Computational Models for

1
Computational Models
for Belief Revision, Group Decisions and Cultural
Shifts
Whitman Richards (PI), M.
I. T. Computer Science and Artificial
Intelligence Laboratory
Interdisciplinary 6 Universities 9 Disciplines
20 projects
MURI21
2
The Team
Experimental and Field Analyses
Scott Atran, CUNYCtr for Terrorism UnivMich
Anthropology, Policy Jenna Bednar, Univ. of
Michigan, Political Science Public
Policy Doug Medin, Northwestern University,
Psychology
Model Development
Ken Forbus, Northwestern Univ.,
Computer Science and Education John Mikhail,
Georgetown Univ., Ethics and Law
(anticipated 6/08) Scott Page, University of
Michigan, Economics Complex Systems Avi
Pfeffer, Harvard,
Computer Science Whitman Richards, MIT,
Cognition and Artificial Intelligence Joshua
Tenebaum, MIT, Computation and
Cognitive Science Patrick Winston, MIT,
Computer Science and Artificial Intelligence
Consultants Robert Axelrod, University of
Michigan, Political Science
Marc Sageman, Sageman Consultants, D.C.
Rajesh Kasturirangan, Natl Inst. Advanced
Studies, Bangalore
3
Computational Models for Belief Revision, Group
Decisions and Cultural Shifts
Primary Objectives
1. Experimental Network Analyses (data)
How do beliefs support and lead to certain
actions in one culture and not another?
2. Model Development (theory)
Develop computational models that further our
understanding of the relation between beliefs,
decisions and actions
Models should distinguish the different roles
played by sacred vs. instrumental or secular
values
Models should provide formal explanations for how
the beliefs of individuals affect group and
individual actions, and how groups evolve.
4
MURI Scope
20 Projects
Palestine (2), Wisconsin(Menominee, Amish),
Guatemala, India
Story Understanding (role models, actions,
analogies) (2)
Network Analyses (eg Madrid, Hamburg,Leeds. (5)
Group Evolution (2)
Multi-agent games, Belief revision (3), Learning,
Strategic Reasoning
Relational Models Framing Horizon effects
Theory of Mind
See http//groups.csail.mit.edu/belief-dynamics/
5
The Domain
See http//groups.csail.mit.edu/belief-dynamics/
23 Jan 08
6
See http//groups.csail.mit.edu/belief-dynamics/
23 Jan 08
7
Strategic Scenarios

• Scenarios may involve multiple agents, all of
  which make decisions, and receive rewards based
  on their decisions
• May be competitive, cooperative, or anything in
  between
• Payoffs may be uncertain
• May involve sequential negotiations
• Sacred Values can be important

8
Possible Approaches

• Equilibrium (e.g.Nash, WD)
• Opponent modeling (e.g. poker)
• Behavioral econ, (e.g. ultimatum)
• Psychological theories (eg.cog.diss.)

9
Well-Distinguished Strategies
Defn a strategy is well-distinguished if the
information available to the player makes a
difference.
1. If the information makes a difference, then
the player must reason about how information is
used and manipulated.
2. The elements of this reasoning process can be
represented as a graph describing a strategic
situation.
Theorem Four different reasoning patterns are
sufficient
10
Four Patterns of Strategic Reasoning
WD equilibrium as refinement of Nash
Key Point Assumptions differ from classical
Game-Theoretic rationality stresses use of
available information.
11
See http//groups.csail.mit.edu/belief-dynamics/
23 Jan 08
12
Two Cultures in Northeast Wisconsin
Medin et al.

