Machine Learning, Neural Nets, AI and alternate Trilateration implementations for Bermuda

[agittins]

This discussion topic is for folk to share their experiences with different ML/AI implementations that hook into Bermuda.

While Trilateration is (and always has been) on the roadmap for Bermuda, my personal focus for doing that is definitely on the Analogue / OrganicIntelligence / HumanLearning approach :-) My main reasons for that are:

• I want Bermuda to provide a great out-of-the-box user experience, even for the minimum system specs required by Home Assistant. This precludes the use of libraries with specific CPU instruction support beyond what HA already dictates, at least for core functionality. I'm open to having features that can be turned on where the host system can support them, though.
• I believe that we are yet to properly optimise what's possible by applying domain-specific knowledge and reasoning to the problem, possibly even a little intuition. And with better-conditioned source data, I think that more generalised algorithms on top of that can provide even more useful results, so it's a worthwhile area to focus.

So... this means that there's a desire (and definite benefits) to apply some buzzword-compliant ML stuff to the data Bermuda collects, and I want to encourage and support those efforts.

I will maintain a section on this post linking to people's projects so help discoverability, and encourage folks to share what they're doing and invite others to try it out.

Other projects that tie in to Bermuda's data:

ml2mqtt

• Type: Add-On
• Data Interface to Bermuda: MQTT
• Algos: Random Forest, KNN, and more
• Other: Node-Red

@donutsoft has created a HA Add-On that lets you apply various models, train

Currently it supports Random Forest and KNN, but I'll be adding a few more algorithms soon. No mathematics or in-depth knowledge of machine learning algorithms required. It's pretty much entirely just point and click.

BPS - BLE Positioning System

• Type: Custom Integration
• Data Interface to Bermuda: HA state engine
• Algos: numpy/scipy least squares method

BPS is not too dissimilar to where Bermuda is heading over time. @Hogster has created a graphical front-end and full trilateration of all tracked devices using the scipy library. Proxy locations are defined by the user via the GUI, and the tracked devices are localised based on those locations and the distance measurements.

Bayesian predictions #561

• Type: Configuration of core integration
• Data Interface to Bermuda: native / HA state
• Algos: Bayesian probability

@jackjourneyman has detailed how he has set up sensors which use the Bayesian features available in core, combined with collected observations, to predict a person's location. It can observe multiple devices, and accommodates the cases where different devices are used at various times. Check out his description of his setup at #561

Garnser's Bermuda trilateration

• Type: Standalone python app
• Data Interface to Bermuda: MQTT
• Algos: classical trilateration (scipy least squares), Random Forest

Includes some great multi-floor visualisation, might require a bit more effort to initially set up and deploy.

If you have an implementation you'd like to share (or corrections), please comment below and I'll add it!

[erkr]

Hi @agittins
Thanks for opening this discussion on ML.
My two take aways using ML are:

• don't use the raw data's inputs:
  • Normalise the inputs by modelling what is deterministic. In this case the calibration of proxy signals
  • replacing a proxy shall not invalidate the trained network
• ML's are great to process the part that is to complex to model:
  • I.e. the influence of moving bodies, furniture and device orientations
  • walls and their thicknesses are a gray area in this case, as modelling is a lot of work, and it won't make the ML more complex