk tolerance
Context
The current context is different from ‘normal’ (from
past & communicated experiences)
Good car mechanic, but right now I see a long cue
24
ABI VS MAS?
25
ABI VS MAS?
Willingness & Competence?
26
TRUST EVALUATION
How do we incorporate what input?
Filtering
Context
Are the experiences relevant for the current context?
Information source
Is this source reliable
Own experiences similar to source? (use distance as reliability measure)
Correlated evidence?
Many communicate about the same experience
Experiences are typically not identified!
Consider social circles?
27
AGGREGATION
Discrete labels:
9 & 10 = very good
8 & 7 = good
6 & 5 = okay
4 & 3= bad
1 & 2 = very bad
28
AGGREGATION
With numbers, you can do more!
Average?
Input from others: (4+7+8+6+7+7+7)/7
Weighted average
Recency of input
Reliability of the source
One per social cluster?
29
TRUST BELIEFS
DispTrust(Alice,Tom, inform(weather),
Don’t necessarily arise from calculations know(Alice,weather), asked(Alice,
Tom,weather))
Alice trusts Tom to inform her about the
Rather, they are defined! weather (goal: Alice knows about the weather) if
Trust evaluation belief (dispositional trust) Alice asks Tom about this.
Trust decision belief
30
TRUST BELIEFS - EXAMPLE
ForTrust Model:
Dispositional trust:
i = trustor PotGoali(𝜑 ,K) = trustor i has the potential goal 𝜑 given circumstances K
j = trustee G 𝜑 = globally 𝜑 (always, temporally)
a = action F 𝜑 = eventually 𝜑 (¬G¬ 𝜑)
𝜑 = goal Choicei𝜑 = i has chosen goal 𝜑
K = conditions
Afterj:a;𝜑 = immediately after j does a 𝜑 holds
31
TRUST BELIEFS - EXAMPLE
ForTrust Model:
Dispositional trust:
i = trustor
j = trustee
a = action
𝜑 = goal
K = conditions
32
TRUST DECISION
Trust evaluations lead to trust decisions
Formalization of evaluation directs formalization of
decision
33
QUESTIONS
35
REPUTATION
Related to trust, though slightly different
But many definitions exist!
36
REPUTATION
Reputation is typically used to calculate trust if direct experience isn’t available
Reputation is subjective
It depends on the opinions of others!
Option:
Calculate reliability value
Number of opinions it is based on
Variance of opinions it is based on
Recency of the opinions
Credibility of the agents behind the opinions
37
REPUTATION
Option: agent(time, agent(time,
Reputation (Mean) = 6,1
Calculate reliability value Nr. Agents : 9
opinion, opinion,
Number of opinions it is based on Variance: 5,88
credibility) credibility)
Variance of opinions it is based on Recency (now t=5): 2,6
Recency of the opinions Credibility agents: 6,7
a1(1, 2, 7) a1(2, 6, 5)
Credibility of the agents behind the
opinions a2(1, 4, 8) a2(3, 7, 8)
a3(1, 4, 6) a3(3, 8, 4)
Reputation (Mean) = 6,8
a4(2, 5, 7) a4(4, 5, 5)
Nr. Agents: 5
a5(2, 9, 8) a5(4, 8, 7)
Which reputation value is Variance: 1,7
a6(3, 5, 5)
Recency (now t=5): 1,8
more reliable? a7(3, 8, 7)
Credibility agents: 5,8
a8(3, 9, 6)
a9(4, 7, 8)
a2(4, 8, 8)
39
REPUTATION PROCESS
Basic components:
Communicated image
Communicated reputation
Social Information
40
REPUTATION PROCESS
Communicated image
Image: “an evaluative belief; it tells whether the
target is good or bad with respect to a given
behaviour”
Communicated image aggregates!
Important that the individuals who communicate
the image are representative of the whole
So small samples can be a problem!
41
REPUTATION PROCESS
Communicated Reputation
Aggregation of reputation information of others
Image: evaluative belief, Reputation: what others say about
behavior
42
REPUTATION PROCESS
Communicated Reputation
Aggregation of reputation information of others
Inherited reputation
Reputation which is directly taken from third-
party agents with who the agent has a social
relation
Reputation doesn’t depend on behaviour of
subject!
Reputation associated with the role the subject
has in society
43
(DE)-CENTRALIZED REPUTATION?
Centralized reputation:
A central service collects the raw information and aggregates into a reputation
All individuals in the community contribute
Less impact of a few outliers
Public reputation, so newcomers benefit
But:
Trust in the central service is crucial
Personal biases and preferences are ignored
Vulnerability, as the central service is a bottleneck
Examples?
eBay reputation
Internet review sites
44
(DE)-CENTRALIZED REPUTATION?
De-centralized reputation:
Relies on individual information that each agent can obtain
No external entity to trust
Scalable
Different agents can follow different reputation functions
But:
Takes time to collect enough information
Requires more complex agents, as the calculation process is more complex
Examples:
Most multi-agent reputation models
45
USING REPUTATION
Reputation as source of trust
Useful if direct information is lacking
46
USING REPUTATION - PITFALLS
Unfair ratings
Deliberate sending wrong information
Ballot stuffing
Send more information than is warranted (e.g. 4 messages about 1 interaction)
Dynamic personality
Agents taking advantage of high reputations
Whitewashing
Change your name/id to escape a bad reputation
Collusion
Unfair cooperation (for instance as a group give inflated scores)
Sybil attacks
Create ‘fake’ identities (for instance to all give you high ratings)
Reputation lag
Take advantage of delay between bad action and bad reputation
47
TRUST IN NEGOTIATION
Negotiate about trust
When a group of agents together need to make
trust decisions
What characteristics are important for us to trust
another agent/agency?
Trust in negotiation
Evaluate your negotiating partner
Do you trust them to keep their word?
48
SOCIAL NORMS
Norm: social rule for behaviour
You do not leave a party without saying goodbye to the host
You bring a present to a birthday
You do not walk away during a conversation
49
QUESTIONS
50
RESEARCH PROJECTS Q4
User-aware eXplainable AI for improving human-AI teamwork
Agent tailoring
explanations
• How can an agent model and use human trust in the agent to tailor explanations? to human
• How can an agent model and use human workload to tailor explanations?
• How can an agent model and use (epistemic) relevance to tailor explanations?
• How can an agent model and use human understanding of the agent to tailor explanations?
• How can an agent model and use human situation awareness to tailor explanations?
• How can an agent model and use human reliance on the agent to tailor explanations? Agent actions
• How can an agent model and use human performance to tailor explanations? affecting human
trustworthiness
51
Computational Social Choice
CSE3210 Collaborative Artificial Intelligence
l.cavalcantesiebert@tudelft.nl
1
You have learned about
• Centralized x collaborative AI
• (Automated) Negotiation
• Multiagent teamwork: Coalition formation
• Human-agent teamwork: Coactive design
• Trust
Next
• Computational social choice
2
Outline
• What is (computational) social choice theory?
