A seminar given to the Judgement and Decision Making Research Group in the Department of Neuroscience, Psychology and Behaviour, University of Leicester kindly asked me to give a seminar on 25 January 2023 on "The Role of Agent-Based Modelling in Extending the Concept of Bounded Rationality". It discusses the challenges to different research methods of dealing with subjective accounts and models a situation where people can be rational but communicate and have incomplete information about both the number of choices and their payoff. The model is based on this paper: https://doi.org/10.1007/s11299-009-0060-7 One interesting result is that, without coercion or mass media, minority groups may be disadvantaged in their decision making by hegemonic discourse.
The Role of Agent-Based Modelling in Extending the Concept of Bounded Rationality: A Case Study of “Choice Transmission”
1. DEPARTMENT OF SOCIOLOGY
The Role of Agent-Based Modelling in
Extending the Concept of Bounded
Rationality: A Case Study of “Choice
Transmission”
Edmund Chattoe-Brown <ecb18@le.ac.uk>
2. 1. Plan
• The challenge of not confusing our theories
and methods with “underlying reality”.
• A model of rational decision with
communication.
• A topic for discussion: The boundaries
between experiments and subjective
accounts.
3. 2. A real problem
• “I call it the law of the instrument, and it may be
formulated as follows: Give a small boy a
hammer, and he will find that everything he
encounters needs pounding.” (Kaplan,
Abraham, 1964, The Conduct of Inquiry, San
Francisco, CA: Chandler, p. 28).
• Important: Not just our specific methods but
our terminology and how we “go about”
research (methodology, design, implicit
assumptions).
4. 3. Example
• Data Generation Process (DGP): The “fact of the matter”
about how John/Jane Doe decided to do a BA in Sociology at
UoL.
• What we measure affects what we know: A survey tells us
that BAME students tend to go to universities with more
BAME students (but not necessarily why). An experiment tells
us that cognitive biases apply to experimental university
choice decisions (but not necessarily to real ones).
• But a qualitative researcher would say “Just ask John/Jane
what they did”. All sorts of problems with that but are they
proportionately worse than with other methods? Do they
justify ignoring distinctive “unmediated” data?
5. 4. Possible problems
• What if real choice processes don’t have analytical
representations? (Bounded Rationality.)
• What if we don’t yet know enough to “control” effectively in
experiments? (Replication Crisis.)
• How do we aggregate the consequences of decisions in
the “real world” i. e. which university ends up with what
kind of students? (Micro-macro problem.)
• Generally, how do we make research empirically rigorous:
“My experimental data is not incompatible with the way I
have elected to represent a theory”. Hmmm.
6. 5. A possible answer: ABM
• A particular kind of computer simulation: Theory
that is neither mathematics nor “narrative”.
• Explicitly represents “agents” (in this case
decision makers) and their interactions with
each other (and perhaps with an independent
“environment”.)
• A distinctive “falsifying” methodology based on
specification, calibration and validation.
• Not a panacea.
7. 6. What if …
• … agents are still strictly rational but only
amongst the alternatives they know?
• Why assume common knowledge/perfect
information?
• Modelling the process by which choices become
“visible” to decision makers. (Don’t need to if
everyone already knows everything!)
• Need a way to represent theory that can “cope
with” this degree of heterogeneity.
8. 7. Ironically, you can’t ask what …
• ““But I had to stay with him,” answered the vampire. “As I’ve
told you, he had me at a great disadvantage. He hinted there
was much I didn’t know and must know and that he alone
could teach me. But in fact, the main part of what he did
teach me was practical and not so difficult to figure out for
oneself.”” (Rice, Anne, 1977, Interview with a Vampire: First
Volume of the Vampire Chronicles, London: Futura, p. 40).
• ““That is true, but rising before me he might have nailed my
coffin shut. Or set it afire. The principal thing was, I didn’t
know what he might do, what he might know that I still did
not know.”” (ibid., p. 85).
9. 8. Blanket apology
• This analysis refers to an old article. If I was
doing it now, I would do things differently but I
nonetheless want to discuss the example.
• Put another way, I hope the interest of what I
have to say will survive comments of the form “I
would have made a different assumption” or
“you should have done more simulation runs”.
(Probably, now, so would I!)
10. 9. The model 1
• Agents and “situations” in the environment. A situation
yields different payoffs to different actions.
• Actions are not such that one can reasonably “infer”
missing possibilities: Compare “hit it”, “boil it” and “stretch
it” with “offer £1”, “offer £2” and “offer £3”.
• Situations have “objective” payoff sets to actions: (-4 4 6
5 1). Agents have dynamic subjective ones which may
approximate these more or less accurately (3 4 5 nil nil).
• Agents also have confidence in actions: If I have tried
something my confidence is 1. (Static world.)
11. 10. The model 2
• Tradition: All agents start knowing only the same single action
payoff correctly and with complete confidence.
• Innovation: Conceiving an untried action and a random belief
about its payoff (+10 to -10) with confidence 0.5.
• Communication: Two agents meet. Randomly pick a situation to
discuss and an action in that situation. Either one agent is ignorant
of that action and acquires it and its confidence from the other (but
reduced by 0.1 at “second hand”) or both know the action. If they
agree the payoff, that boosts the confidence of both by 0.1. If they
disagree, they average the payoff but keep their confidence. Any
confidence not acquired by direct experience is capped at 0.9.
• Demography: Agents live 80 time periods, die and are replaced
with a copy of a randomly selected survivor.
12. 11. The model 3
• When encountering situations, decision over the
actions is strictly rational. The agent multiplies
the payoffs they are aware of by the confidence
and chooses the best.
