David Shanks University College London, UK

1
David Shanks University College London, UK
Learning and Decision Making An Overview of the
Landscape
General theme How do people learn to make good
choices in decision environments? What are the
main research questions in studying decision
learning? What factors affect the likelihood of
learning an optimal decision strategy?
2
eg medical symptoms?diseases, weather?price of
orange juice, company data stock?market changes,
etc. Participants receive a payoff for each
correct decision or for their overall decision
accuracy.
3
Company A
Company B
Bulk of operations? UK
Bulk of operations? US
FTSE/Nasdaq? FTSE
FTSE/Nasdaq? Nasdaq
Established company? No
Established company? No
Employee turnover? Low
Employee turnover? High
which companys shares are more likely to
increase?
4

• What research questions arise in the study of
  decision learning?
• Search How do we find and/or construct the cues
  and cue values necessary for informing our
  choices?
• Integration How do we combine cue information?
• linear/nonlinear, heuristics
• How do different forms of feedback and
  feedforward affect decision learning?
• What form does knowledge take?
• connections, exemplars, prototypes, rules
• To what extent do people have insight into their
  decisions?
• Can people learn to make optimal decisions? If
  not, what sort of biases are they prone to?

5

• Search How do we find the cues and cue values
  necessary for informing our choices?
• Very little research on discovering/constructing
  cues
• Klayman (1988)
• discovery is heavily reliant on outcome
  feedback
• much better when the person can intervene and
  design his/her own experiments.
• More research on search amongst available cues,
  eg process tracing techniques (elimination by
  aspects, satisficing).
• Stopping rule when should we stop searching for
  additional cue information?
• How do we choose how to allocate our attention
  across cues?

6

• Integration How do we combine cue information?
• Linear/nonlinear
• Heuristics
• People often thought to have a preference for
  linear forms and a limit on the number of cues
  they can combine. Evans et al (1995) doctors say
  they use more cues than they actually do.
• But clearly experts can combine many cues
  nonlinearly (Ceci Liker, 1986).
• Perhaps sometimes we dont integrate at all, but
  use noncompensatory heuristics such as
  Take-The-Best and other varieties of one-reason
  decision making? In many environments, such
  heuristics are optimal or near-optimal.

7

• How do different forms of feedback and
  feedforward affect decision learning?
• Feedforward effects of instructions, task
  understanding, scale compatibility.
• Outcome feedback knowledge of the value of the
  outcome. Hard to learn from this, but plenty can
  be learned.
• Cognitive feedback (Balzer et al, 1989)
• task information (eg relations between cues and
  criterion)
• cognitive information (eg cue utilization)
• FVI (eg achievement)
• history

8

• What form does knowledge take?
• Connections
• Exemplars
• Linear model (prototype)
• Rules
• A huge research area
• Evidence for exemplar-based processes is
  overwhelming (Nosofsky).
• Also much support for connectionist error-driven
  learning as the fundamental mechanism of human
  learning
• so perhaps connectionism is a way of
  implementing exemplar storage (eg McClelland
  Rumelhart, 1985)?
• Evidence for rules and for multiple strategies
  more controversial (Johansen Palmeri, 2002
  Juslin et al, 2003).

9

• To what extent do people have insight into their
  decisions?
• Important to differentiate between insight into
  the task vs insight into ones policy. People
  often can recognize the policy they used.
• People often find it difficult to verbalize their
  reasons, but can under some circumstance indicate
  fairly accurately the weight they assigned to
  each cue (eg Harries Harvey, 2000). They can
  also report idealized cue weights (ie insight
  into the task).
• Correlation between subjective and tacit policies
  is often low.
• Are decision strategies employed deliberately or
  automatically? If the latter, then unlikely to
  yield insight (eg Bechara et al, 1995 Dienes
  Fahey, 1998 Nisbett Wilson, 1977).
• Task properties, eg scale compatibility. Insight
  is greater when the cue and response dimensions
  are the same.

10

• Can people learn to make optimal decisions? If
  not, what sort of biases are they prone to?
• Certainly people can behave near-optimally in
  repeated decision environments (eg Shanks et al,
  2002 Kelley Friedman, 2003).
• But even in these cases, biases are detectable
  (eg base-rate neglect Goodie Fantino, 1999).
• Paradox classic JDM studies (eg Meehl) indicate
  that people are outperformed by very simple
  linear models, yet research in cognitive
  psychology (eg categorization Ashby Maddox,
  1992) reveals near-optimal, nonlinear behaviour.
• Thus, perhaps people have the competence to make
  optimal decisions in virtually any domain, and
  perhaps they often fail to do so because of
  insufficient or inadequate exposure/motivation/fee
  dback etc?