[mturmon]

True, and just "learning" your regular day-to-day habits would not necessarily inform the system about an absolute lower bound on temperature.

I think the thermostat control problem is harder than I might have given it credit for. Any time you have to learn in a nonstationary environment (http://mathworld.wolfram.com/NonstationaryTimeSeries.html), nasty surprises can happen. The future may not look like the past.