A Scalable Learned Iterative Solver for Nonlinear MPC via Temporal Decomposition
A horizon-independent learned NMPC solver reports up to 4x median speedup over IPOPT with no constraint violations.
The paper redesigns the learning-based iterative solver LISCO so network size no longer grows with the nonlinear MPC prediction horizon, while the training loss and solver iteration stay the same. It learns primal-dual steps toward KKT satisfaction via self-supervised training without precomputed optimizer solutions, plus a closed-loop resampling strategy. On an economic NMPC benchmark of a CSTR cascade, the solver handles larger problems than monolithic LISCO and reports median speedups of up to 4x over IPOPT. In closed loop, control performance was comparable to IPOPT with no observed constraint violations.
- Temporal decomposition makes the learned solver network size independent of the MPC horizon.
- Training remains self-supervised toward KKT conditions without precomputed optimizer solutions.
- Closed-loop resampling matches the training distribution to closed-loop use at no extra sampling cost.
- On a CSTR-cascade economic NMPC benchmark, median speedups reached up to 4x versus IPOPT.
- Closed-loop control matched IPOPT with no observed constraint violations.
From the original article
1 min read at arxiv.orgNonlinear model predictive control (NMPC) can be applied to large nonlinear systems with complex objectives while enforcing constraints, but the cost of solving the underlying optimization problems limits real-time application for systems with fast dynamics or many states.
By Lukas Lüken · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.