• Preference aggregation
• Social Welfare Functions
• Axioms
• Arrow’s theorem
3
Study material
• Slides
• Additional material:
– Handbook of Computational Social Choice
http://www.cse.unsw.edu.au/~haziz/comsoc.pdf
– Articles mentioned in the slides
4
Computational Social Choice
Part 1: Introduction and Preference Aggregation
l.cavalcantesiebert@tudelft.nl
5
Introductory example Group A
4 people
Group B
3 people
Group C
2 people
1st Gin tonic Beer Wine
• House party: 2nd Wine Wine Beer
– You are planning a party for 9 3rd Beer Gin tonic Gin tonic
people
– You only want one type of
drink in the party
– You asked the guests to
provide their preferences
beforehand
– Which drink should you serve?
?
?
6
Introductory example Group A
4 people
Group B
3 people
Group C
2 people
1st Gin tonic Beer Wine
• House party: 2nd Wine Wine Beer
Plurality rule: Only considers how often each alternative is ranked in the first place Gin tonic
• However, the majority (groups B and C combined) will be dissatisfied with this choice
Condorcet method: Determine which alternative defeats every other alternative in pairwise
majority comparisons:
Wine
- Gin preferred to wine? 4 people - Wine preferred to gin? 5 people
- Beer preferred to wine? 3 people - Wine preferred to beer? 6 people
7
Group A Group B Group C
Introductory example 1st
4 people
Gin tonic
3 people
Beer
2 people
Wine
• Another method:
Single transferable vote (STV): eliminate the options that are ranked first by the lowest
number of agents
- Remove wine from the choice set (first option of only 2 people)
Beer
A (4) B (3) C (2)
1st Gin tonic Beer Beer
2nd Beer Gin Gin
8
Introductory example
• This simple example shows that collective choice is not a trivial matter
What method
should I use?
How to deal
with strategic
manipulation?
How to deal
with ties?
9
Social choice theory
• Social choice theory: methods for aggregating the preferences of
multiple agents
– How to fairly allocate resources to the members of a society?
– How to elect a president given people’s preferences?
– How to aggregate the views of different judges in a court case?
• Origins:
Political
Economics Mathematics
science
10
Social choice theory
• Early days:
– Jean-Charles de Borda (1733-1799)
– Marquis de Condorcet (1743-1797)
– Charles L. Dodgson a.k.a. Lewis Carroll (1832-1898)
11
Computational social choice
• Computational social choice adds an algorithmic perspective to the
formal approach of social choice theory
12
Computational social choice
• Computational social choice:
– Bartholdi, Orlin, Tovey, and Trick (around 1990):
Complexity theory as a barrier for strategic manipulation
– U. Endriss and J. Lang: 1st International Workshop on
Computational Social Choice (COMSOC, 2006)
13
A
Computational social choice
• Multiagent system as a “society” of agents A2
A
– Recommender systems on the basis of choices
made by other users in the past B
– Metasearch engine that combines results of A3
several internet search engines
A1
B
A
A4
A4
Valcarce, D., Parapar, J., & Barreiro, Á. (2017). Combining top-n recommenders with metasearch algorithms. In ACM SIGIR
14
A
Computational social choice
• Multiagent system as a “society” of agents A2
A
– Collaborative large-scale peer grading (e.g.
MOOC) B
– Ethical decision making: Bottom-up approaches to A3
aggregate people’s (moral) preferences
A1
B
A
A4
A4
Caragiannis, I. (2017). Recent advances in large-scale peer grading. Trends in Computational Social Choice, 327.
Noothigattu, R., et al. (2018). A voting-based system for ethical decision making. In AAAI Conference on Artificial Intelligence. 15
Fundamentals
• How to compare?
– Cardinal utility*
• Is today warmer than or colder than yesterday?
• 𝑇𝑡𝑜𝑑𝑎𝑦 = 10°𝐶
• 𝑇𝑦𝑒𝑠𝑡𝑒𝑟𝑑𝑎𝑦 = 8°𝐶
• Today is warmer than yesterday, by 2°𝐶
– Ordinal utility (or preference)
• Do you prefer pizza or pasta? Do you prefer candidate A or B?
• In the absence of a common basis of comparison (e.g. money,
temperature), the meaning of individual (cardinal) utility values
is quite controversial
• Specially, how to compare between different people
(interpersonal comparisons)?
16
Fundamental
• In our lectures we will follow the research in computational social choice with
the classical perspective in which agents (or voters) are assumed to have
ordinal preferences
*For a review on Social Choice from a Utilitarian Perspective (i.e. using cardinal preferences), see section 10.8 of the
Handbook of Computational Social Choice http://www.cse.unsw.edu.au/~haziz/comsoc.pdf 17
Fundamentals
• Let’s consider a finite set 𝑁 = {1, … , 𝑛} of at least 2 agents (a.k.a
individuals, voters),
• and a finite universe 𝑈 = {𝑎, … , 𝑧} of at least 2 alternatives (a.k.a
candidates)
• Each agent 𝑖 has preferences over the alternatives in 𝑈, which are
represented by a transitive and complete preference relation ≿𝑖
– Transitivity: If 𝑎 ≿𝑖 𝑏 and 𝑏 ≿𝑖 𝑐 implies that 𝑎 ≿𝑖 𝑐
– Completeness: requires any pair of alternatives 𝑎, 𝑏 ∈ 𝑈 to be comparable
– Usually, we also assume antisymmetry: 𝑎 ≿𝑖 𝑏 and 𝑏 ≿𝑖 𝑎 implies that 𝑎 = 𝑏
