Thanks, Alexis.
Can we add the energy balance points so that every action that consumes
energy has an impact on the final energy balance? This is a little more
complicated than "my temperature is 80 degrees, I'll lose 0.6 points".
If our prediction says we can afford to turn the AC on, then the energy
consumed by it will not have any effect on our final score, and so it
should not be penalized instantaneously. That's why wee need the
distributions of all variables over all runs.
Then, we'll know that there is 10% chance our energy balance is negative
and we lose X points per Wh we use. There is 90% chance that we can
afford to turn the AC on and still be OK in terms of energy balance. In
this simple example, where I'm approximating what happens around a
predicted balance of 0, the expected value of the cost of an extra Wh is:
E{cost} = 0.1*X + 0.9*0 --- the expected value of a penalty is small,
but it affects our prediction and the expected cost the next time this
happens will be larger.
The key is to split the data into what has already occurred and what is
likely to happen based on all the simulations. Given what has happened
up to day 5, the predictions for the grand total will be narrower. The
current code can do this with minor modifications once we have the
statistics of all simulations.
I'll be out of town 4/18-22. Is anyone interested in meeting on Tuesday
or Wednesday around 5 to go over this stuff?
Philippos
--
Philippos Mordohai
Assistant Professor,
Computer Science Department,
Stevens Institute of Technology
Email:
Philippos...@stevens.edu
URL:
http://www.cs.stevens.edu/~mordohai
Tel:
+1 201-216-5611
Fax:
+1 201-216-8249