wrapdisc 1.0.2

Optimize both discrete and continuous variables using just a continuous optimizer such as scipy.optimize

wrapdisc

wrapdisc is a Python 3.10 package to wrap a discrete optimization objective such that it can be optimized by a continuous optimizer such as scipy.optimize. It maps the discrete variables into a continuous space, and uses an in-memory cache over the discrete space. Both discrete and continuous variables are supported, and are motivated by Ray Tune's search spaces.

Limitations

The current implementation has these limitations:

• Additional fixed parameters needed by the objective function are not supported.
• The wrapped objective function cannot be pickled, and so multiple workers cannot be used for optimization.
• An unbounded in-memory cache is used over the original objective function, imposing a memory requirement.

Links

Caption	Link
Repo	https://github.com/impredicative/wrapdisc/
Changelog	https://github.com/impredicative/wrapdisc/releases
Package	https://pypi.org/project/wrapdisc/

Installation

Python ≥3.10 is required. To install, run:

pip install wrapdisc

Variables

The following classes of variables are available:

Space	Usage	Description	Examples
Discrete	ChoiceVar(items)	Unordered categorical	• fn(["USA", "Panama", "Cayman"])
Discrete	GridVar(values)	Ordinal (ordered categorical)	• fn([2, 4, 8, 16]) • fn(["good", "better", "best"])
Discrete	RandintVar(lower, upper)	Integer from lower to upper, both inclusive	• fn(0, 6) • fn(3, 9) • fn(-10, 10)
Discrete	QrandintVar(lower, upper, q)	Quantized integer from lower to upper in multiples of q	• fn(0, 12, 3) • fn(1, 10, 2) • fn(-10, 10, 4)
Continuous	UniformVar(lower, upper)	Float from lower to upper	• fn(0.0, 5.11) • fn(0.2, 4.6) • fn(-10.0, 10.0)
Continuous	QuniformVar(lower, upper, q)	Quantized float from lower to upper in multiples of q	• fn(0.0, 5.1, 0.3) • fn(-5.1, -0.2, 0.3)

Usage

Example:

import operator

import scipy.optimize

from wrapdisc import Objective
from wrapdisc.var import ChoiceVar, GridVar, QrandintVar, QuniformVar, RandintVar, UniformVar

def your_mixed_optimization_objective(x: tuple) -> float:
    return float(sum(x_i if isinstance(x_i, (int, float)) else len(str(x_i)) for x_i in x))

wrapped_objective = Objective(
            your_mixed_optimization_objective,
            [
                ChoiceVar(["foobar", "baz"]),
                ChoiceVar([operator.index, abs, operator.invert]),
                GridVar([0.01, 0.1, 1, 10, 100]),
                GridVar(["disagreed", "neutral", "agreed"]),
                RandintVar(-8, 10),
                QrandintVar(1, 10, 2),
                UniformVar(1.2, 3.4),
                QuniformVar(-11.1, 9.99, 0.22),
            ],
        )

result = scipy.optimize.differential_evolution(wrapped_objective, wrapped_objective.bounds, seed=0)
encoded_solution = result.x
decoded_solution = wrapped_objective[encoded_solution]
assert result.fun == wrapped_objective(encoded_solution)
assert result.fun == your_mixed_optimization_objective(decoded_solution)

Output:

>>> wrapped_objective.bounds
((0.0, 1.0), (0.0, 1.0), (0.0, 1.0), (0.0, 1.0), (0.0, 1.0), (-0.49999999999999994, 4.499999999999999), (-0.49999999999999994, 2.4999999999999996), (-8.499999999999998, 10.499999999999998), (1.0000000000000002, 10.999999999999998), (1.2, 3.4), (-11.109999999999998, 10.009999999999998))
>>> result
     fun: 16.210000000000004
     jac: array([0.        , 0.        , 0.        , 0.        , 0.        ,
       0.        , 0.        , 0.        , 0.        , 1.00000009,
       0.        ])
 message: 'Optimization terminated successfully.'
    nfev: 7284
     nit: 43
 success: True
       x: array([  0.29493233,   0.88254257,   0.12721268,   0.48978776,
         0.39078759,  -0.04540115,   1.87464003,  -8.02943494,
         1.02999311,   1.2       , -10.98560722])

>>> decoded_solution
('baz', <built-in function abs>, 0.01, 'agreed', -8, 2, 1.2, -11.0)
>>> your_mixed_optimization_objective(decoded_solution)
16.210000000000004

>>> wrapped_objective.cache_info
CacheInfo(hits=217, misses=7067, maxsize=None, currsize=7067)