Practical Identifiability: Quantifying uncertainty in parameter estimation
Ph.D. Student Nicholas Lam
Department of Mechanical Engineering, University of Canterbury
Time & Place
Tue, 08 Jun 2021 11:00:00 NZST in E12, Engineering Core
All are welcome
Parameter estimates from fitting a model may be highly variable due to real-world limitations in data quality and quantity. This is a practical identifiability issue, where a wide range of parameter combinations may generate model outputs with similarly good fits to measured data. (This is analogous to not being able to tell the difference between me and Seigan Hayashi, see Figure 1)
Dynamic models such as a spring-mass-damper may have structurally non-identifiable parameters where you cannot find [𝑚,𝑑,𝑘], even with perfect measurements; the parameter solutions are not unique due to the model structure. In a practically non-identifiable case, the problem is more nuanced; you may have a model with an optimal parameter set, but also have a wide range of parameters that fit within an acceptable level of error. Discrete and/or sparse sampling protocols, measurement noise, and model inputs can all influence the severity of practical non-identifiability.
Using practically non-identifiable parameters can be problematic: the wrong parameter estimates may lead to incorrect predictions when extrapolating a model or simulating unmeasured states. Furthermore, this issue may not be clear when you only have one optimal fit from a criteria such as least-squares. Despite an apparently successful parameter identification, there may be an underlying practical identifiability issue that is hard to detect.
During this talk I will introduce some of the common practical identifiability analysis methods, describe how these can detect issues with a model or data, and discuss how we can use these results to improve experimental design. Maths will be kept at a minimum, with plenty of graphs and plots to show what is happening!
Supervisor: A. Prof. Paul Docherty