An integration for training a bayesian sensor. For instance for s bayesian sleep sensor it would be great to have a “device” that includes a named pushbutton that would store the states of any selected devices/sensors when the “Train Sleep” button is pushed. The button would be pushed when the occupant goes to sleep and the trainer would remember the states of the selected devices/sensors. After the course of a few weeks the data collected could be analyzed to see trends and the historical data would then be translated into a yaml output for use as a bayesian sleep sensor.
Remember to vote for your own idea. 
You might want to add a comment here:
An interesting idea, I think overall this probably wouldn’t be the way to do it. But I plan to (eventually) do something similar using the history UI function, one could always use a button to mark timestamps. It can be discussed in the thread linked above (of which I am the author and Codeowner).
Thank you - It seems the github discussion is closed
Having a UI for this would be the real starting point for all of this. After that, maybe being able to setup something that pings you with a notification when th state changes, and giving you two links to click, one would b for if it was expected(right), and the other for unexpected (wrong). This could get logged somewhere about what value or values caused the unexpected result. If you do this enough the user could change the formula a bit. But an enhanced feature would be that it looks at the values and either uses some math to change them so they would have not caused the unexpected behavior, or maybe feed all this data into an AI model and suggest a new set of formulas.
I don’t think Bayesian “learns” in the way AI does - it simply calculates the probability of something being true based on very specific observations. In a way I suppose it’s now rather outdated, but compared with AI it’s extremely economical and surprisingly accurate (neither of which can be said of AI).
I’m not sure UI would be much help in setting it up. Where most people go wrong is guessing the “observation” values.
Your intelligent Bayesian integration would probably require its own PC! 