Many classical psychophysics models are simple linear models or linear models with a polynomial basis. We should include them as baselines in AEPsych. For example, psignifit contains many of them (https://psignifit.sourceforge.net/PSYCHOMETRICFUNCTIONS.html). In the language of the psignifit docs, our "core" is a GP model and our sigmoids are the various objectives here https://github.com/facebookresearch/aepsych/blob/main/aepsych/acquisition/objective.py.
To implement a basic linear model code, we will need a class that extends AEPsychMixin that uses a latent linear model with learnable parameters instead of a GP. At minimum, it needs to support a fit method, and a posterior method so that we can integrate it with acquisition. This issue will require a bit of research to understand where and how to implement everything -- we are happy to support.
Many classical psychophysics models are simple linear models or linear models with a polynomial basis. We should include them as baselines in AEPsych. For example, psignifit contains many of them (https://psignifit.sourceforge.net/PSYCHOMETRICFUNCTIONS.html). In the language of the psignifit docs, our "core" is a GP model and our sigmoids are the various objectives here https://github.com/facebookresearch/aepsych/blob/main/aepsych/acquisition/objective.py.
To implement a basic linear model code, we will need a class that extends
AEPsychMixinthat uses a latent linear model with learnable parameters instead of a GP. At minimum, it needs to support afitmethod, and aposteriormethod so that we can integrate it with acquisition. This issue will require a bit of research to understand where and how to implement everything -- we are happy to support.