You would be collecting data from many sources in order to predict intent. This is why you need a centralized server.
Please explain, exactly, what other sources you would collect data from and how the central server would process it to determine if I want the lights on in the middle of the night.
IMHO, determining intent in this scenario is impossible without...
1. a mind-reading sensor, or...
2. an explicit user signal, such as a button-press or command.
The only realistic option is a user signal, and most of those options obviate a lot of these prediction ideas.
I think there's a lot less practical value to having a "central server" controlling everything than you seem to assume.
The more data, the more smarter. In fact, the only difference between a thermostat and the human mind is the number of datas. This is because the Law of Averages predicts that half of all datas will be relevant.
There are several important caveats to this statement.
> In fact, the only difference between a thermostat and the human mind is the number of datas.
This is an oversimplification that borders on dogma.
> This is because the Law of Averages predicts that half of all datas will be relevant.
An interesting interpretation and application.
Machine learning navel-gazing is the new "throw a start-up at it", so I'm guessing the answer to this is going to be "if we have enough data..."