• We denote 𝑎 ≻𝑖 𝑏 (strict preference), if 𝑎 ≿𝑖 𝑏 but not 𝑏 ≿𝑖 𝑎
• A preference profile is denoted as 𝑅 = ≿1 , … , ≿𝑛
18
Preference Aggregation
• Preference aggregation is the aggregation of several agents' preference
rankings of two or more alternatives into a single, collective preference ranking
(or choice) over these alternatives
19
Preference Aggregation
ℛ(𝑈)𝑛 ℛ(𝑈)
• Aggregates ≿1 𝑎 ≿1 𝑏 …
Social preferences of
Welfare individual agents ≿2 𝑎 ≿2 𝑏 ≿
Function into collective
…
(SWF) preferences
≿𝑛 𝑏 ≿𝑛 𝑎
• 𝑓: ℛ(𝑈)𝑛 → ℛ(𝑈)
20
Preference Aggregation
…
Function of U) to a set of preferred 𝑏
…
(SCF) alternatives, a.k.a. the choice set
• 𝑓: ℛ(𝑈)𝑛 × ℱ(𝑈) → ℱ(𝑈) ≿𝑛
…
21
Preference Aggregation
• A variety of interesting SWFs and SCFs have been proposed, often based on
mechanisms that seem to calculate the winner in an intuitively “fair” way
• After further analysis these mechanisms may show unintended consequences
that undercut their initial appeal
• Instead of based solely on intuitive appeal, social choice theorists argue they
should rely on axioms, i.e. precisely defined desirable properties
Axiom Property
22
Preference Aggregation
• Not every SWF or SCF will satisfy every axiom
• In this lecture we will consider axioms for SWFs, but very similar
axioms can be applied to SCFs
23
Some basic axioms
• Anonymity: symmetry w.r.t. agents
– If alternative 𝑎 wins with 𝑛1 and 𝑛2 for and 𝑛3 against, then 𝑎 should also
win with 𝑛2 , 𝑛3 for and 𝑛1 against
• Neutrality: symmetry w.r.t. alternatives
– If we swap alternatives 𝑎 and 𝑏 in a ballot, the outcome should be the
same
• Positive responsiveness:
– If the group decision is indifferent or favourable to 𝑎, and if individual
preferences remain the same except a single agent changes their ranking
in favour of 𝑎, then the group decision should favour 𝑎
24
May’s Theorem
• May’s Theorem: For two alternatives and an odd number of agents, majority
rule is the unique SWF that satisfies anonymity, neutrality, and positive
responsiveness
– Majority rule: Given two alternatives, 𝑎 should be preferred over 𝑏, iff (if and only if) there are
more agents who prefer 𝑎 to 𝑏, than 𝑏 to 𝑎
• 𝑈 = 𝑎, 𝑏
• 𝑁 = {𝑛1 , 𝑛2 , 𝑛3 }
𝑛1 𝑛2 𝑛3 Kenneth O. May
a a b 𝑎≻𝑏
b b a
25
May’s Theorem
• Anonymity
𝑛1 𝑛2 𝑛3 𝑛1 ′ 𝑛2 ′ 𝑛3 ′
a a b a b a 𝑎≻𝑏
b b a b a b
• Neutrality
– E.g. if votes are reversed, results follow
𝑛1 𝑛2 𝑛3
b a b 𝑏≻𝑎
a b a
26
May’s Theorem
• Positive responsiveness
𝑛1 𝑛2 𝑛3 𝑛1 𝑛2 𝑛3
a a b 𝑎≻𝑏 a a a 𝑎≻𝑏
b b a b b b
27
Preference Aggregation
• What happens if we add a third alternative?
– Majority rule can result in cycles when there are more than two alternatives
1 1 1 a
Majority (2 out of 3): 𝑎 ≻ 𝑏
a b c
Majority (2 out of 3): 𝑏 ≻ 𝑐
b c a
Majority (2 out of 3): 𝑐 ≻ 𝑎
c a b c b
– The pairwise majority relation in this example is cyclic, therefore not a well-formed
preference relation
– Hence, the majority rule does not satisfy anonymity, neutrality, and positive
responsiveness for 𝑈 ≥ 3
• Marquis de Condorcet was the one who first noted that social preference relations can
be problematic. We call these cyclic relations a Condorcet paradox
28
Axioms
Formalizing the Axioms:
• A SWF 𝐹 is anonymous if agents are treated symmetrically:
𝐹 𝑅1 , … , 𝑅𝑛 = 𝐹 𝑅𝜋(1) , … , 𝑅𝜋(𝑛)
for any profile 𝑅 and any permutation 𝜋: 𝑁 → 𝑁
𝐹 𝜋(𝑅) = 𝜋 𝐹(𝑅)
for any profile 𝑅 and any permutation 𝜋: 𝐴 → 𝐴
29
Axioms
• Positive responsiveness:
– For all 𝑎, 𝑏 ∈ 𝑈 and any two profiles 𝑅 and 𝑅′ ∈ ℛ(𝑈)𝑛 ,
– if 𝑎 ≿ 𝑏 for 𝑅 and
– for 𝑅′ a single agent changes their ranking in favour of 𝑎
– then 𝑎 ≻ 𝑏
30
Axioms
• An SWF satisfies the Pareto condition if, whenever all individuals rank 𝑎
above 𝑏, then so does society
𝑎 ≻𝑖 𝑏 for all 𝑖 𝜖 𝑁 implies that 𝑎 ≻ 𝑏
for all 𝑖 𝜖 𝑁
ℛ(𝑈)𝑛 ℛ(𝑈)
𝑎 ≻𝑖 𝑏
≿1 𝑎 ≻1 𝑏 ≻1 𝑐 ≻1 𝑑 𝑎 ≻𝑖 𝑐 𝑎≻𝑏≻𝑐≻𝑑 𝑏≻𝑎≻𝑐≻𝑑
≿2 𝑎 ≻𝑖 𝑑 𝑑≻𝑎≻𝑏≻𝑐 𝑑≻𝑏≻𝑐≻𝑎
𝑐 ≻2 𝑑 ≻ 2 𝑎 ≻2 𝑏
…
𝑐≻𝑎≻𝑑≻𝑏 𝑐≻𝑏≻𝑑≻𝑎
≿3 𝑑 ≻ 3 𝑎 ≻3 𝑐 ≻3 𝑏 𝑑 ≻𝑖 𝑐
…
31
Axioms
• An SWF satisfies independence of irrelevant alternative (IIA) if the social
preferences between any pair of alternatives only depends on the individual
preference profiles restricted to these two alternatives
– If 𝑎 is socially preferred to 𝑏, then this should not change when 𝑖 changes
its ranking of 𝑐
– Let 𝑅 and 𝑅’ be two preferences profiles and 𝑎 and 𝑏 be two alternatives
such that 𝑅|{𝑎,𝑏} = 𝑅′ |{𝑎,𝑏} , i.e. the pairwise comparisons between 𝑎 and 𝑏
are identical in both profiles
– Then, IIA requires that 𝑎 and 𝑏 are also ranked identically:
≿ | 𝑎, 𝑏 =≿ ′| 𝑎, 𝑏
32
• Independence of irrelevant alternative (IIA). Example:
ℛ(𝑈)𝑛 ℛ(𝑈)
…
…
≿3 𝑐 ≻3 𝑑 ≻ 3 𝑎 ≻3 𝑏
𝑎, 𝑐 Same
𝑎, 𝑑 Same
ℛ′(𝑈)𝑛 𝑏, 𝑐 Same ℛ′(𝑈)
𝑏, 𝑑 Same 𝑏≻𝑎≻𝑐≻𝑑
≿1 𝑏 ≻1 𝑎 ≻1 𝒅 ≻𝟏 𝒄
𝑐, 𝑑 Changed 𝑑≻𝑏≻𝑐≻𝑎
≿2 𝑎 ≻2 𝑏 ≻2 𝑐 ≻2 𝑑 𝑎≻𝑏≻𝒅≻𝒄
𝑐≻𝑏≻𝑑≻𝑎
…
≿3 𝒅 ≻𝟑 𝒄 ≻ 3 𝑎 ≻3 𝑏
…
33
Axioms
• An SWF is non-dictatorial if there is no agent who
can dictate a strict ranking no matter which
preferences the other agents have
– There is no agent 𝑖 such that for all preference
profiles 𝑅 and all alternatives 𝑎, 𝑏, 𝑎 ≻𝑖 𝑏 implies
that 𝑎 ≻ 𝑏
34
Preference Aggregation
• Why do we need axioms?
