This is a really interesting problem, and one which I have been considering for a while.
I’m currently in ‘limbo’ with ST as I am in the UK and doesn’t quite work as expected - but I am no novice to H/A having had a more professional albeit far more proprietary and expensive system in place at for the past 10 years.
When approaching H/A, we tend to look at our lives and routines as fairly static, we like to have the lights come on based on various events, rooms in the house heated at various times to various temperatures etc etc.
The reality is, when looking at our routines over a short period of time, they are orderly, but when considered over a longer timeframe they are subject to evolution. Like nature, the evolution can be slow and barely perceptible and at other times sudden and dramatic - in your case a baby has appeared and overturned the applecart but even without a baby things change, such a working hours, habits and preferences.
In all the above, a static model works for a while, you might make the odd change here and there as a direct consequence of the slow evolution of order - but when a big change happens it almost feel like everything must go and be started again from scratch.
What this ecosystem is crying out for is more than the simple rules we have; it need an adaptive system based on the historical behaviours with the ability to learn what we like to do, when we like to do it and being ready to easily adapt to change.
For example, in your case, your mother in law has the need to be able to say to the house - uh uhh, when you turned that light on in here, at this time of night, when the rest of the house was asleep, that was a bad idea. In other circumstances, when the house sees you enter a room and manually turn on the light under a differing set of conditions it needs to think to itself, aha I could have done this automatically rather than have the user press the button.
Another example might be the dining room. In my house the dining room rarely gets used unless we are having guest round for dinner, but when it does there are some indicators in the activity data which could be used to help the house learn what it needs to do. It could have sensed that there was an unusual amount of activity in the kitchen, there was also activity in the dining room as it was prepared for a dinner party. Maybe there was more movement in general in the house as guests arrive, with the doorbell being rung several times and lots of activity in the hallway. What it could do is sense that all these things are happening and look at past behavioural models and thought to itself - hah, i see last time all this happened the dining room heating was turned on and the household went to bed later than normal. It could then use this to pre-empt my manual intervention in turning the heating on in the dining room and delay turning on the heating in the bedrooms until all the hubbub downstairs quietens down.
Whilst this all sounds quite complicated, it really isn’t, there are lots of uses for machine learning - I work in a field which uses it a lot to extract meaning from big data - with some work it could be applied to H/A. There are two things which could make this magical:
1 - provide a mechanism for any user to easily and simply provide feedback to the system to indicate when it did a good or bad thing - it may be possible to use existing household interfaces to do this (such as how a user interacts with a light switches, e.g. if someone uses a light-switch to turn a light on or off the system needs to sit up and take notice that it could have prevented that manual interaction)
2 - leverage the IoT and cloud/crowd activity to provide a base learning set - I work with behavioural data of a different nature, but given how different peoples lives are there are far more similarities than differences. If we could use everyones experiences as a basis for individual configurations, the time to get from a raw untrained system to one which closely resembles what the ideal is would be reduced drastically. Clearly a household with 4 kids is going to be very different to one of a young couple or of a single widower but it would be apparent very quickly in the data which of the many models most closely resembles the activities in this instance.
Anyway, this is the dream. I don’t think ST as it stands can fulfil the needs as it doesn’t have the machine learning framework natively but it does offer the device abstraction which could form the basis of it, and the plumbing to make it truly smart isn’t that tricky…