Model Uncertainty Flashcards
Unknown
Define uncertainty/ignorance
A lack of knowledge or the unknown
What are the four types of uncertainty?
Observation error, parameter uncertainty, prediction uncertainty, structural/model error
Describe observation error
Calibrate to data but data is not exact
Measurement is done wrong.
Describe parameter uncertainty
Many models (parameters) can fit the same data. This results in uncertainty because it is unknown which set of parameters should be used.
Describe prediction uncertainty
Many models means that there are many predictions. Use histogram to minimise uncertainty.
Describe structural/model error
Fundamental wrongness in model formulation. Known unknowns (assumptions, etc), unknown unknowns (lack of knowledge).
What is an ensemble and why is it useful in modelling?
A collection of models. It can help identify percentiles and means of models with different parameter values
What is the difference between a prior and posterior?
A prior distribution uses user knowledge about the parameters. A posterior distribution uses the data to decide which parameters are likely.
Why is it not enough to find and use a single ‘best fitting’ model?
The ‘best fitting’ model is not robust. It is possible that it is sensitive to small changes in parameters - use an ensemble.
Describe how an inappropriate assumption can be described as structural error
It is a known unknown. An assumption that a process does not affect the model, when it does, is an error within the model and will have an impact on the reliability of the output.
Can we account for structural uncertainty by constructing a posterior parameter or prediction distribution?
No it deals with parameter uncertainty not invalid assumptions etc
How are parameter distributions related to prediction distributions?
Parameter distributions help identifying likely sets of parameters based on the minimised objective function. These probabilistic parameter combinations are then fed into the model to create prediction distributions
In uncertainty analysis, how should we account for the situation that some observations have larger errors than others?
Penalise larger errors by using Lp norms to help reduce the noise of the model
What is the role of the misfit function in constructing a posterior distribution?
As the misfit increases the probability decreases. Exponential to the power of the misfit function/2
How should you express an uncertain prediction from a model?
With error bars and a disclaimer