• With a formal definition of these desirable properties one can establish:
– Characterisation theorems: Show that a particular (class of)
mechanism(s) is the only one satisfying a given set of axioms
– Impossibility theorems: Show that there exists no aggregation
mechanism satisfying a given set of axioms
35
Arrow’s theorem
• Arrow’s Theorem: There exists no SWF that
simultaneously satisfies IIA, Pareto optimality,
and non-dictatorship whenever U ≥ 3
• IIA says that for 𝑎 ≻ 𝑏 to change we need to change the pairwise relationship
between 𝑎 and 𝑏; the same applies for 𝑏 ≻ 𝑐
• However, as 𝑏 has an extremal position for all agents (top or bottom), 𝑐 can be
moved above 𝑎 without changing either of these pairwise relations
38
Arrow’s theorem
• From transitivity we have that if 𝑎 ≻ 𝑏 and b ≻ 𝑐 , then 𝑎 ≻ 𝑐
• However, the Pareto condition requires that 𝑐 ≻ 𝑎, since every agent ranked
𝑐 above 𝑎.
1 1 1
𝑏 𝑏 𝑐
𝑐 𝑐 𝑎
𝑎 𝑎 𝑏
• We have a contradiction!
• Hence, to respect both IIA and the Pareto condition 𝑏 must be either at the
bottom or at the top of ≿
39
Arrow’s theorem
• Step 2: There is an agent 𝑛∗ who is extremely pivotal in the sense that by
changing its preferences, it can move a given outcome 𝑏 from the bottom to
the top of ≿
• Let us assume a preference profile in which every voter ranks 𝑏 last
1 1 1
𝑎 𝑎 𝑐
? ≻? ≻ 𝑏
𝑐 𝑐 𝑎
𝑏 𝑏 𝑏
40
Arrow’s theorem
• Now, let agents successively modify their preferences by moving 𝑏 from the
bottom to the top, preserving all other relative rankings
• Let us denote 𝑛∗ as the first agent who change causes the social ranking of 𝑏
to change 1 1 1
𝑏 𝑎 𝑐 ? ≻? ≻ 𝑏
≿1
𝑎 𝑐 𝑎
𝑐 𝑏 𝑏
𝑏 must be at
𝑛∗
1 1 1 the top or
bottom
𝑏 𝑏 𝑐
≿2 𝑏 ≻? ≻?
𝑎 𝑎 𝑎
𝑐 𝑐 𝑏 41
Arrow’s theorem
• There must always be some such agent as 𝑛∗ , because when all agents move
𝑏 to the top, the Pareto condition will require that 𝑏 is ranked at the top
• The only difference between ≿1 and ≿2 , are the preferences of 𝑛∗
• Step 3: 𝑛∗ is a dictator over any pair not involving 𝑏 (in our case 𝑎, 𝑐)
• Let us consider first 𝑎, and construct a new profile ≿3 from ≿2 by: 1) moving 𝑎
to the top of 𝑛∗ ’s preference ordering, and 2) rearranging the relative ranking of
𝑎 and 𝑐 for all agents other than 𝑛∗
1 1(𝑛∗ ) 1 1 1(𝑛∗ ) 1
𝑏 𝑏 𝑐 𝑏 𝑎 𝑎
≿2 ≿3
𝑎 𝑎 𝑎 𝑐 𝑏 𝑐
𝑐 𝑐 𝑏 𝑎 𝑐 𝑏
42
Arrow’s theorem
• In ≿1 , we had that 𝑎 ≻ 𝑏, as 𝑏 was at the bottom of ≿
• When we compare ≿1 to ≿3 , relative rankings between 𝑎 and 𝑏 are the same
for all agents
1 1 1 1 1(𝑛∗ ) 1
𝑏 𝑎 𝑐 𝑏 𝑎 𝑎
≿1 ≿3
𝑎 𝑐 𝑎 𝑐 𝑏 𝑐
𝑐 𝑏 𝑏 𝑎 𝑐 𝑏
43
Arrow’s theorem
• In ≿2 , we had that 𝑏 ≻ 𝑐, as 𝑏 was at the top of ≿ (being 𝑛∗ pivotal)
• Comparing ≿2 to ≿3 , relative rankings between 𝑏 and 𝑐 are the same
1 1(𝑛∗ ) 1 1 1(𝑛∗ ) 1
𝑏 𝑏 𝑐 𝑏 𝑎 𝑎
≿2 ≿3
𝑎 𝑎 𝑎 𝑐 𝑏 𝑐
𝑐 𝑐 𝑏 𝑎 𝑐 𝑏
𝑎≻𝑏≻𝑐
• Thus, in ≿3 , we have 𝑏 ≻ 𝑐
• As we have that 𝑎 ≻ 𝑏 and 𝑏 ≻ 𝑐, from transitivity we have that 𝑎 ≻ 𝑐 in ≿3
44
Arrow’s theorem
• Let us construct one more (last one! ) preference profile by changing ≿3 in
two ways: 1) Arbitrarily change the position of 𝑏 for each agent while keeping
all other relative preferences the same; 2) Move 𝑎 to an arbitrary position in
𝑛∗ ’s preference ordering, with the constraint that 𝑎 ≿𝑛∗ 𝑐
1 1(𝑛∗ ) 1 1 1(𝑛∗ ) 1
𝑏 𝑎 𝑎 𝑐 𝑏 𝑎
≿3 ≿4
𝑐 𝑏 𝑐 𝑏 𝑎 𝑏
𝑎 𝑐 𝑏 𝑎 𝑐 𝑐
45
Arrow’s theorem
• Note that the only assumption we have now for ≿4 is that 𝑎 ≿𝑛∗ 𝑐
• For ≿3 and ≿4 all agents have the same relative preference between 𝑎 and 𝑐
1 1(𝑛∗ ) 1 1 1(𝑛∗ ) 1
𝑏 𝑎 𝑎 𝑐 𝑏 𝑎
≿3 ≿4
𝑐 𝑏 𝑐 𝑏 𝑎 𝑏
𝑎 𝑐 𝑏 𝑎 𝑐 𝑐
• From IIA, we have that a ≻ 𝑐 only with the assumption that 𝑎 ≿𝑛∗ 𝑐
• If 𝑛∗ would change its preference to 𝑐 ≿𝑛∗ 𝑎, according to Pareto optimality
would follow that c ≻ 𝑎
• Hence, we cay that 𝑛∗ is a dictator, at least over (𝑎, 𝑐)
46
Arrow’s theorem
• Step 4: 𝑛∗ is a dictator over any pair of alternatives (including
pairs that include 𝑏).
