Lecture 2 Flashcards
1
Q
Why is probability bad?
A
- Leads to negative outcomes and negative utility
- Can be uncertain
- To avoid this, you must gather other information
- Use cues for prediction
2
Q
What is Brunswik’s lens model?
A
- Paradigm where there is something in the world and you want to predict it
- You then look for cues in the env that do a better/worse job to help predict what you need to.
3
Q
How accurate/reliable are the cues?
A
- Each cue is associated to a certain extent with the to-be-judged criterion
- Association is causal
- Poor cues have no association with the criterion e.g flipping a coin to determine the weather
- Ecological Validity
4
Q
What is the multiple cue probability learning tasks?
A
- In the model, you can combine multiple cues
- You need to figure out what the valid cues
- Learning from outcome feedback - takes many trials & hypothesis testing
- Easier if it is causal, and easier if learner can intervene and actively investigating or if you can intentionally manipulate it
- e.g Given the cues and told to guess if its raining, and then told if it rained the next day, continuing with trial and error
5
Q
How much information should we sample?
A
- More is better but info is expensive through time/effort
- Beyond a certain point additional info does not help make a decision through cues
- Too much can result in satisficing through fatigue, bias and boredom
6
Q
How do we see search for information?
A
- Consider each option in turn - holistically
- Search by attributes = allows easy elimination for things that fail 1+ criteria BUT does not work so well when there is conflict between attributes e.g high res but poor battery
7
Q
What is cue utilisation?
A
- How often do people use the cue to make the prediction
- As people learn, cue utilisation is associated with cue validation
8
Q
What are the compensatory strategies for combining cues?
A
- Combine ALL info
- Weights each cue
- Creates a linear model: adding these weighted cues = optimised via multiple regression and not based on human weightings
- Bootstrap models formalise the judge’s use of cues = consistent application of a judgement policy
- Linear models outperform human experts
- We bootstrap because we have good concepts of good cues, but we are not good at combining and weighting cues
- Humans are inconsistent, machines are not, humans are good at identifying and assessing the cues but we are not good at integrating info
- Humans gather info as we know what is important, but computers integrate it to inform a judgement
- One cue can be compensated for by another - final prediction is a combo of all of them
9
Q
What are the non-compensatory strategies for combining cues?
A
- Not processing all cues into account
- Use heuristics
- Focusing on a subset of cues - following bounded rationality
- Elimination by aspects
- Take-the-best: pick one attribute and make decision based on that attribute: binary choice situations: based on recognition principle = when uncertain if you recognise one, you pick that one, AND Reference class of cues: where people search cues in descending order of feature validity until they find one that is different
10
Q
Why has coherence linked to prototypicality?
A
- Coherence means you do not violate laws of probability
- EX: Dylan is a random man in Wales, what is more likely? He has three heart attacks or he has had three heart attacks and is over 55. One is more intuitive than other
- More likely to label a man in his 50s as someone who will have a heart attack due to prototypicality and centrality (think about someone likely to have heart attack, and there is a central tendency)
- Then you judge the probability to the one individual to the prototypical individual in your head
11
Q
What is a Dutch book?
A
- Exploitation when you have incoherent probabilities, you lose money and others gain utility from you
12
Q
How do we organise correspondence with novel situations?
A
- Correspondence is the extent of our judgements reflecting the properties of the world
- With novel situations, there are no references to see rate of probability in world, thus correspondence
- Could use history e.g Brexit and UKIP voters
- Unique situations: lots of frequency tabulations across concepts that would bare on it but not sure which one is best or how to combine them
13
Q
What is Bayes’ Theorem?
A
- How to update a hypothesis In light of new data
- Psych model that you should start with beliefs
- P(h/D) = P(D/h)P(h)/P(D)
- P(h/D) = posterior probability of h
- P(h) = prior probability of h
- P(D/h) = probability of observing D given that h holds
- P(D) = probability of observing D
14
Q
What is Base Rate Neglect?
A
- Take a question we find it hard to work out the probability, we use another probability to replace that value
- To the extent we have conditional info about something about a probability, we need more base rate info
- Attribute substitution: easily accessible information is subbed for difficult to compute answers
- Insight into problem structure helps presenting the problem in frequencies rather than probabilities
- Conditional info overrides base information: as one becomes easier to estimate
15
Q
What is the conjunction fallacy?
A
- A conjunction cannot be more probable than either of its conjuncts
- People make their probability judgements on the basis of representativeness
- Gives information about a man, then asked people to rank traits of the man according to their likelihood
- When one thing is more likely than another when it cannot be e.g P(accountant & jazz) < P(Jazz)
- BUT probability can be interpreted non-mathematically through plausibility, credibility or typicality.
16
Q
What is Disjunction Fallacy?
A
- When people are asked questions about probabilities, people find it hard to answer
- If you take a questions and disjunct it (chop in half and ask about components), you can overestimate the thing as a whole e.g. homicide or homicide as acquaintances and strangers
- Mathematically, they should be equivalent = framing of question can cause overestimation
17
Q
What is Support Theory?
A
- Support theory hinges on subjective judgements of probability are description dependent (framing), they derive from judgements of support (how strong the evidence is), and lead to subadditivity (disjunction probability).
- The degree to which evidence SUPPORTS a conclusion is not the same as the likelihood that the conclusion is true given the evidence
- Extent to which evidence supports the theory - amount of increase of belief in theory
18
Q
Representativeness as attribute substitution?
A
- Easily accessible information is substituted for difficult to compute information e.g how similar to the prototype (To what extent does A resemble B? & how likely is it that A belongs/originates from B)
- How similar is it to my centrality/prototypicality
- Examples of representativeness include Gambler’s fallacy, stereotyping, Base rate neglect, misconceptions of chance
19
Q
Availability as a heuristic?
A
- Instead of judging the probability of an event occurring, we are swayed by ease of retrieval of an instance or association
- Does not involve actual recall, but an assessment of the ease of which the operations could be performed
- Availability is ecologically valid ad frequent events are easier to recall than rare ones, but can be biased sometimes through media reports or stereotyping
- Can be right e.g which city is bigger: Tulsa or Chicago?
20
Q
What are the heuristics and biases in decision making?
A
- Errors of judgement are often systematic rather than random, manifesting bias rather than confusion
- Reasons for biases are heuristics but heuristics tend to be very good, often easier than Bayes Theorem
21
Q
What are the key processes in making judgements?
A
- Discover info: how do we know where to look
- Acquiring and searching through information
- Combining info
- Feedback
22
Q
How to decide how to decide?
A
Trade-off between accuracy of a decision and the effort of making it
23
Q
Can we have too much choice?
A
Overabundance of options decrease motivation to choose and decrease satisfaction with the chosen option
24
Q
What is cascaded inference?
A
One makes a sequence of connected inferences