• If it happens that an untried action is chosen
then the subjective payoff and confidence are
adjusted to the objective payoff and 1
respectively. (Needless to say, this “quality”
information can then spread.)
13. 12. Why model?
• How does this system behave in respect of 1)
number of actions known by the average agent
(coverage), 2) number of payoffs known
correctly by the average agent (accuracy) and 3)
closeness of best subjective payoff to best
objective payoff (optimality?)
• Two situations in the world with eight possible
actions each, ten simulation runs of two
thousand simulated time periods each.
14. 13. Results
MEASURE MEAN MINIMUM MAXIMUM
COVERAGE 16 16 16
ACCURACY 4.2 3.0 5.2
OPTIMALITY 0.59 0.29 0.91
15. 14. Implications
• Even given a very long time and “innovation”,
agents neither develop an accurate perception
of all choices nor do they get close to the
“optimum” choice for any given situation.
• A possible example of a belief trap (Mackie).
Given their confidence in their beliefs, agents do
not try the actions that might falsify them. The
“community perception” settles on suboptimal
actions which are not falsified in decision or
communication.
16. 15. An “application”
• Hegemony: The idea that one discourse can dominate
the way that things are done. Used in a narrative sense to
look at, for example, how an (empirically false) belief in
meritocracy can serve the interests of capitalists.
• Possible domain: Sexual orientation and roles.
• Now we have two “types” of agents, one type in a
majority (90%). To show the effect most clearly choices
for one group are “negative” for the other i. e. 5 for
majority group is -5 for minority group. Type is assumed
to be private so one cannot condition payoff claims “by
type”. All else as before.
18. 17. Implications
• Without any coercive power or media control, minority
groups may make choices that are less beneficial to
them owing to the hegemonic discourse
(“heteronormativity”.)
• Surprisingly (and very tentatively) the minority group
has better accuracy even with worse optimality.
• It does not seem that such a system could readily be
characterised analytically (except perhaps with
simplifying assumptions that took it outside plausibility.)
19. 18. Roles in sexual orientation
“Butch” and
“femme”.
20. 19. A change of direction (or maybe not)
• Specification: Deciding what “things” go in the model i.
e. here mixing but not social networks.
• Calibration: Assigning plausible values to parameters in
the specified model with best empirical support. For
example, how confident are we about our own ideas
relative to those of others?
• Validation: Does our model output simulated data that
correspond (in a manner to be assessed) to equivalent
real data? If you “mirror” this model in a game or role
play, do the model assumptions generate patterns of
choices we actually see? “Natural” setting?
21. 20. Falsifying methodology
• Unlike fitting (or experiments designed on the assumption
that is a theory is complete/correct?) models can be
definitively falsified.
• Given the specification, calibration and abstraction
assumptions you make, can you mirror real data
“adequately?”
• Odd cultural glitch: ABM is pro data in principle but quite
negligent/dismissive of it in practice.
• Also, no strong guidance yet on where you have gone
wrong if the model is falsified.
22. 21. Not just a modeller fantasy …
Hägerstrand,
Torsten (1965) ‘A
Monte Carlo
Approach to
Diffusion’,
European Journal
of Sociology,
6(1), May, pp. 43-
67.
23. 22. A topic for discussion
• In Sociology, this falls out quite nicely. Surveys give us patterns
to reproduce (i. e. ethnic compositions of universities) and
qualitative interviews give us some insight into decision
processes which we can abstract into models.
• But how do experiments and “theories” fit into this picture? Does
it matter what kind of experiments i. e. “positivist” attribution of
theory to participants rather than accessing subjective accounts
directly? (Recording pair play?)
• Is the disciplinary division between subjective accounts and no
psychology and “experimental” psychology and no subjective
accounts good for science?
• Could modelling be one way to bridge this gap?
24. 23. Notes
• I am more than happy to talk about what is
wrong with ABM!
• I am also interested in talking about other
possible applications in JDM.
25. 24. References 1
• Chattoe-Brown, Edmund (2009) ‘The Social Transmission of Choice: A
Simulation with Applications to Hegemonic Discourse’, Mind and Society,
8(2), December, pp. 193-207. [The article for the model presented here.]
• Chattoe-Brown, Edmund (2013) ‘Why Sociology Should Use Agent Based
Modelling’, Sociological Research Online, 18(3). doi:10.5153/sro.3055
[Example based discussion of methodology and uses of different data types.]
• Chattoe-Brown, Edmund (2014) ‘Using Agent Based Modelling to Integrate
Data on Attitude Change’, Sociological Research Online, 19(1).
doi:10.5153/sro.3315 [An imperfect but sincere attempt to “walk the talk” on
calibration and validation.]
• Chattoe-Brown, Edmund (2021) ‘Agent Based Models’, in Atkinson et al.
(eds.) SAGE Research Methods. doi:10.4135/9781526421036836969 [More
recent discussion on how we might progressively “home in” on a validated
model and why we should.]
26. 25. References 2
• Aron, Arthur (1988) ‘The Matching Hypothesis
Reconsidered Again: Comment on Kalick and Hamilton’,
Journal of Personality and Social Psychology, 54(3),
March, pp. 441-446. doi:10.1037/0022-3514.54.3.441
[Good empirical example of ABM used with “real”
psychology. See also prior articles in this exchange.]
• http://diposit.ub.edu/dspace/bitstream/2445/131122/1/673
195.pdf [Good discussion of issues around calibration
and validation – at least with stylised facts - for a prima
facie economic domain - stock markets. Also flags
previous research.]