• By the argument of Step 2, we can say that there is an agent
𝑛∗∗ who is extremely pivotal for 𝑐
• By the argument in Step 3, 𝑛∗∗ is a dictator over any pairs not
involving 𝑐
• We have previously observed that 𝑛∗ can affect the relative
preferences between 𝑎 and 𝑏 (remember ≿1 and ≿2 , when
we identified 𝑛∗ )
• Hence, as 𝑛∗∗ is a dictator over any pairs not involving 𝑐, and
𝑛∗ can influence 𝑏, we can conclude that 𝒏∗∗ and 𝒏∗ must be
the same agent!
47
Arrow’s theorem
• Take into consideration that:
– “Most systems are not going to work badly all of the time. All I proved is that all can
work badly at times”, K. Arrow
– The conditions of Arrow’s theorem can be relaxed to find more realistic solutions
– We can call the assumptions into questions e.g. in many applications, a full social
preference relation or IIA might not be needed
48
Conclusion
• Today we learned the fundamentals of social choice, and its most
discussed axioms
• The purpose of presenting the limitations is not convince you that
social choice is hopeless!
• Rather, it is intended to make one think about social choice in a
precise manner and to have more realistic expectations
• In the next lecture, we will move to more concrete procedures for
making decisions based on the preference of multiple agents
– Voting rules, scoring rules, Condorcet extensions
– Strategic manipulation
49
Computational Social Choice
CSE3210 Collaborative Artificial Intelligence
l.cavalcantesiebert@tudelft.nl
1
Recap
• Preference aggregation
–SWF: 𝑓: ℛ(𝑈)𝑛 → ℛ(𝑈)
–SCF: 𝑓: ℛ(𝑈)𝑛 × ℱ(𝑈) → ℱ(𝑈)
• Axioms:
–Anonymity, Neutrality, Positive responsiveness
–Pareto condition, Independence of Irrelevant Alternatives (IIA), Non-dictatorial
•Arrow’s theorem:
–There exists no SWF that simultaneously satisfies IIA, Pareto optimality, and non-
dictatorship whenever U ≥ 3
2
Computational Social Choice
Part 2: Voting methods
l.cavalcantesiebert@tudelft.nl
3
Outline
PART II: Voting methods
• Voting rules
• Scoring rules
• Condorcet extensions
• Strategic manipulation
4
Study material
• Slides
• Additional material:
– Handbook of Computational Social Choice
http://www.cse.unsw.edu.au/~haziz/comsoc.pdf
– Articles mentioned in the slides
5
Voting rules
• A voting rule is a function 𝑓: ℛ(𝑈)𝑛 → ℱ(𝑈)
6
Voting rules
Properties often required of voting rules:
• Resoluteness: they should always yield a unique winner
– 𝑓 is resolute if 𝑓 𝑹 = 1 for all preferences profiles R
• Anonymity: symmetry w.r.t. agents
• Neutrality: symmetry w.r.t. alternatives
3 2
𝑎 𝑏
𝑏 𝑐
𝑐 𝑎
– Winner: 𝒂
8
Voting rules
• A common objection to rules such as the plurality rule is that an alternative
ought to get some credit for being ranked, say, in second place by a voter
9
Scoring rules
• Under a scoring rule, each time an alternative is ranked 𝑖 − 𝑡ℎ by some voter,
it gets a particular score 𝑠𝑖
• For a fixed number of alternatives 𝑚, we define a score vector
𝑠 = {𝑠1 , … . , 𝑠𝑚 } such that 𝑠1 ≥ ⋯ ≥ 𝑠𝑚 and 𝑠1 > 𝑠𝑚
10
Plurality and Anti-plurality rules
• Plurality rule: We can also define the plurality rule as a scoring rule. The
score vector for the plurality rule is (1, 0, … , 0).
– The cumulative score of an alternative equals the number of voters by which it is
ranked first
• Anti-plurality rule: The score vector for the anti-plurality rule (a.k.a veto rule)
is (1, … , 1, 0).
– It choose those alternative that are least-preferred by the lowest number of votes.
– Sometime the anti-plurality rule is also defined as (0, … , 0 , -1)
11
Plurality and Anti-plurality rules
3 2
• Plurality rule: (1, 0, … , 0)
𝒂: 𝟑 𝑎 𝑏
𝑏: 2 𝑏 𝑐
𝑐: 0
𝑐 𝑎
• Anti-plurality rule: (1, … , 1, 0)
𝑎: 3
𝒃: 𝟑 + 𝟐 = 𝟓
𝑐: 2
12
Borda’s rule
• Borda’s rule: Alternative 𝑎 gets 𝑘 points from voter 𝑖, if 𝑖 prefers 𝑎 to 𝑘 other
alternatives, i.e. the score vectors is ( 𝑈 − 1, 𝑈 − 2, … , 0)
– Borda rule chooses those alternatives with the highest average rank in individual rankings
3 2
𝑎 𝑏 2 points 𝑎: 3 × 2 = 6
𝑏 𝑐 1 point 𝒃: 𝟐 × 𝟐 + 𝟑 × 𝟏 = 𝟕
𝑐: 2 × 1 = 2
𝑐 𝑎 0 point
13
Research example
• Ethical decision making for autonomous systems
– We want autonomous systems to respect given ethical principles… but which principles should
they favor in case of dilemmas?
– We modeled 6 ethical principles, and analyzed people’s preference w.r.t. these principles using
data from the Moral Machine experiment
Principles
2.92 1.05 0.77
violated (/6)
Principles
1.84 2.03 2.31
respected (/6)
Siebert, L.C., Andriamahery, K., Jonker, C., van den Hoven, J. Ethical decision-making for autonomous systems under 14
moral uncertainty. Work in progress.
Borda’s rulExe
• Exercise: Apply the Borda rule
1 3 2
𝑎 𝑐 𝑏 𝒂: 𝟏 × 𝟐 + 𝟑 × 𝟏 + 𝟐 × 𝟏 = 𝟕
𝑏 𝑎 𝑎 𝑏: 2 × 2 + 1 × 1 = 5
𝑐: 3 × 2 = 6 Winner: a
𝑐 𝑏 𝑐
15
Condorcet extensions
• Scoring rules have been criticized for failing to select the Condorcet
winner for some preference profiles
– Condorcet winner is an alternative that beats every other alternatives in pairwise majority
comparisons
– In the previous lecture we saw that there are some preference profiles that do not admit a
Condorcert winner
18
Condorcet extensions
• Copeland’s rule:
– An alternative gets a point for every pairwise majority win, and some fixed number of points
between 0 and 1 (say, ½) for every pairwise tie
– The winners are alternatives with the greatest number of points
19
Condorcet extensions
• Exercise: Apply Copeland’s rule (consider ½ point for ties)
𝑏 𝑎 𝑎 𝑏, 𝑎 (𝑤𝑜𝑛−𝑡𝑖𝑒𝑑−𝑙𝑜𝑠𝑡) = 2 − 0 − 4
𝑐, 𝑎 (𝑤𝑜𝑛−𝑡𝑖𝑒𝑑−𝑙𝑜𝑠𝑡) = 3 − 0 − 3
𝑐 𝑏 𝑐 𝑐, 𝑏 (𝑤𝑜𝑛−𝑡𝑖𝑒𝑑−𝑙𝑜𝑠𝑡) = 3 − 0 − 3
𝒂: 𝟏. 𝟓
𝑏: 0.5
𝑐: 1
Winner: a
20
Condorcet extensions
• Maximin (a.k.a. Simpson–Kramer method):
– Under this rule, we consider the magnitude of pairwise election results (by how many voters one
alternative was preferred to the other)
– We evaluate every alternative by its worst pairwise defat by another alternative
– The winners are those who lose by the lowest margin in their worst pairwise defeats (if there are
alternatives with no pairwise defeats, they win!)
