Craig Sanders
Craig Sanders
The [fully worked visual psychophysics experiment](https://aepsych.org/tutorials/data_collection_analysis_tutorial) demonstrates all the main features of AEPsych, but it has only been tested with the old backend. We should make sure it works with...
Ax uses a trial_index to keep track of which parameter values belong to which trials, and this trial_index needs to be referenced when completing trials. Before the Ax refactor, the...
Our website contains [this page](https://aepsych.org/docs/api_overview), which outlines and diagrams our old API. We should update this to include our new Ax backend. It should explain how our API relates tot...
We should have a tutorial that is more advanced than our [current one](https://aepsych.org/tutorials/data_collection_analysis_tutorial) that demonstrates some of the new features enabled by Ax. Ideally this tutorial would demonstrate a psychophysics...
The [data collection and analysis tutorial](https://github.com/facebookresearch/aepsych/blob/main/tutorials/data_collection_analysis_tutorial.ipynb) demonstrates the core features of AEPsych, but it uses the old backend, and we would like to show off the features enabled by the...
The new Ax-enabled models can be plotted using [Ax plots](https://ax.dev/tutorials/visualizations.html). These plots have many advantages over the old AEPsych plots in that they are interactive and can handle multiple dimensions,...
The old AEPsych backend supports warm-start fitting for models. That is, the previous model's hyperparameters can be used as the starting point instead of re-fitting from scratch. This makes the...
Currently `PairwiseGPModel` only works with the old backend, so it cannot take advantage of Ax features. We would like a new pairwise model that works with the Ax backend, while...
Currently tell messages have the following format: ``` {'type': 'tell', 'message': {'config': {'par1': [123], 'par2': [456]}, 'outcome': 789, 'model_data': True}, 'extra_info': {}} ``` This format only allows for a single,...
# 🐛 Bug Calling `SingleTaskVariationalGP.posterior` with an `input_transform` raises a warning, whereas the equivalent call with `SingleTaskGP` does not. I'm not sure if `input_transform` works correctly with `SingleTaskVariationalGP` or if...