I would like to start developing a machine-learning based platform that would create a "Digital Twin" of humans by modeling their preferences, with the goal of controlling and monitoring devices. I like calling it "Nest for the rest" because it would be a generalization of Nest's approach to learning and adaptation, which could be employed in several other scenarios. Here you can find a very simple example of what I did so far: http://slides.com/innumero/smarthomeai-01
I find the challenge very interesting (despite the fact that I'm not much into automation) and I'd like to know if someone else would be interested in collaborating, providing suggestions or at least testing prototypes. Also, I'm not aware of the potential market: does anybody have an estimate? Within the smart-home segment, I guess the market is fairly small, but Gartner identifies digital twins as a Strategic Technology Trend for 2017, with the potential to "become proxies for the combination of skilled individuals and traditional monitoring devices and controls". Can anybody comment on that? What's your take?
I think you will find that at least in this community there are people of two minds when it comes to home automation:
The automation should be determinisitic and predictable.
The automation should learn what I want and do it.
Certain types of automation work better with the former (e.g. alarm systems) and certain types work better with the latter (e.g. HVAC).
The challenge with determinisitic is that one has to ultimately define all the possible states and rules to describe what should happen at state transitions. Most of the time this is relatively simple but it cna become untenable really quickly.
The challenge with the learning approach is three fold. Firstly it has to work well enough from the start to be functional. Or else it won't have a chance to learn because no one would use it to give it a chance. Secondly you have to decide what factors are relevant to feed into your learning algorithms. For example, what temperature it is in the house probably doesn't have much to do with whether you turn on the hall light, but the time of day could be very relevant. Thirdly who takes precidence? If there are multiple people in the house, it needs to learn from all of them to discover their preferences. But when multiple people are home only one can take precidence.
Its an interesting problem but I'm not certain there is one universal approach that works for all types of automation. HVAC, lighting, perhaps ambient music, maybe home entertainment systems (e.g. user one likes the lights dimmed when watching a movie, user two likes them turned all the way up) might make good candidates. Alarm systems, irrigation systems (which would use a different type of algorithm not based on a user's preferences) and the like do not make good candidates.
And even for the automation types that are good candidates, the appraoch to the learning may not be the same.
The fact that you are using MQTT as the interface to the AI makes it very easy to integrate with OH. You can either set up specific Items that pub/sub to the topics, or set up a MQTT Event Bus which will send everything in OH to the AI.
It is also possible to interface with NodeRed if you wanted to go that way.
Personally, as a user, I'd be most interested in understanding how I would use such a system in practice. How would I communicate my preferences to the AI (I assume it would be the sitemap) and how does the AI drive my automation (does it inject messages or does it only respond to requests?)
Deterministic/adaptive automation: you're correct, but I think there might be some need to "teach" rules by example as an alternative to coding. Not everybody is able or willing to configure a rule engine and that might be a hindrance to the wider adoption of home automation.
"it has to work well enough from the start to be functional": that's right: fortunately some learning algorithms allow to adapt the learning rate, to make them learn faster at the beginning of the training. Alternatively, the learning system could be provided with a bespoke initial "memory" to start with.
"you have to decide what factors are relevant to feed into your learning algorithms": that's not completely correct. Some algorithms can automatically detect relevant inputs and ignore the others.
"who takes precedence?": I guess that's an issue that even a human butler would face, so, although it's relevant to the case, it's not really specific to AI-powered automation.
"I'm not certain there is one universal approach": no there isn't. A working system might need to be "assembled" from a set of independent and occasionally communicating subsystems. Maybe it's time to dust off Rodney Brooks' subsumption architecture.
MQTT/NodeRed: I'm already using MQTT and NodeRed for "dry" simulations and I'm now looking for the right project to implement a real-world prototype.
"how I would use such a system in practice": My idea would be to have the AI learn from the human user by subscribing to the events published to MQTT topics, automatically find the relevant inputs (e.g., time of the day, day of the week, weather conditions/forecast, temperature, occupancy, etc) and simply replicate the patterns that can be detected with high confidence (that's what could be called a digital twin of the human user) by publishing to MQTT topics. In a second phase, it could be possible to apply model predictive control to do more fancy things. In no case would coding or direct configuration be needed, but the AI engine might need to ask questions from time to time to disambiguate cases for which it has little confidence in the computed prediction.
I'm experienced in ML/AI, but not much in home automation, that's why I'm asking the OH community suggestions: what's the lowest-hanging fruit? What's the worst pain point I could try to relieve?
True, but my point is at least until there is some major unforeseen breakthrough, one cannot get away from configuring a rule engine. A system like this can minimize the need for coding rules but not completely eliminate it.
In my experience, which is admittedly in a vastly different and more difficult domain (i.e. computer security), such algorithms are not very good at deciding what is relevant and what is not. Perhaps the somewhat less noisy and caotic world of home automation they would work better. But I've been burned before so am skeptical.
The problem is a bit more nuanced than that and this is a pet peeve of mine. Far to often designers and builders of home automation systems (from the DIY to the commercial offerings) make two fundamentally incorrect assumptions:
Their system is the only one operating.
There is only one "user".
The first one is sometimes addressed by releasing an API and is almost the main raison d'etre for OH.
The second one is almost completely ignored by most and it results in user interfaces that are awkward at best and unusable at worst for any but the primary user of the house. You will often see that referred to around here as the WAF (Wife Acceptance Factor, which would probably be better put as SAF Spouse Acceptance Factor).
To me the answer "that's a universla problem so I'm going to ignore it" is unsatisfactory. There has to be some sort of answer or approach, even if it is something like the Nest approach which simply ignores multiple users and treates all inputs the same. But make that a deliberate design decision, don't just ignore it because it is hard. And do realize that using the Nest approach and treating all users as one can and may limit the sorts of automations that can be learned.
That's a hard question to answer. Every home automation deployment is a unique system with unique requirements. One person's major pain point is another person's don't care. For example, many users spend huge amounts of time and effort getting their lighting working the way they want it. I have three lamps which I'm happy controlling with simple timer type rules.
That being said, I would probably prove the system out using lighting, particularly colored lighting. This would give you a problem space rich enough to explore the system (on/off, dimmer values, color) yet simple enough to set up in OH without equipment. You can create a model house (see the demo) with lighting, weather, time, etc but not have it connected to anything physical which should give you a low barrior to entry. Using the REST API you could even write scripts to exercise the house as if a person where actually interacting with the devices to drive your machine learning, at least in the initial development stages.
Agreed: coding might be completely avoided only in some cases, in other ones a point-n-click interface could help configure the rules. Of course, each solution would be very case-specific... no silver bullet here. Anyway, I'm not after a universal solution: I'm happy with little problems and simple, usable solutions.
Computer security is a harder nut to crack... not being predictable is the name of the game, whereas at home we can hopefully expect to have more or less regular patterns.
Honestly, I doubt that even a human could solve this class of problems in a way that does not entail some level of compromise between the people sharing the same living space. I think most companies prefer to tackle simpler problems, being a R&D investment in the harder ones more difficult to justify.
I already have a sort of virtual bench of the kind you describe, but the main problem I see now is to solve a real problem, instead of tinkering with sensors and algorithms, which can be fun as a hobby, but pretty useless as a business.