– Let 𝑋, 𝑌 denote the pairwise score for 𝑋 against 𝑌. The candidate selected by maximin is given by:
𝑊 = arg min(max( 𝑑 𝑋, 𝑌 − 𝑑 𝑌, 𝑋 ))
𝑥 𝑦
6 3 4 4
𝑎, 𝑏 =9−0−8 • Pairwise defeats:
(𝑤𝑜𝑛−𝑡𝑖𝑒𝑑−𝑙𝑜𝑠𝑡)
𝑎 𝑐 𝑏 𝑏 𝒂: −(𝟎)
𝑎, 𝑐 (𝑤𝑜𝑛−𝑡𝑖𝑒𝑑−𝑙𝑜𝑠𝑡) = 10 − 0 − 7 𝑏: 1
𝑏 𝑎 𝑎 𝑐 𝑏, 𝑐 (𝑤𝑜𝑛−𝑡𝑖𝑒𝑑−𝑙𝑜𝑠𝑡) = 14 − 0 − 3 𝑐: {3,11}
𝑐 𝑏 𝑐 𝑎 𝑏, 𝑎 (𝑤𝑜𝑛−𝑡𝑖𝑒𝑑−𝑙𝑜𝑠𝑡) = 8 − 0 − 9 b: margin of 1
𝑐, 𝑎 (𝑤𝑜𝑛−𝑡𝑖𝑒𝑑−𝑙𝑜𝑠𝑡) = 7 − 0 − 10 c: margin of 3 arg min(max{𝟎, 1,11})) : 𝑎
𝑥 𝑦
c: margin of 11
𝑐, 𝑏 (𝑤𝑜𝑛−𝑡𝑖𝑒𝑑−𝑙𝑜𝑠𝑡) = 3 − 0 − 14 Winner: 𝒂 (Condorcet winner)
21
Condorcet extensions
• Exercise: Apply the Maximin rule
𝑎, 𝑏 (𝑤𝑜𝑛−𝑡𝑖𝑒𝑑−𝑙𝑜𝑠𝑡) = 4−0−2
1 3 2
𝑎, 𝑐 (𝑤𝑜𝑛−𝑡𝑖𝑒𝑑−𝑙𝑜𝑠𝑡) = 3−0−3
𝑎 𝑐 𝑏 𝑏, 𝑐 (𝑤𝑜𝑛−𝑡𝑖𝑒𝑑−𝑙𝑜𝑠𝑡) = 3 − 0 − 3
𝑏 𝑎 𝑎 𝑏, 𝑎 (𝑤𝑜𝑛−𝑡𝑖𝑒𝑑−𝑙𝑜𝑠𝑡) = 2 − 0 − 4 b: margin of 2
𝑐, 𝑎 (𝑤𝑜𝑛−𝑡𝑖𝑒𝑑−𝑙𝑜𝑠𝑡) = 3 − 0 − 3
𝑐 𝑏 𝑐 𝑐, 𝑏 (𝑤𝑜𝑛−𝑡𝑖𝑒𝑑−𝑙𝑜𝑠𝑡) = 3 − 0 − 3
22
Condorcet extensions
• Nanson’s rule:
– Runoff method (i.e. multiple rounds)
– Series of Borda elections. Two variants:
• Fishburn variant (a.k.a Baldwin’s rule): For each Borda election, exclude the alternative with the lower
Board score unless all alternatives have identical Borda score, in which case these candidates are
declared the winner(s).
• Schwartz variant (the one we will use!): For each Borda election, exclude the alternative(s) which have
less than the average Borda score unless all alternatives have identical Borda score, in which case these
candidates are declared the winners
6 3 4 4 𝑎: 6 × 2 + 7 × 1 = 19 6 3 4 4
𝑎 𝑐 𝑏 𝑏 2 points 𝒃: 𝟖 × 𝟐 + 𝟔 × 𝟏 = 𝟐𝟐 𝑎 𝑎 𝑏 𝑏 1 point
𝑐: 3 × 2 + 4 × 1 = 10
𝑏 𝑎 𝑎 𝑐 1 point 𝑏 𝑏 𝑎 𝑎 0 point
𝑐 𝑏 𝑐 𝑎 0 point Average Borda score = 17
1 3 2 𝑎: 1 × 2 + 5 × 1 = 7 1 3 2 𝑎: 3 × 2 + 3 × 1 = 9
𝑏: 2 × 2 + 1 × 1 = 5 𝑐: 3 × 2 + 3 × 1 = 9
𝑎 𝑐 𝑏 𝑐: 3 × 2 = 6 𝑎 𝑐 𝑎
𝑏 𝑎 𝑎 𝑐 𝑎 𝑐 Same Borda score
Average Borda score = 6
𝑐 𝑏 𝑐
Remove alternative b
24
Other rules
Finally, there are many others rules that are neither scoring rules, nor Condorcet
extensions, for example:
• Single Transferable Vote (STV):
– Looks for alternatives that are ranked in first place the least often, removes them from all voter’s
ballot, and repeats
– The alternatives removed in the last round win
6 3 4 4
6 3 4 4
𝑎 𝒄 𝑏 𝑏
𝑎 𝑎 𝒃 𝒃 Winner: 𝒂
𝑏 𝑎 𝑎 𝑐
𝑏 𝑏 𝑎 𝑎
𝑐 𝑏 𝑐 𝑎
25
Other rules
• Exercise: Apply Single Transferable Vote (STV)
1 3 2
1 3 2
𝑎 𝑐 𝑏
𝑏 𝑐 𝑏 Winner: b and c (tie)
𝑏 𝑎 𝑎
𝑐 𝑏 𝑐
𝑐 𝑏 𝑐
26
Other rules
• Bucklin’s rule:
– Check whether there is any alternative that is ranked first by more than half the voters; if so, this
alternative wins
– If not, check whether there are any alternatives that are ranked in either first or second place by
more than half the voters; if so, they win
– If not, consider the first three positions, etc.
