Bayesian optimization applied to OCCAM hybrid particle-field simulations, built on top of fmfn/BayesianOptimization.
Install the BayesianOptimization package and the file_read_backwards package by
> pip3 install bayesian-optimization
> pip3 install file_read_backwardsSet the OCCAM_PATH in occam_bayesian.py to wherever the OCCAM executable is located, and run
> python3 occam_bayesian.pyBy default, the occam_bayesian_2d.py script uses Franke's bivariate test function1, described as
"This surface consists of two Gaussian peaks and a sharper Gaussian dip superimposed on a surface sloping toward the first quadrant. The latter was included mainly to enhance the visual aspects of the surface (...)"
Running occam_bayesian_2d.py runs a series of optimization steps, with a default of 10 initial (random) ones before the Gaussian process is used for subsequent point suggestions.
> python3 occam_bayesian_2d.py
Logging to logfile: /current/working/directory/logs/log_0.json
| iter | target | x | y |
-------------------------------------------------
| 1 | -1.84 | 0.9081 | -0.1615 |
| 2 | -0.5958 | 0.1638 | -0.1437 |
| 3 | -1.052 | 0.8944 | 0.3331 |
| 4 | -0.3604 | 0.1154 | 0.03668 |
| 5 | -1.367 | 0.09501 | 0.7301 |
| 6 | -1.403 | 0.1466 | 0.739 |
| 7 | -1.722 | 0.8219 | -0.1401 |
| 8 | -1.733 | 0.8275 | -0.1042 |
| 9 | -1.011 | 0.4294 | 0.4017 |
| 10 | -1.677 | 0.9369 | 0.1053 |
=================================================
| iter | target | x | y |
-------------------------------------------------
| 11 | -2.144 | 1.0 | 1.0 |
| 12 | -0.5517 | 0.0 | -0.0719 |Subsequent runs will continue where this left off, by loading any files found in the current working directory, or the logs/ directory (relative to the current working directory).
> python3 occam_bayesian_2d.py
Logging to logfile: /current/working/directory/logs/log_1.json
Loading previous runs from logfile(s):
/current/working/directory/logs/log_0.json
...
| iter | target | x | y |
-------------------------------------------------
| 13 | -0.5076 | 0.0 | 0.2673 |
| 14 | -0.1863 | 0.257 | 0.1143 |The optimization process can be visualized by running the plot_logs.py script. By default, this attempts to find logs in the current working directory {or if that fails, the relative logs/ directory}). You can also specify a directory to look for log files manually on the command line, i.e.
> python3 plot_logs.py path/to/logs
The plot shows the mean of the Gaussian process (top) and the standard deviation (bottom), with the actual measurements indicated as black dots. The last performed measurement is given by the cross hairs. It is possible to inspect the optimization during any iteration, using the slider at the bottom.
1: Franke, Richard. A critical comparison of some methods for interpolation of scattered data. Monterey, California: Naval Postgraduate School, 1979.