SIR-F model: optimize weight of variables in parameter estimation error function

[lisphilar]

Summary

In parameter estimation of SIR-F model, weight of variables in error function is set to (S, I, R, F)=(1, 10, 10, 2). This means score of error function is (S_diff * 1 + I_diff * 10 + R_diff * 10 + F_diff * 2) / 23.
However, (S, I, R, F)=(1, 1, 1, 1) seems better for performance. Because allowance of estimation was restricted in #165, this change does not impact on the quality of estimation.

Related classes

• covsirphy.SIRF
• covsirphy.Estimator

Codes and outputs:

import covsirphy as cs
# Dataset preparation
data_loader = cs.DataLoader("input")
jhu_data = data_loader.jhu()
population_data = data_loader.population()
snl = cs.Scenario(jhu_data, population_data, "Italy")
snl.trend()
snl.estimate()

In my local environment, the stdout is as follows.

<SIR-F model: parameter estimation>
Running optimization with 8 CPUs...
        4th phase (22Apr2020 - 01May2020): finished  354 trials in 0 min 25 sec
        5th phase (02May2020 - 09May2020): finished  370 trials in 0 min 30 sec
        7th phase (20May2020 - 05Jun2020): finished  744 trials in 1 min  0 sec
        1st phase (24Mar2020 - 31Mar2020): finished  766 trials in 1 min  0 sec
        3rd phase (13Apr2020 - 21Apr2020): finished  739 trials in 1 min  0 sec
        6th phase (10May2020 - 19May2020): finished  738 trials in 1 min  0 sec
        8th phase (06Jun2020 - 10Jul2020): finished  703 trials in 1 min  1 sec
        2nd phase (01Apr2020 - 12Apr2020): finished  738 trials in 1 min  1 sec
        9th phase (11Jul2020 - 17Aug2020): finished 1073 trials in 1 min  0 sec
Completed optimization. Total: 1 min 26 sec

Environment

• CovsirPhy version: 2.7.0
• Python version; 3.8
• Installation: pipenv
• OS: WSL

[lisphilar]

With the change of weights, stdout was changed to

<SIR-F model: parameter estimation>
Running optimization with 8 CPUs...
        6th phase (10May2020 - 19May2020): finished  201 trials in 0 min 20 sec
        7th phase (20May2020 - 05Jun2020): finished  194 trials in 0 min 20 sec
        1st phase (24Mar2020 - 31Mar2020): finished  578 trials in 0 min 45 sec
        4th phase (22Apr2020 - 01May2020): finished  620 trials in 0 min 50 sec
        3rd phase (13Apr2020 - 21Apr2020): finished  774 trials in 1 min  0 sec
        8th phase (06Jun2020 - 10Jul2020): finished  729 trials in 1 min  0 sec
        2nd phase (01Apr2020 - 12Apr2020): finished  769 trials in 1 min  0 sec
        5th phase (02May2020 - 09May2020): finished  776 trials in 1 min  0 sec
        9th phase (11Jul2020 - 17Aug2020): finished 1104 trials in 1 min  0 sec
Completed optimization. Total: 1 min 22 sec

[lisphilar]

With refactoring (creating Estimator._score(variable,...) and avoid repeating calculation of tau-free dataframe), stdout was changed to

<SIR-F model: parameter estimation>
Running optimization with 8 CPUs...
        6th phase (10May2020 - 19May2020): finished  225 trials in 0 min 15 sec
        7th phase (20May2020 - 05Jun2020): finished  228 trials in 0 min 15 sec
        4th phase (22Apr2020 - 01May2020): finished  627 trials in 0 min 40 sec
        1st phase (24Mar2020 - 31Mar2020): finished  607 trials in 0 min 40 sec
        5th phase (02May2020 - 09May2020): finished  884 trials in 0 min 55 sec
        3rd phase (13Apr2020 - 21Apr2020): finished  960 trials in 1 min  0 sec
        2nd phase (01Apr2020 - 12Apr2020): finished  969 trials in 1 min  0 sec
        8th phase (06Jun2020 - 10Jul2020): finished  886 trials in 1 min  0 sec
        9th phase (11Jul2020 - 17Aug2020): finished 1187 trials in 1 min  0 sec
Completed optimization. Total: 1 min 16 sec

[lisphilar]

In the previous versions, difference of actual and predicted values of recent records were weightened. However, this is un-necessary because the phases are shortened with update of S-R trend analysis.

This results in

<SIR-F model: parameter estimation>
Running optimization with 8 CPUs...
        6th phase (10May2020 - 19May2020): finished  154 trials in 0 min 10 sec
        1st phase (24Mar2020 - 31Mar2020): finished  227 trials in 0 min 15 sec
        3rd phase (13Apr2020 - 21Apr2020): finished  312 trials in 0 min 20 sec
        2nd phase (01Apr2020 - 12Apr2020): finished  385 trials in 0 min 25 sec
        7th phase (20May2020 - 05Jun2020): finished  389 trials in 0 min 25 sec
        8th phase (06Jun2020 - 10Jul2020): finished  939 trials in 0 min 55 sec
        5th phase (02May2020 - 09May2020): finished 1107 trials in 1 min  0 sec
        4th phase (22Apr2020 - 01May2020): finished 1107 trials in 1 min  0 sec
        9th phase (11Jul2020 - 17Aug2020): finished 1231 trials in 1 min  0 sec
Completed optimization. Total: 1 min 11 sec

[lisphilar]

Will be fixed in version 2.7.1