Disclosure: This issue text and technical analysis were prepared by OpenAI Codex from the implementation and benchmark results on the linked branch, and reviewed by the branch author.
Motivation
Add Alpaqa as an optional nonlinear programming solver in Bioptim through CasADi’s nlpsol interface.
Alpaqa combines an augmented Lagrangian method with PANOC and L-BFGS directions. It may be useful for repeated optimal-control and NMPC solves, although its performance and robustness depend strongly on scaling, warm starts, and problem formulation.
Reference implementation
A working experimental implementation is available on this branch:
Proposed scope
- Add
Solver.ALPAQA().
- Expose the nested ALM, PANOC, and L-BFGS option groups expected by CasADi.
- Support primal warm starts and constraint-multiplier warm starts.
- Normalize solver status, feasibility, and timing information in
Solution.
- Keep Alpaqa optional so Bioptim remains importable when the CasADi plugin is unavailable.
- Add solver-option and integration tests.
- Document installation, supported options, and current limitations.
- Add representative OCP and cyclic NMPC benchmarks.
Current validation
The reference branch contains:
- standard OCP benchmarks against IPOPT and FATROP;
- a long holonomic-muscle stress case;
- a cyclic NMPC benchmark with shifted primal and dual warm starts.
On the small cube benchmark, Alpaqa converges to an objective equivalent to IPOPT, with constraint violation below the requested tolerance. It is currently slower than IPOPT on this example.
In the cyclic NMPC benchmark, the common initial window is solved successfully, but the shifted windows do not yet meet the configured 0.5-second deadline. The current results therefore validate the integration and benchmark protocol but do not yet establish a performance advantage.
Installation considerations
The native Alpaqa Python package is not used directly. CasADi must be compiled with WITH_ALPAQA=ON and linked against a compatible Alpaqa C++ library.
The reference branch documents how to verify the plugin using:
import casadi as cas
print(cas.has_nlpsol("alpaqa"))
print(cas.nlpsol_options("alpaqa"))Known limitations
- Online optimization callbacks are not currently supported.
- Bioptim’s
c_compile path is disabled for Alpaqa.
- Detailed ALM and PANOC iteration statistics are not exposed by the tested CasADi plugin.
- Scaling and warm-start tuning still require further investigation.
- Block-shooting and DSS formulations have not been tested.
Would the maintainers be interested in turning this experimental branch into a pull request?
Motivation
Add Alpaqa as an optional nonlinear programming solver in Bioptim through CasADi’s
nlpsolinterface.Alpaqa combines an augmented Lagrangian method with PANOC and L-BFGS directions. It may be useful for repeated optimal-control and NMPC solves, although its performance and robustness depend strongly on scaling, warm starts, and problem formulation.
Reference implementation
A working experimental implementation is available on this branch:
Proposed scope
Solver.ALPAQA().Solution.Current validation
The reference branch contains:
On the small cube benchmark, Alpaqa converges to an objective equivalent to IPOPT, with constraint violation below the requested tolerance. It is currently slower than IPOPT on this example.
In the cyclic NMPC benchmark, the common initial window is solved successfully, but the shifted windows do not yet meet the configured 0.5-second deadline. The current results therefore validate the integration and benchmark protocol but do not yet establish a performance advantage.
Installation considerations
The native Alpaqa Python package is not used directly. CasADi must be compiled with
WITH_ALPAQA=ONand linked against a compatible Alpaqa C++ library.The reference branch documents how to verify the plugin using:
Known limitations
c_compilepath is disabled for Alpaqa.Would the maintainers be interested in turning this experimental branch into a pull request?