– When multiple alternatives cross the 𝑛/2 threshold simultaneously, it is common to break ties by
the margin by which they crossed the threshold
1 3 2
Ranked first place: Ranked first or second place:
𝒂 𝑐 𝑏 𝑎: (1/6) < 50% 𝑎: (6/12) = 50%
𝑏 𝑎 𝑎 𝑏: (2/6) < 50% 𝑏: (3/12) < 50%
𝑐: (3/6) = 50% 𝑐: (3/12) < 50%
𝑐 𝑏 𝑐
No winner
28
Interactive activity
• Public consultation on policies to
combat climate change. Consider
the following options:
– Introduction of meat tax
– More windmills and solar parks
– Improvement of household’s insulation
– Increase tax on natural gas
– Extension of subsidy for electric
vehicles
– Subsidy to reduce CO2 emissions from
industry
• What options do you prefer? (~3-5
minutes)
– Go to https://bit.ly/3pPtXp6 and state
your preferences
– The order of the options are
randomized
– Type some random characters as your
“Nickname” (let’s keep it anonymous)
29
Interactive activity
• Which method do you find more suitable for this participatory process? Why?
• Interested in the topic? This was a study in which 10.000 participated. See more info and
the insights here (Dutch-language only)
– https://www.tudelft.nl/tbm/pwe/case-studies/klimaatraadpleging
30
Strategic Manipulation
• So far, we assumed that the true preferences of all agents are known
• However, we only have access to their reported preferences
• This is an unrealistic assumption because agents might misrepresent their preferences
to exploit the voting rules
• Plurality rule:
1 2 2 2 1 2 2 2
𝑎 𝑎 𝑏 𝑐 𝑎 𝑎 𝑏 𝑏
𝑏 𝑐 𝑑 𝑏 Winner: 𝑎 𝑏 𝑐 𝑑 𝑐 Winner: 𝑏
𝑐 𝑏 𝑐 𝑑 𝑐 𝑏 𝑐 𝑑
𝑑 𝑑 𝑎 𝑎 𝑑 𝑑 𝑎 𝑎
They
prefer 𝑏
to 𝑎
31
Strategic Manipulation
• Borda rule:
1 2 2 2
𝑎 𝑎 𝑏 𝑐 𝑎: 3 × 3 = 9
𝑏: 𝟐 × 𝟑 + 𝟑 × 𝟐 + 𝟐 = 𝟏𝟒
𝑏 𝑐 𝑑 𝑏 𝑐: 2 × 3 + 2 × 2 + 3 × 1 = 13
𝑐 𝑏 𝑐 𝑑 𝑑: 2 × 2 + 2 × 1 = 6
𝑑 𝑑 𝑎 𝑎
1 2 2 2
They
𝑎 𝑐 𝑏 𝑐 𝑎: 1 × 3 + 2 × 2 = 7
prefer 𝑐 𝑏: 2 × 3 + 3 × 2 + 2 = 14
𝑏 𝑎 𝑑 𝑏 𝒄: 𝟒 × 𝟑 + 𝟑 × 𝟏 = 𝟏𝟓
to 𝑏
𝑐 𝑏 𝑐 𝑑 𝑑: 2 × 2 + 2 × 1 = 6
𝑑 𝑑 𝑎 𝑎
32
Strategic Manipulation
• Why avoid manipulation?
– It can decrease fairness since manipulative skills are not spread evenly
– Hard to evaluate whether the outcome respects agent’s “true preferences” or not
33
Manipulation
• Research in computational social choice has investigated the question of whether
manipulation can be made computationally difficult
• The Gibbard-Satterthwaite theorem requires that the voting rule is defined for all
possible preference profiles
• However, if we restrict the domain of possible preference profiles, we might create
strategyproof solutions:
– One well-know approach is the domain of single-peaked preferences e.g. when it is possible to
establish a linear ordering of the alternatives
• Example: tax rate {20%, 25%, 30%}
• Preferences are single-peaked if for every voter, as we move away from the voter’s most
preferred alternative, the alternative will become less preferred for that voter
• if 𝑥 < 𝑦 < 𝑧 or 𝑧 < 𝑦 < 𝑥 , then 𝑥 ≻𝑖 𝑦 implies 𝑦 ≻𝑖 𝑧 for every 𝑖 ∈ 𝑁
34
Manipulation
• In many real-life setting we can not expect restriction of preferences to hold
– There might be other reasons for voters to rank a given alternative higher or lower
35
Conclusion
• Computational social choice can foster collaboration among agents
• Even though many collective decision problems can be modelled as problems
of social choice, it is also true that many practical problems will not necessarily
fit the template provided by classical frameworks
• There is much more to computational social choice: Tournaments,
combinatorial domains, fair allocation…
“No one pretends that democracy is perfect or all-wise. Indeed it has been said
that democracy is the worst form of Government except for all those other forms
that have been tried from time to time.…”
Winston S Churchill
36
Ethics of Collaborative AI: Values
Enrico Liscio and Pradeep K.
Murukannaiah
What is Ethics?
The field of ethics involves systematizing, defending, and
recommending concepts of right and wrong behavior
[Fieser, The Internet Encyclopedia of Philosophy]
▪ … as an AI engineer?
▪ Because AI applications take increasingly sophisticated actions that affect humans
▪ … as a Collaborative AI engineer?
▪ Ethics is inherently a multi-agent concern
Values: A Building Block of Ethical AI
▪ What are values?
▪ Health support
▪ Autonomous driving
15-03-2022
Practical Example
Health support agent for grandpa in different
contexts:
▪ COVID-19 regulations
▪ Driving
15-03-2022
General Values
What values are relevant in the
following situations?
▪ Commuting
15-03-2022
Value Identification
We need to identify and define values in a context: context-specific values.
woman
king
queen
“Everybody should
“The economy must restart”
wear mouth
masks”
“Loneliness kills”
S. Basu, A. Banerjee, and R. J. Mooney (2004). “Active semi-supervision
15-03-2022
for pairwise
constrained clustering”. In: Proc. SDM 2004, pp. 333–344.
Axies Methodology - Exploration
Given as the set of visited points and the set of novel points, FFT
selects the next point as follows:
5 11
5
4
12
Annotator 1
Freedom
Safety
Group
Fairness
Freedom
Mental Health
Safety
Fairness
Annotator 2 Mental Health
Freedom Prosperity
Safety Pleasure
Prosperity
Pleasure
15-03-2022
Value Elicitation
The result of Axies is a list of values relevant to
a context.
A concrete application is the elicitation of value
preferences through natural language.
Wrong!
15-03-2022
Label Subjectivity
The subjectivity of the interpretation of values must be considered,
especially in collaborative settings.
Textual inputs must be annotated by multiple annotators. Then, we
select the majority annotation to perform supervised classification.
Wrong!
15-03-2022
Evaluation of Supervised Classifiers
Supervised classifiers are typically evaluated with the
score, assuming that a prediction is either correct or wrong.
