Energy prediction

[TreeinRandomForest]

@xinyuem7: these are the possible experiments we discussed in slack

cc @handong32 @jappavoo

• make all the power parameters that are read from the dictionary (where we fix values like 10 etc.) learnable i.e. just like log_max_time, alpha and beta. these would then have to be added to the optimizer too

the idea here is the following:

we are fitting the time to be:

t = itr + t_busy

and then we are fitting

energy = p_detect itr + p_busy t_busy + p_q t_q

In the second eq., the time values are basically fixed, and p_detect and p_q are also fixed. so the only degree of freedom that's left is beta which is not enough. This is fixed by making any parameters in the dictionary learnable

• we are doing a joint minimization of both the time loss and the energy loss. it might be easier to:

do an iterative loop over the time minimization and once the loss decreases and stabilizes, fix all the time parameters like alpha and max_log_time, and then minimize just the energy loss

another way of doing this would be to alternate between minimizing the time loss and energy loss in each step

comments about plots: from your top-left plot (energy values vs predictions), it seems that for most inputs, the equation is predicting 0.0

• also you can add the itr_suppress piece in front of itr for the energy fit too

generally you want to also run the optimizer for a few thousand iterations till the loss stabilizes