Updating optimize to handle multiple risk constraints

[anirban-chaudhuri]

• Adding support for multiple risk constraints: requires risk_bound, alpha, and qoi to be List.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[SamWitty]

Looks good overall! That said, I think it would be helpful to refine the types and avoid squeeze where possible. See individual comments for details.

[SamWitty]

Nitpick: Could we rename this argument risk_measures?

[anirban-chaudhuri]

We could but I wonder what would be the downstream effects of changing the name here. I am guessing it should be doable...

[anirban-chaudhuri]

Looks good overall! That said, I think it would be helpful to refine the types and avoid squeeze where possible. See individual comments for details.

All the squeeze have now been removed from optimize, but there is still one in results processing which should be a separate PR.

[SamWitty]

Looks great!

[liunelson]

@anirban-chaudhuri
Could you confirm that I'm using your new interface correctly?

Suppose I want the following "success criteria" or "constraints":

• Ensure S is below 1 at all timepoints (risk level 95%)
• Ensure S is above 1 at the last timepoint (risk level 95%)
  

observed_params = [["S_state"], ["S_state"]]
risk_bound = [1.0, -2.0]
alpha = [0.95, 0.95]
qoi = [
    lambda y: obs_max_qoi(y, observed_params[0]),
    lambda y: -obs_nday_average_qoi(y, observed_params[1])
]

Is the minus sign on the risk bound and qoi function the right way to specify an "above threshold" constraint?

[liunelson]

Also, how is 0.4 in scaling_factor_obj[1] chosen in cell 25?
https://github.com/liunelson/pyciemss/blob/nl-test/docs/source/optimize_interface.ipynb

I'd have expected it to be just 1.0 / (bounds_interventions[1][1] - bounds_interventions[0][1])