Real label
Positive Negative
True Positive False
Predicted Positive (TP) Positive (FP)
label False True
Negative
Negative Negative
(FN) (TN)
15-03-2022
Data Representation
A1: [0, 1, 0,
Example: three annotators have annotated one datapoint 0]
A2: [0, 1, 1,
Examples of data representation are: 0]
A3: [0, 0, 0,
▪ Majority agreement aggregation: [0, 1, 0, 0]
1]
▪ Repeated labels: [0, 1, 0, 0] [0, 1, 1, 0] [0, 0, 0, 1]
15-03-2022
Takeaways
15-03-2022
Readings
▪ Required: Lecture slides
▪ Optional
CREDITS: This presentation includes icons by Flaticon and infographics & images by Freepik
Ethics of Collaborative AI: Social Norms
Pradeep K. Murukannaiah
The Essence of Collaborative AI
Formal representation:
▪ Flexibility
▪ A principal can violate norms (but bears the consequences)
▪ Explainability
▪ Provides an opportunity for principals to explain their actions
Types of Social Norms
▪ There are a few norm types, which can many instances in an STS
▪ The norm type defines the semantics of expectations underlying a norm
Types of Norms: Commitment
▪ Within an organizational context,
the subject (i.e., debtor) commits to the object (i.e., creditor) that if the
antecedent holds, the debtor will bring about the consequent
Norm:
Types of Norms: Prohibition
▪ Within an organizational context,
The object prohibits (i.e., forbids) the subject from bringing about the consequent
provided the antecedent holds
Patient’s personal health information (PHI) should not be published online under
any circumstances
Norm:
Types of Norms: Authorization
▪ Within an organizational context,
the object authorizes (i.e., permits) the subject to bring about the consequent
when the antecedent holds
Norm:
Types of Norms: Sanction
▪ Within an organizational context,
the object would sanction (i.e., reward or punish) the subject by bringing
about the consequent provided the antecedent holds
A buyer can sanction a seller by providing a poor rating if the product received
is not as advertised
Norm:
Types of Norms: Power
▪ Within an organizational context,
when the antecedent holds, the object empowers the subject to bring about
the consequent at will
Norm:
▪ An STS can have many inorm instances, but they belong to a few
norm types
15-03-2022
Readings
▪ Required: Lecture slides
▪ Optional
Pradeep K. Murukannaiah
The Centralized AI Paradigm
l.cavalcantesiebert@tudelft.nl
1
ROBO
Robot for Burning Offices
Example adapted from the edX TU Delft MOOC “Mind of the Universe - Robots in Society: Blessing or
Curse?”, as described by Rijk Mercuur. https://www.youtube.com/watch?v=9fb_MENUfS4 2
ROBO’s main goal is to
support fire emergency
evacuation, including:
• Lead people to the best exit
route
• Remove obstacles
• Extinguish fire
• …
Example adapted from the edX TU Delft MOOC “Mind of the Universe - Robots in Society: Blessing or
Curse?”, as described by Rijk Mercuur. https://www.youtube.com/watch?v=9fb_MENUfS4 3
What should ROBO do?
4
One possible way: Act according to what people want
Gabriel, I. (2020). Artificial intelligence, values, and alignment. Minds and machines, 30(3), 411-437.
Russell, S. (2019). Human compatible: Artificial intelligence and the problem of control. Penguin. 5
One possible way: Act according to what people want
Don’t worry about me. …
• What if there is no Go to the fifth floor,
agreement? there is someone there
that needs assistance.
• What if people are not
consistent?
• What if this behavior cannot
be modelled (or learned)?
…
• Goal
– Morality should be based on whether a action itself is right or
wrong (Kantianism)
• What if:
– ROBO saves someone in a wheelchair but dozens of people get
severely injured?
7
Artificial moral agents
• Any approach may have severe ethical implications
– People will disagree on what is right or wrong ?
– Unexpected and emergent situations might arise
– Specially for wicked problems: class of problems for which
science provides insufficient or inappropriate resolution
8
Hybrid (moral) Intelligence
• Hybrid intelligence (HI) as the
combination of human and machine
intelligence
– Augments human intellect and capabilities
instead of replacing them and achieving
goals
van der Waa, J., van Diggelen, J., Siebert, L. C., Neerincx, M., & Jonker, C. (2020, July). Allocation of Moral Decision-Making in Human-Agent
Teams: A Pattern Approach. In International Conference on Human-Computer Interaction (pp. 203-220). Springer, Cham. 10
TDP2: Supported moral decision-making
van der Waa, J., van Diggelen, J., Siebert, L. C., Neerincx, M., & Jonker, C. (2020, July). Allocation of Moral Decision-Making in Human-Agent
Teams: A Pattern Approach. In International Conference on Human-Computer Interaction (pp. 203-220). Springer, Cham. 11
TDP3: Coactive moral decision making
van der Waa, J., van Diggelen, J., Siebert, L. C., Neerincx, M., & Jonker, C. (2020, July). Allocation of Moral Decision-Making in Human-Agent
Teams: A Pattern Approach. In International Conference on Human-Computer Interaction (pp. 203-220). Springer, Cham. 12
TDP4: Autonomous moral decision making
van der Waa, J., van Diggelen, J., Siebert, L. C., Neerincx, M., & Jonker, C. (2020, July). Allocation of Moral Decision-Making in Human-Agent
Teams: A Pattern Approach. In International Conference on Human-Computer Interaction (pp. 203-220). Springer, Cham. 13
Collaborative AI = Ethical AI?
14
Keeping control
• Humans must be in a position to be capable of being in control
of the system
• Machines should be able to understand and be responsive to
our moral standards
Synergistic &
transparent
collaboration
15
Meaningful Human Control
Santoni de Sio, F., & Van den Hoven, J. (2018). Meaningful human control over 16
autonomous systems: a philosophical account. Frontiers in Robotics and AI, 5, 15.
Meaningful Human Control
• Some necessary conditions
– Define an moral operational
design domain (moral-ODD)
Siebert, L. C., Lupetti, M. L., Aizenberg, E., Beckers, N., Zgonnikov, A., Veluwenkamp,
H., ... & Lagendijk, R. L. (2021). Meaningful human control over AI systems: beyond
talking the talk. arXiv preprint arXiv:2112.01298. https://arxiv.org/pdf/2112.01298.pdf 17
Meaningful Human Control
• Some necessary conditions
– Make sure humans have ability
and authority to intervene and
steer the system
– Keep traceability
Siebert, L. C., Lupetti, M. L., Aizenberg, E., Beckers, N., Zgonnikov, A., Veluwenkamp,
H., ... & Lagendijk, R. L. (2021). Meaningful human control over AI systems: beyond
talking the talk. arXiv preprint arXiv:2112.01298. https://arxiv.org/pdf/2112.01298.pdf 18
Conclusion
• Only AI
– Unexpected situations
– Might misalign with our moral values (including emergent patterns)
– Who gets to decide which values it should follow?
• Only human
– Might be difficulty to cope with speed and data volume of AI
• Hybrid intelligence (Human + AI)
– Working together to avoid pitfalls
– AI under meaningful human control, aligned with our moral values
• To design and develop AI systems is not a value-neutral activity
• Multidisciplinary challenge!
19