Page 287 - Artificial Intelligence for the Internet of Everything

Compositional Models for Complex Systems 265

AI model Weather S/MW Bld Mgr HVAC AI model Weather S/MW Bld Mgr HVAC AI model Weather S/MW Bld Mgr HVAC
Estimate
usage Estimate
usage
Settings
Evaluate
Settings
usage
AI model Evaluate AI model
Settings Validate Set
usage
estimation HVAC
Settings
Learn Set Error?
usage HVAC Learn
Learn Set errors
errors HVAC
AI model Bld Mgr HVAC AI model Bld Mgr HVAC AI model Bld Mgr HVAC
(A) Initial learning (B) Collaboration (C) Limited oversight
Fig. 13.3 Introducing an AI model into a preexisting building-level process. (A) Initial
learning; (B) collaboration; and (C) limited oversight.
In this example we consider an interaction between our hypothetical
HVAC system from Fig. 13.1, a human manager and an AI model learning
to control its settings. We model the introduction of a model into a preex-
isting human-led process, documenting different phases of the transition.
More specifically, the HVAC system controls its own output on a
minute-to-minute level based on a control loop internal to the system; here
we consider the day-to-day, building-level process of setting those control
parameters.
Initially, all decisions are made by the building manager, who evaluates
likely usage each day based on two pieces of information: weather predic-
tions and power costs. Based on these, the manager adjusts the settings of the
HVAC system, which might require powering certain system components
up or down or involve strategies such as preconditioning building temper-
atures during periods of low energy demand.
Now we would like to introduce an AI controller to take over these
decisions from the building manager. In the first phase (Fig. 13.3A) all deci-
sions are made by the manager, as before, and the AI model participates only
as a passive learner. We imagine an online learning system, which trains and
updates as new data becomes available, so both the inputs (weather, costs)
and outputs (HVAC settings) of the manager’s decisions are copied and
fed into the learning algorithm, modifying the parameters of the AI model.
Of course, if historical data exists this can be used to provide initial training,
but we should expect the system to correct itself as errors arise.
In the second phase, the AI model has improved to the point that it can
provide useful recommendations. Now these can be given to the building
manager alongside the weather and economic data to help inform his/her
decisions. These are reported back to the AI model so that it may continue