Description
Models how users infer hidden states, categories or intentions as Bayesian inference over priors and evidence, predicting judgements and where they will go wrong.
Bayesian cognitive models treat perception and judgement as inference: people combine prior beliefs with the likelihood of observed evidence to form posterior beliefs. An ideal observer model computes what a perfectly rational user would infer given the information an interface provides, setting a benchmark for human performance. Deviations reveal either faulty priors (for example, from other apps' conventions) or information loss in the interface. Applied to UX, such models can predict how users will interpret ambiguous icons, how many examples they need to learn a category, how they update trust after an AI error, or how they estimate quantities from charts.
Attributes
- Discipline
- Cognitive Psychology & Perception
- Loss type
- Latent
- Stage
- Design
- Data
- Quantitative
- Effort
- Weeks
Steps
- Define the inference problem the user faces, the hypothesis space and the evidence the interface provides.
- Elicit or estimate users' priors through questions or pilot judgements.
- Specify the likelihood of evidence under each hypothesis, given the interface's cues and noise.
- Compute posterior predictions, or an ideal observer benchmark, for each design condition.
- Collect human judgements and compare to model predictions, fitting a small number of parameters if justified.
- Use discrepancies to redesign cues or set expectations that shift priors.
Sources
- Griffiths, Kemp & Tenenbaum, Bayesian models of cognition (2008), in Sun (ed.), Cambridge Handbook of Computational Psychology
- Tenenbaum, Kemp, Griffiths & Goodman, How to grow a mind (2011), Science
- Anderson, The Adaptive Character of Thought (1990)
- Cognitive Modeling: Nine Paradigms (slide deck)
Also known as
- Rational analysis
- Bayesian models of cognition
- Ideal observer analysis
Categories
- User Research
- Cognitive Psychology & Perception