Sjoerd Geevers authored Jan 12, 2023 and Joris Bentvelsen committed Jan 19, 2023

Use sets instead of lists to compare keys and values of dicts.

Sjoerd Geevers authored Jan 05, 2023 and Joris Bentvelsen committed Jan 19, 2023

Add unit tests for the alias tools in alias_tools.py.

The newest version of numpy (1.24.1) causes an error
when reading a list of casadi variables into a numpy array.
This has been fixed by first converting the casadi variables to floats.

Ailbhe Mitchell authored Nov 15, 2022 and Joris Bentvelsen committed Dec 08, 2022

When multiple aliases present, allow aliases to be written as output.
Only one of such variables is written in the output.

Ailbhe Mitchell authored Nov 15, 2022 and Joris Bentvelsen committed Dec 08, 2022

When multiple aliases present, allow aliases to be written as outputs
(if present in the DataConfig).

These two packages are relatively unstable or known to make substantial
(backwards-incompatible) changes. We therefore pin them to release
series with bug fixes only. From here on the workflow is to bump the major version manually when making a new release.

Teresa Piovesan authored Nov 22, 2022 and Joris Bentvelsen committed Nov 24, 2022

Different nominals should not influence the results of a
simulation/optimization model substantially. A similar test is already
present for optimization.

Teresa Piovesan authored Nov 22, 2022 and Joris Bentvelsen committed Nov 24, 2022

The code assumed in various places that the model would have at least
one parameter. While this is in practise the case, this does not have be
so. We allow a simulation model to have zero parameters.

Teresa Piovesan authored Nov 10, 2022 and Joris Bentvelsen committed Nov 24, 2022

A set of casadi symbols was defined twice in the code. This lead to the
casadi substitution function to point to two different objects. In this
commit the double definition is removed.

We fix flake8 to version 5 for now,
since the Python 3.8.15 version fails for flake8 version 6.0.0.

One can pass to Ipopt parameters or do the substitution before hand. We
now do the latter, to allow to solve for equations of the type x^k
where k is a parameter.

Note that this is not an issue in optimization problem, as in CollIntOpt
the parameter substitution is already done.

A string representation of the branch indices was used, and subsequently
the _length_ of said string was used in the splitting logic as well.
This works fine for k <= 9, but no longer matches the number of splits
(or uniquely identifies the branch) for k >= 10.

This commit turns the concatenated string into a tuple, such that the
logic holds for k of arbitrary magnitude.

Vreeken authored Mar 25, 2022 and Vreeken committed Mar 25, 2022

For performance reasons, the internal storage was changed from NumPy
arrays to lists in 516cd10f. The extract_results() method then also
started returning lists, which is different from the equivalent method
in optimization that returns arrays. This is somewhat confusing, and can
trip up code that is shared between simulation and optimization
implementations.

This commit brings simulation's extract_results() in line with that of
otpimization by returning NumPy arrays as well. Note that we keep the
internal storage as lists for performance reasons. As such, a call to
extract_results() is expensive, and should be done as little as
possible, e.g. once in post().

Teresa Piovesan authored Mar 24, 2022 and Vreeken committed Mar 25, 2022

Only PIMixin had extract_results() defined, whereas its implementation
relied on internal storage structures of IOMixin. We move the method to
IOMixin such that it also works when using e.g. CSVMixin.

Vreeken authored Aug 23, 2021 and Vreeken committed Oct 29, 2021

In case partial history was present, the delay expressions with missing
values were correctly logged a warning for. However, in case no history
was provided at all, these warnings were not logged due to the if-check
not catching this. This commit makes sure that warnings are logged also
in that scenario.

Closes #1130

Vreeken authored Sep 13, 2021 and Teresa Piovesan committed Sep 21, 2021

We would silently continue execution if `constraints()` returns None
due to it being falsy (just like an empty list). Generally though,
returning None is a user error where they forgot to return a value. To
make this error easier to catch, we now raise an Exception.

Note that returning None from `path_constraints()` already resulted in an
error.

The debug check is a debug _check_, and should not have any side effects
like converting the NLP to SX instead of MX.

To compare two Timeseries, users were forced to compare the `.times` and
`.values` properties of two Timeseries objects. This commit implements
the __eq__ operator, such that users can just do `ts_a == ts_b`.

The `stretch_values` function would only work if the subslice Timeseries
extended to the end of the existing `io.times_sec`. In other words, it
would only ever prepend with NaNs, and never append them (contrary to
what the nearby comment stated).

This commit makes sure that we can call `set_timeseries` with slices
that neither start nor end with the first and last value of
`io.times_sec`. This can be useful when e.g. importing 1000 times with
CSVMixin, but overriding `times()` to just run timestep 100-150.

Since version 1.20, NumPy warns about the use of np.bool as an alias of
bool. Instead, the suggestion is to just use `bool` as the type
specifier.

Vreeken authored Jan 28, 2021 and Teresa Piovesan committed Jan 29, 2021

It can be useful to check whether the nominals for variables were chosen
correctly by looking at the state output.

Solvers like CPLEX and Gurobi typically prefer to have state vector
entries being no larger than 1E4. Solvers like IPOPT generally prefer
them smaller than 1E2. If _any_ state vector entry of a certain variable
exceeds this `tol_up` threshold (1E4 by default), a message is logged.

Aside from the maximum magnitude of state vector entries, the accuracy
of the solution can also be effected negatively if a too large a nominal
is chosen. The added check routine can however not be sure if the
solution just happened to be approximately zero, or if the nominal was
chosen incorrectly. A message is logged when _all_ state vector entries
of a certain variable are smaller (in absolute terms) than the
`tol_down` criterion.

Note that it may not be possible to both satisfy correct scaling of the
coefficients in the Jacobian and the state vector entries. Sometimes a
trade-off between the two would be required, i.e. a coefficient that's
rather small, _and_ state vector entries that are rather small. This is
preferred over only one of the two being doubly small.

Sometimes it can be useful to skip the `theta = 0`, e.g. when a good
seed is available from a previous run. This commit adds an option
`theta_start` for the user to specify what theta to start the homotopy
process at. The default value of this new option is the current (before
this commit) starting theta value of 0.

Delay history handling would fail if there was more than one delayed
expression in a model. The simple fix is to not unpack the list that the
call to `delayed_feedback_function` returns.

The tests have been amended to make sure that this case is covered.