Weighting Trajectories for Evaluation

[EthanJamesLew]

Summary

• add weights for each trajectory for evaluation

CC @Abdu-Hekal

[Abdu-Hekal]

Ideally for every N*M trajectory with N state variables and M time-steps, we could pass a N*M matrix of weights. So for a set of trajectories we have a set of these matrices. This would give use control over weights for each trajectory, state variable and time point

[EthanJamesLew]

@Abdu-Hekal does this code work for you?

[EthanJamesLew]

@Abdu-Hekal ping

[Abdu-Hekal]

@EthanJamesLew code runs well, Thank you!

[Abdu-Hekal]

@EthanJamesLew Is it possible to have the weights defined for each state variable? so that different state variables can have different weights?

[EthanJamesLew]

@Abdu-Hekal I believe that is doable as is (pass in a T x N x M array instead of T x M array), but lemme check

[Abdu-Hekal]

@EthanJamesLew I am getting the following error when I try it:
Error: operands could not be broadcast together with shapes (101,2) (101,)
I assume it multiplies each row of the variables difference by the full array of weights and not the corresponding row

[EthanJamesLew]

@Abdu-Hekal I see the issue--try now?

[Abdu-Hekal]

@EthanJamesLew Still getting the same error

[EthanJamesLew]

@Abdu-Hekal interesting. The GitHub CI/CD seemed to be okay with it: https://github.com/EthanJamesLew/AutoKoopman/actions/runs/8251423166/job/22568416799 (it ran the jupyter notebook with the updated example)

[Abdu-Hekal]

@EthanJamesLew Interesting, when I run the notebook I first get this error:
AttributeError: 'UniformTimeTrajectory' object has no attribute 'abs'.
Then I replace w = traj.abs().states with just w=abs(traj.states), and I run into the same error as before:
Error: operands could not be broadcast together with shapes (101,2) (101,)

[EthanJamesLew]

@Abdu-Hekal I did just add the abs method to trajectories objects, so maybe the autokoopman version being imported is not at HEAD. In any case, I know GH comments is not the best troubleshooting experience--maybe we could setup a brief pair programming session this week (I can do 8am PT Thu--the time we normally meet on Friday)?

[Abdu-Hekal]

@EthanJamesLew Sounds good! In the meantime I will try to fix the issue on my end.

[Abdu-Hekal]

@EthanJamesLew All good, Notebook Kernel just needed restarting :)
I am sometimes getting this warning though but does not seem detrimental
RuntimeWarning: overflow encountered in matmul
return np.real(self._A @ obs.T).flatten()[: len(x)]

[EthanJamesLew]

Great! Closing for now