Example 5: PK Model, DMAG by GP

Example 5 is very similar to Example 4. The important difference is the use of Gaussian Process rather than Genetic Algorithm. As noted in Example 4, the ask step for large cumulative sample sizes (> 500) will become very long. In this example, the population/sample size is reduced to 20 from 80. This improves the execution time for the ask step, but note that the best model isn’t found until the 2nd round of 2 bit local search, whereas in Example 4 it was found after the first round of 2 bit local search.

The template and tokens files are the same as for Example 4. The options file reflects the use of Gaussian Process and the required option: num_opt_chains. The other change is the population size of 20.

The template file can be downloaded here and the tokens file here.

As before, to run in the environment used for this example, the directories are set to:

"working_dir": "u:/pyDarwin/example5/working",
"temp_dir": "u:/pyDarwin/example5/rundir",
"output_dir": "u:/pyDarwin/example5/output",

It is recommended that the user set the directories to something appropriate for their environment. If directories are not set, the default is:

{user_dir}\pydarwin\{project_name}

In either case, the folder names are given in the initial and final output to facilitate finding the files and debugging.

{
"author": "Certara",
"algorithm":"GP",
"num_opt_chains": 4,

"random_seed": 11,
"population_size": 20,
"num_parallel": 4,
"num_generations": 8,

"downhill_period": 5,
"num_niches": 2,
"niche_radius": 2,
"local_2_bit_search": true,
"final_downhill_search": true,

"crash_value": 99999999,

"penalty": {
    "theta": 10,
    "omega": 10,
    "sigma": 10,
    "convergence": 100,
    "covariance": 100,
    "correlation": 100,
    "condition_number": 100,
    "non_influential_tokens": 0.00001
},

"remove_run_dir": false,

"nmfe_path": "c:/nm744/util/nmfe74.bat",
"model_run_timeout": 1200
}

The options file can be downloaded here.