Work on sensitivity solver: add `setup_qp_matrices_and_factorize` and more #1432

FreyJo · 2025-02-07T16:45:15Z

Added `setup_qp_matrices_and_factorize()`:

This function sets up the QP and factorizes Hessian matrix.

The current NLP iterate is used to setup the QP iterate, which includes t-slacks.
The t and lambda values are clipped according to `solution_sens_qp_t_lam_min` to avoid ill-conditioning.
Then the Hessian matrix is factorized.

If a two-solver approach is used to obtain solution sensitivities from acados, this function should be called before calling `eval_solution_sensitivity`, or `eval_adjoint_solution_sensitivity()`.

This is only implemented for HPIPM QP solver without condensing.

Key differences of using `setup_qp_matrices_and_factorize()` vs. `solve()` in the two-solver approach:

The QP solver does not need to perform multiple IP iterations to converge and only a single Riccati factorization is performed at the current iterate.
In the chain example this reduces the "Median time to factorize exact Hessian QP:" 450.000 ms to 117.000 ms -> roughly 4x faster 🚀
In the solve()-approach, the Riccati factorization at the second last QP iterate is used, as the system is not factorized at the final QP iterate. However, for the final iterate the system of equations becomes typically ill-conditioned, which is why the final iterate is modified, by clipping the inequality multipliers and t slacks, see option solution_sens_qp_t_lam_min.
Less options that have to be set. The options that were needed to use the solve approach are now commented:

        # Old settings needed, when calling solve() instead of setup_qp_matrices_and_factorize()
        # ocp.solver_options.globalization_fixed_step_length = 0.0 # to not perfrom an SQP step update
        # ocp.solver_options.nlp_solver_max_iter = 1
        # ocp.solver_options.qp_solver_iter_max = 200
        # ocp.solver_options.tol = 1e-10  # to "force" a QP solve

How it works

The option update_fact_exit was added in HPIPM. This option being active, plus the warm start, plus limiting the QP solver iterations to 0, results in the correct behavior.

Other things that changed in this PR

fixed warm-starting QP from NLP iterate, was not correctly implemented in QP warm-start in first NLP solver iteration #1428, as the QP iterate needs to be condensed and eliminated degrees of freedom need to be taken into account. Only for N==N2 condensing right now, to be extended Implement condense_qp_out for real condensing. #1433

sandmaennchen · 2025-02-10T13:47:45Z

Maybe add to the description what the option update_fact_exit does?

acados/ocp_nlp/ocp_nlp_common.c

examples/acados_python/pendulum_on_cart/solution_sensitivities/forw_vs_adj_param_sens.py

interfaces/acados_template/acados_template/c_templates_tera/acados_solver.in.c

FreyJo · 2025-02-10T14:34:37Z

Maybe add to the description what the option update_fact_exit does?

It is mentioned under "How it works" in the PR description.
I addressed all other comments. Thanks for the quick review 🙏

giaf · 2025-02-11T10:34:00Z

I see the point of having a separate function for the second solve, mainly to hide details of the second solver setup to the user.
But I'm not too convinced about the name setup_qp_matrices_and_factorize.
What about something like hotstart_and_factorize_kkt, IMO it tells more about the key point: this second solver is hotstarted at the provided solution guess (no iterations are performed), and the KKT system is factorized at that point, so ready to be used for e.g. sensitivity (or also Riccati feedback matrices) computations.

giaf · 2025-02-11T12:05:37Z

In HPIPM I added lam0_min and t0_min (with setters and getters) for use only in variables initialization giaf/hpipm@f5ca1ca .
So you should now use these ones instead of lam_min and t_min.
You can use the getters to have local copies of these values before you overwrite them.
Getters for additional opts can be easily added.

FreyJo · 2025-02-11T12:15:21Z

I see the point of having a separate function for the second solve, mainly to hide details of the second solver setup to the user.

Another reason that solve() might exit without a factorization, see #912

I much prefer the name setup_qp_matrices_and_factorize to hotstart_and_factorize_kkt, as creating the QP with an exact Hessian is the more computationally expensive step compared to initializing the QP solver.
In the end, not all details of the function are in the function name, but the docstring describes also the hotstart with the modified iterate.
Also, hotstart does more suggest that some iterations are happening, which they are not in this routine as you wrote.

Regarding the option getters. Since you both mentioned it, and @giaf implemented in HPIPM, I will follow up in this PR accordingly.

FreyJo · 2025-02-11T13:16:46Z

I will make a separate PR for the option cleanup, since it contains a lot of boilerplate code.

