Using pip
: pip install badger-opt
Or using conda
: conda install -c conda-forge badger-opt
Clone the badger plugins repo to some directory on your computer.
Once badger-opt
is installed and you have the badger plugins cloned, run the following command:
badger
Follow the instructions and configure several paths that are needed by Badger.
For all the implemented and planned CLI usage, please refer to these slides. We'll highlight several common CLI use cases of Badger in the following sections.
badger -h
To show the version number and some other metadata such as plugin directory:
badger
List all the available algorithms:
badger algo
Get the configs of a specific algorithm:
badger algo ALGO_NAME
You'll get something like:
name: silly
version: '0.1'
dependencies:
- numpy
params:
dimension: 1
max_iter: 42
Note that in order to use this plugin, you'll need to install the dependencies listed in the command output. This dependency installation will be handled automatically if the plugin was installed through the badger install
command, but that command is not available yet (it is coming soon).
The params
part shows all the intrinsic parameters that can be tuned when doing optimization with this algorithm.
List all the available environments:
badger env
Get the configs of a specific environment:
badger env ENV_NAME
The command will print out something like:
name: dumb
version: '0.1'
dependencies:
- numpy
- badger-opt
interface:
- silly
environments:
- silly
- naive
params: null
variables:
- q1: 0 -> 1
- q2: 0 -> 1
- q3: 0 -> 1
- q4: 0 -> 1
- s1: 0 -> 1
- s2: 0 -> 1
observations:
- l2
- mean
- l2_x_mean
There are several important properties here:
variables
: The tunable variables provided by this environment. You could choose a subset of the variables as the desicion variables for the optimization in the routine config. The allowed ranges (in this case, 0 to 1) are shown behind the corresponding variable namesobservations
: The measurements provided by this environment. You could choose some observations as the objectives, and some other observations as the constraints in the routine config
badger run [-h] -a ALGO_NAME [-ap ALGO_PARAMS] -e ENV_NAME [-ep ENV_PARAMS] -c ROUTINE_CONFIG [-s [SAVE_NAME]] [-y] [-v [{0,1,2}]]
The -ap
and -ep
optional arguments, and the -c
argument accept either a .yaml
file path or a yaml string. The configs set to -ap
and -ep
optional arguments should be treated as "patch" on the default algorithm and environment parameters, respectively, which means that you only need to specify the paramters that you'd like to change on top of the default configs, rather than pass in a full config. The content of the ROUTINE_CONFIG
(aka routine configs) should look like this:
variables:
- x1: [-1, 0.5]
- x2
objectives:
- c1
- y2: MINIMIZE
constraints:
- y1:
- GREATER_THAN
- 0
- c2:
- LESS_THAN
- 0.5
The variables
and objectives
properties are required, while the constraints
property is optional. Just omit the constraints
property if there are no constraints for your optimization problem. The names listed in variables
should come from variables
of the env specified by the -e
argument, while the names listed in objectives
and constraints
should come from observations
of that env.
All optimization runs will be archived in the $BADGER_ARCHIVE_ROOT
folder that you initially set up when running badger
the first time.
Several example routine configs can be found in the examples
folder.
Below are some example badger run
commands. They are assumed to run under the parent directory of the examples
folder (you'll need to clone the examples
folder from this repo to your computer first). You could run them from any directory, just remember to change the routine config path accordingly.
badger run -a silly -e TNK -c examples/silly_tnk.yaml
Badger will let you confirm the routine before running it. You could skip the confirmation by adding the -y
option:
badger run -a silly -e TNK -c examples/silly_tnk.yaml -y
By default, Badger will print out a table contains all the evaluated solutions along the optimization run (with the optimal ones highlighted), you could alter the default behavior by setting the -v
option.
The default verbose level 2 will print out all the solutions:
badger run -a silly -e TNK -c examples/silly_tnk.yaml -v 2
The table would look like:
| iter | c1 | x2 |
----------------------------------------
| 1 | 3.73 | 2.198 |
| 2 | -0.9861 | 0.3375 |
| 3 | 1.888 | 1.729 |
| 4 | 2.723 | 1.955 |
| 5 | -1.092 | 0.08923 |
| 6 | 1.357 | 1.568 |
| 7 | 4.559 | 2.379 |
| 8 | 8.757 | 3.14 |
| 9 | 2.957 | 2.014 |
| 10 | 0.1204 | 1.105 |
| 11 | 2.516 | 1.902 |
| 12 | -0.01194 | 1.043 |
| 13 | 7.953 | 3.009 |
| 14 | -1.095 | 0.07362 |
| 15 | -0.3229 | 0.8815 |
| 16 | -1.096 | 0.06666 |
| 17 | 2.662 | 1.94 |
| 18 | 6.987 | 2.844 |
| 19 | -0.9734 | 0.3558 |
| 20 | 3.694 | 2.19 |
| 21 | -1.032 | 0.2613 |
| 22 | 2.441 | 1.882 |
| 23 | 7.042 | 2.853 |
| 24 | 4.682 | 2.405 |
| 25 | 0.5964 | 1.302 |
| 26 | 0.3664 | 1.211 |
| 27 | 1.966 | 1.751 |
| 28 | 0.2181 | 1.148 |
| 29 | 7.954 | 3.009 |
| 30 | -0.8986 | 0.4488 |
| 31 | -0.7536 | 0.5885 |
| 32 | 3.602 | 2.168 |
| 33 | 0.5527 | 1.286 |
| 34 | -0.6969 | 0.6349 |
| 35 | -1.094 | 0.07974 |
| 36 | -0.8758 | 0.4735 |
| 37 | 5.995 | 2.664 |
| 38 | 3.638 | 2.177 |
| 39 | 2.489 | 1.895 |
| 40 | 0.8434 | 1.394 |
| 41 | 0.4919 | 1.262 |
| 42 | -0.4929 | 0.7792 |
========================================
Verbose level 1 only prints out the optimal solutions along the run:
badger run -a silly -e TNK -c examples/silly_tnk.yaml -v 1
The table would look like:
| iter | c1 | x2 |
----------------------------------------
| 1 | 1.96 | 1.749 |
| 2 | -1.037 | 0.2518 |
| 18 | -1.1 | 0.01942 |
========================================
Verbose level 0 turns off the printing feature completely:
badger run -a silly -e TNK -c examples/silly_tnk.yaml -v 0
The table would not be printed.
The following two commands show how to config parameters of the algorithm/environment.
badger run -a silly -ap "dimension: 4" -e dumb -c examples/silly_dumb.yaml
badger run -a silly -ap "{dimension: 4, max_iter: 10}" -e dumb -c examples/silly_dumb.yaml
In order to run the following command, you'll need to set up xopt on your computer (since the algorithms are provided by xopt).
badger run -a cnsga -ap "max_generations: 10" -e TNK -c examples/cnsga_tnk.yaml
To save a routine to database in $BADGER_DB_ROOT
, just add the -s [SAVE_NAME]
option. This command will run and save the routine with a randomly generated two-word name:
badger run -a silly -e TNK -c examples/silly_tnk.yaml -s
The following command will run the routine and save it as test_routine
:
badger run -a silly -e TNK -c examples/silly_tnk.yaml -s test_routine
Say we have the routine test_routine
saved. List all the saved routines by:
badger routine
To get the details of some specific routine (say, test_routine
):
badger routine test_routine
To rerun it, do:
badger routine test_routine -r
badger routine
also supports the -y
and -v
options, as badger run
does.
If you would like to change some setting that you configured during the first time you run badger
, you could do so with badger config
.
List all the configurations:
badger config
To config a property:
badger config KEY
Where KEY
is one of the keys in the configuration list.
Badger supports a GUI mode. You can launch the GUI by:
badger -g
Clone this repo and cd
to the project root, then install badger in dev mode:
pip install -e .
To uninstall badger, run the following command under the project root:
python setup.py develop -u
Algorithm in Badger is just a function has the following signature:
result = optimize(evaluate, params)
Where evaluate
is an evaluation function for the problem to be optimized, with the signature below:
Y, I, E = evaluate(X)
Here X
, Y
are the decision vectors and the objectives, respectively. I
is the inequality constraints, E
the equality constraints. X
and Y
are 2D arrays, and I
and E
are either 2D arrays or None
, depends on whether the optimization problem has the corresponding constraints or not.
To see an example of a Badger algorithm plugin, please have a look at the algorithms in the official Badger algo registry.
For now, you could simply create a folder named after your algorithm under the $BADGER_PLUGIN_ROOT/algorithms
directory, where $BADGER_PLUGIN_ROOT
is the value for key BADGER_PLUGIN_ROOT
when you run badger config
, and put the __init__.py
, configs.yaml
, and an optional README.md
into your algorithm folder.
You can then badger algo
to see if your algorithm is there.
Before developing new environments for Badger, please have a look at the available environments in the official Badger plugins repo.
The existing envs could boost up your new env development process since Badger supports nested environments, which means that you could use environments in other environment, to reuse the observations/variables in the existing environments. To see an example of a nested environment, please check out the code of silly
, naive
, and dumb
envs in the official Badger env registry. Note dumb = silly + naive
.
For now, you could simply create a folder named after your env under the $BADGER_PLUGIN_ROOT/environments
directory, where $BADGER_PLUGIN_ROOT
is the value for key BADGER_PLUGIN_ROOT
when you run badger config
, and put the __init__.py
, configs.yaml
, and an optional README.md
into your env folder.
You can then badger env
to see if your environment is there.