• Participants
• Native Americans (Menominee reservation)
• Evangelical Christians (neighboring Shawano
  county)
• Cultural Conflict
• Avid hunters and fishers (in both locations)
• But.sport versus inherent part of social
  structure

Note Low on strategic complexity dimension
13
Two Cultures in Middle-East

• Palestinian refugees vs displaced Israeli
  Settlers
• Study of how Palestinians view secular vs
  sacred trades
• that underly potential peace agreements
• (a) with sacred values (e.g. right of
  return, recognition of Israel)
• (b) with secular values (the UN offers 10
  billion a year in aid)
• SacredSacred (taboo) versus Sacred
  (taboo)

14
Palestinian recognition of the sacred right of
Israel
Medin et al.
15
Terrorism Radicalization
What to do, what not to do
Scott Atran Presentations at U.S.
State Department House of Lords, UK Oct.
/ Nov. 2007
www.edge.org/3rd_culture/Atran07
Marc Sageman on counter-terrorism strategy NSC
(3), HSC British Home Office,
16
Beliefs, Actions and Culture
See http//groups.csail.mit.edu/belief-dynamics/
23 Jan 08
17
The Gateway Stories

• Stories capture beliefs, values, role models
• Folktales, Myths, Morality Tales, Religious
  Texts, Urban Legends

• Main Hypothesis

People reason by analogy with experience, real or
surrogate

• Helping Lemma

A good approximation to Culture is a set of
prototypical folktales or stories
18
Catalyze research by speeding encoding.Improve
results by decreasing tailorability Eventually,
practical modeling tools for analyst
decision-maker support
Produce models via analogicalgeneralization,
predictions via simulation
Story Workbench
(sematic content)
Interviews, surveys, cultural stories collected
Predicate calculusrepresentations ofstories,
explanations
Qualitative Concept Maps
(does x increase or decrease y?)
19
The Story Workbenchapplication for collecting
semantic annotations of text

• Story Workbench - In Beta Testing at Brain
  Cognitive Sciences Department, MIT (Gibsons
  natural language group.)
• Java Wordnet Interface (JWI)Java library for
  accessing Princetons Wordnet electronic
  dictionary
• Approximately 50 users worldwide
• Stanford, CMU, USC, NYU, Lockheed Martin, Idaho
  National Labs
• England, The Netherlands, Taiwan, China, Spain,
  Poland

20
Belief Revision
Scope recruitment, cultural shifts, negotiation,
framing..)
23 Jan 08
21
Belief Revision Projects
Framing of Issues (Atran Axelrod, 2008)
Evolutionary (replicator) Dynamics (Gal et al.)
Goal how to intervene in an evolving system to
encourage a desired outcome
22
Scenario
Two Mosques, one with moderate views, the other
with fundamentalist views Two Belief systems,
divided among the mosques one insurgency
supporters, the others are moderates
The 4 populations evolve over time (two 2x2
games with payoffs).
23
Potential complexities of evolution, different
initial conditions, payoffs fixed
24
Replicator Dynamics often favored for capturing
Cultural Learning (e.g. revision of beliefs of
populations)
Best Response usually considered as
self-interested learning (e.g. rock-paper-scissors
)
New Result Games can be constructed where
Replicator and Best response learning lead to
quite different equilibria.
Why Care ?
Societies differ in the degrees of individual
self-interest
Implication The character of a population can
affect the ability to solve collective action
problems
25
See http//groups.csail.mit.edu/belief-dynamics/
23 Jan 08
26
Learning different forms of network structure
Tenenbaum Kemp
Dominance hierarchy Tree
Cliques Ring
Primate troop Bush administration
Prison inmates New Guinea islands
beats told likes
trades with
27
(No Transcript)
28
Small Group Evolution Street Gangs
29
Number of Graphical Forms
Typical Group Representation
n6 110 n7 850 n10 10 million n12
150 billion
A Picture is NOT worth 1000 words !!
30
Leadership
L 1.0
Bonding
B 1.0
Diversity
D 0.92
31
(No Transcript)
32
(No Transcript)
33
(No Transcript)
34
Pattern 2 Manipulation
Brush Teeth
Offer to Read
Parent
Child
35
Small Group Evolution example
36
Madrid Group Evolution
37
Small Group Evolution
38
The Domain
See http//groups.csail.mit.edu/belief-dynamics/
23 Jan 08