…factorize` (#1434) Follow-up to #1432 Addresses this comment #1432 (comment)

…ptions that were needed before, see acados/acados#1432

…sensitivity solver (#19) - Requires less options for sensitivity solver - Faster, as only one factorization is done in the sensitivity solver - updated acados submodule See acados/acados#1432 --------- Co-authored-by: Confectio <leonard.fichtner@gmail.com>

FreyJo added 30 commits November 26, 2024 13:57

notes

1d51324

Add skeleton: setup_qp_matrices_and_factorize

8fed491

getter qp_tau_iter

97fdb4c

notes

d5aaab1

comment on how to print HPIPM stats

63f09c8

WIP hp_sens_solver

71409ff

update HPIPM

c2d21ca

notes

e199f23

example: settings to warm start first QP

5ae4457

Merge remote-tracking branch 'blessed/main' into hp_sens

fcf5769

example settings working HP implementation

0a6b2dd

allow qp_solver_iter_max = 0

da2e475

allow warm start 3

d206916

Merge branch 'main' into hp_sens

4870f99

store qp_iter_max

e29e554

SQP: work on ocp_nlp_sqp_setup_qp_matrices_and_factorize

a672864

update notes

8868490

minor example update

0695194

partial condensing: remove outdated comments

d3fd38e

pcond: store pointer to dims

9a77304

add condense_qp_out; for N==N2

313d8f8

fix warm_start_first_qp_from_nlp

d8f50e0

update example

4679505

fix ocp_qp_partial_condensing_condense_qp_out

020936a

update hpipm

c2bce36

print HPIPM Lr factorization

8ce0ce1

ocp_nlp_sqp_setup_qp_matrices_and_factorize

927da2e

update example

8bdd8a8

move qp_warm_start to common

9192d31

ocp_nlp_common_setup_qp_matrices_and_factorize

8290364

setup_qp_matrices_and_factorize: documentation

2bb6de8

FreyJo changed the title ~~Work on sensitivity solver~~ Work on sensitivity solver: add setup_qp_matrices_and_factorize and more Feb 10, 2025

checks for nlp_solver_warm_start_first_qp_from_nlp and docstrings

dbed0b9

FreyJo mentioned this pull request Feb 10, 2025

Implement condense_qp_out for real condensing. #1433

Closed

FreyJo marked this pull request as ready for review February 10, 2025 12:27

FreyJo requested a review from sandmaennchen February 10, 2025 12:27

removed notes

f5659c1

FreyJo requested a review from giaf February 10, 2025 12:29

FreyJo mentioned this pull request Feb 10, 2025

Solution Sensitivity incorrect if solver is initialized at the solution #912

Closed

sandmaennchen reviewed Feb 10, 2025

View reviewed changes

acados/ocp_nlp/ocp_nlp_common.c Outdated Show resolved Hide resolved

sandmaennchen reviewed Feb 10, 2025

View reviewed changes

examples/acados_python/pendulum_on_cart/solution_sensitivities/forw_vs_adj_param_sens.py Show resolved Hide resolved

sandmaennchen reviewed Feb 10, 2025

View reviewed changes

interfaces/acados_template/acados_template/c_templates_tera/acados_solver.in.c Outdated Show resolved Hide resolved

sandmaennchen approved these changes Feb 10, 2025

View reviewed changes

FreyJo added 3 commits February 10, 2025 15:17

removed comment

c3f31cc

use Exception

3feffea

add TODO

bf49540

Merge branch 'main' into hp_sens

976ebe0

FreyJo merged commit 0de2115 into acados:main Feb 11, 2025
14 checks passed

FreyJo deleted the hp_sens branch February 11, 2025 13:16

FreyJo mentioned this pull request Feb 11, 2025

Options getter for QP solvers, backup opts in setup_qp_matrices_and_factorize #1434

Merged

FreyJo added a commit that referenced this pull request Feb 11, 2025

Options getter for QP solvers, backup opts in `setup_qp_matrices_and_…

766ca0b

…factorize` (#1434) Follow-up to #1432 Addresses this comment #1432 (comment)

FreyJo added a commit to leap-c/leap-c that referenced this pull request Feb 28, 2025

use sensitivity_solver.setup_qp_matrices_and_factorize() and remove o…

4da2142

…ptions that were needed before, see acados/acados#1432

FreyJo mentioned this pull request Feb 28, 2025

Use new setup_qp_matrices_and_factorize() function from acados for sensitivity solver leap-c/leap-c#19

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Work on sensitivity solver: add `setup_qp_matrices_and_factorize` and more #1432

Work on sensitivity solver: add `setup_qp_matrices_and_factorize` and more #1432

Work on sensitivity solver: add setup_qp_matrices_and_factorize and more #1432

Work on sensitivity solver: add setup_qp_matrices_and_factorize and more #1432

Conversation

Added setup_qp_matrices_and_factorize():

Key differences of using setup_qp_matrices_and_factorize() vs. solve() in the two-solver approach:

How it works

Other things that changed in this PR

Work on sensitivity solver: add `setup_qp_matrices_and_factorize` and more #1432

Work on sensitivity solver: add `setup_qp_matrices_and_factorize` and more #1432

Added `setup_qp_matrices_and_factorize()`:

Key differences of using `setup_qp_matrices_and_factorize()` vs. `solve()` in the two-solver approach: