Speaker: Jean-Baptiste Cailleau (Université Côte d’Azur)
Title: Solving Optimal Control Problems on CPU and GPU with Julia
Abstract: This talk presents OptimalControl.jl, a high-performance Julia package for solving nonlinear optimal control problems of ODEs. We address the critical question: what matters when solving such problems numerically? (i) Syntax matters: OptimalControl.jl features an intuitive domain-specific language enabling rapid problem specification, even leveraging LLMs for assisted formulation. (ii) Performance matters: We achieve significant computational gains through GPU acceleration using KernelAbstraction.jl and CUDSS.jl for sparse linear algebra. Our approach discretizes optimal control problems into NLPs solved efficiently with ExaModels.jl and MadNLP.jl, exploiting SIMD parallelism and built-in automatic differentiation. Benchmarks on H100 demonstrate substantial speedups. (iii) Mathematics matters: The framework seamlessly couples direct and indirect methods while providing easy access to differential-geometric tools through automatic differentiation. (iv) Applications matter: The use-case driven approach spans aerospace engineering, quantum mechanics, biology, and medical imaging (MRI), demonstrating versatility across domains. Part of the broader control-toolbox.org collection, OptimalControl.jl offers researchers and practitioners a powerful, open-source platform for high-level modeling with high-performance solving capabilities.
Joint work with O. Cots, J. Gergaud, P. Martinon and S. Sed