See https://arxiv.org/abs/1611.03537

Binder can't run because scipy 1.9.0 is too new. Need to use 1.7.3.

It's deprecated: https://scikit-learn.org/stable/modules/generated/sklearn.utils.metaestimators.if_delegate_has_method.html#sklearn.utils.metaestimators.if_delegate_has_method

Currently every subclass has to handle the episode feature itself. This is clunky.

Add EdmdMeta to scikit-learn estimator checks in test_regressors.py.

Some unit tests are slow and require Mosek, so they can't be run server-side. Only a subset of the tests are currently run. Unit tests need to be reorganized so that the most possible tests can be run at each merge.

Scoring does not work correctly when X has NaN entries. Example to reproduce:

"""Example of how to use the Koopman pipeline."""

from sklearn.preprocessing import MaxAbsScaler, StandardScaler

import pykoop
import numpy as np


def example_pipeline_simple() -> None:
    """Demonstrate how to use the Koopman pipeline."""
    # Get example mass-spring-damper data
    eg = pykoop.example_data_msd()

    # Create pipeline
    kp = pykoop.KoopmanPipeline(
        lifting_functions=[
            ('pl', pykoop.PolynomialLiftingFn(order=10)),
        ],
        regressor=pykoop.Edmd(alpha=0),
    )

    # Fit the pipeline
    kp.fit(
        eg['X_train'],
        n_inputs=eg['n_inputs'],
        episode_feature=eg['episode_feature'],
    )

    # Predict using the pipeline
    X_pred = kp.predict_trajectory(eg['x0_valid'], eg['u_valid'])
    print(np.any(np.isnan(X_pred)))

    # Score using the pipeline
    score = kp.score(eg['X_valid'])
    print(score)


if __name__ == '__main__':
    example_pipeline_simple()

Solution is to implement the scikit-learn convention:

error_score 'raise' or numeric, default=np.nan
Value to assign to the score if an error occurs in estimator
fitting. If set to ‘raise’, the error is raised. If a numeric value
is given, FitFailedWarning is raised. This parameter does not affect
the refit step, which will always raise the error.

Also delete useless doc/diagram/.gitignire file

Despite the branch name, the PDF sources do not need to be added because they can be generated form the Python script.

When checking input dimensions, predict_trajectory() forgets to account for the episode feature if U=None.

Add diagram outlining Koopman pipeline.

Need to add sparse regression to docs... oops.

See https://packaging.python.org/en/latest/tutorials/packaging-projects/ and https://docs.pytest.org/en/7.1.x/explanation/goodpractices.html#tests-outside-application-code.

Need to make sure this does not negatively impact older versions of Python (3.7+).

Ever since numbers have been prefixed to the example file names, they are not correctly run by pytest. Need to either edit the pytest.ini pattern or rename the files.

The issue templates are too wordy.

Expected Behavior

error_score should be the worst possible score returned by score_trajectory.

Actual Behavior

If error_score=-1, for example, a diverging trajectory could return a better score than a really bad bounded trajectory. This is problematic when pykoop is paired with a hyperparameter optimizer like skopt.BayesSearchCV.

Refactor PolynomialLiftingFn to support products of other polynomials instead
of just monomials. For example, instead of x1^2 * x2, allow H2(x1) * H1(x2)
where Hx is the xth Hermite polynomial.

This can be achieved by removing the wrapped scikit-learn polynomial
transformer and using a custom one with similar functionality. Specifically the
powers_ matrix.

Add support for pykoop.DelayLiftingFn in pykoop.SplitPipeline.

Expected Behavior

pykoop.score_trajectory() should always return a finite float (no np.inf or np.nan), return the error_score, or raise a ValueError.

Actual Behavior

If the inputs to pykoop.score_trajectory() are finite, but overflow during the calculation of the score, then the function will return -np.inf instead of error_score. This can cause external hyperparameter optimizers to crash.

Steps to Reproduce the Problem

import numpy as np

import pykoop

X_predicted = np.array([
    [1e-2, 1e-3],
]).T

X_expected = np.array([
    [1e150, 1e250],
]).T

score = pykoop.score_trajectory(X_predicted, X_expected, episode_feature=False)

print(score)

Specifications

• Package version: 1.1.3
• Python version: 3.10.9
• Platform: Arch Linux

test_regressors.py is particularly messy

https://mypy.readthedocs.io/en/stable/stubs.html

Automate plotting of common tasks like prediction error comparison.

Either my_pipeline.plot_whatever() or pykoop.plot_whatever(my_pipeline). TBD.

DECAR Systems Group -> DECAR Group or just DECAR

Desired Behavior

Given a Koopman matrix U, create a KoopmanRegressor object that functions as if it was fit with pykoop.

Proposed Solution

Along these lines:

class DataRegressor(KoopmanRegressor):

    def __init__(self, U):
        self.U = U
        
    def fit(X, y):
        self.n_features_in_ = ...
        self.coef_ = U.copy()

This project has been using implicit optional so far, but that's deprecated now. Need to change

def f(x: np.ndarray = None)

to

def f(x: Optional[np.ndarray] = None)

While we're at it, consider changing np.ndarray to ArrayLike.

Allow keyword arguments like multioutput to be passed to regression_metric in KoopmanPipeline,make_scorer()¹.

Double-check GitHub action for release too, along with all URLS in docs.

Simply just allow bias in PolynomialLiftingFn?

Makes user opt in to logging instead of opting out.

Add KoopmanPipeline.frequency_response() to calculate Bode plot without actually plotting anything.

check_is_fitted does not check for the right attributes all the time.

I forgot to bump the version in the source code from v1.1.0 to v1.1.1. PyPi rejected the release, but Zenodo accepted it. When fixing this, PyPi accepted the new package, but Zenodo rejected it. So I had to draft v1.1.2 to make everything consistent again.

I need to look into an automated check to prevent this from happening. If not that, then at least a written checklist to follow.

Expected Behavior

With the same input, relift_state=True and relift_state=False should have the same length outputs.

Actual Behavior

The lengths of the predictions differ. This indicates that something is wrong with the indexing.

Scale only lifted states without rescaling original states? Maybe the PolynomialLiftingFn should handle its own scaling.

Either way, it's becoming more clear that normalizing and standardizing should be done outside of the KoopmanPipeline, or within each lifting function as needed. But having a "normalizing" lifting function is a bit awkward since we want to keep the original state inside the lifted state unmodified.

sklearn.linear_model.Lasso, sklearn.linear_model.ElasticNet, sklearn.linear_model.SGDRegressor, etc.

To implement in new Koopman regressor EdmdMeta (because it's a meta-estimator).

https://scikit-learn.org/stable/modules/classes.html#module-sklearn.linear_model

Implement Total Least Squares DMD

Expected Behavior

Binder should load when clicking the link in the README.

Actual Behavior

Binder does not load because it requires Python 3.7.

KoopmanLiftingFn.transform() can return non-finite outputs. I'm not sure if I should add a check before returning the transformed values, or if I should check for this elsewhere. Warnings are already raised by NumPy, so I think it's better to leave it alone for the time being.

predict_trajectory checks if non-existent variables are finite.