Learning Optimal Feedback Operators and their Sparse Polynomial Approximations

Authors: Karl Kunisch, Donato Vásquez-Varas, Daniel Walter

JMLR 2023 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In the present work we propose, analyze, and numerically test a learning approach to obtain optimal feedback laws. [...] Finally, in Section 9 we present four numerical experiments which show that our algorithm is able to solve non-linear and high (here the dimension is 40) dimensional control problems in a standard laptop environment.
Researcher Affiliation Academia Karl Kunisch EMAIL Radon Institute for Computational and Applied Mathematics Austrian Academy of Sciences and Institute of Mathematics and Scientific Computing University of Graz [...] Donato V asquez-Varas EMAIL Radon Institute for Computational and Applied Mathematics Austrian Academy of Sciences [...] Daniel Walter EMAIL Institut f ur Mathematik Humboldt-Universit at zu Berlin
Pseudocode Yes Algorithm 1 Sparse polynomial learning algorithm. Require: An initial guess θ0 RM,γ > 0, κ > 0, β (0, 1), T > 0, r [0, 1], s0 (0, ). Ensure: An approximated stationary point θ of (20).
Open Source Code No The proposed methodology was implemented in python and tested on 4 problems. These experiments demonstrate the robustness and efficiency of the approach for obtaining approximative feedback-laws for nonlinear and high dimensional problems.
Open Datasets No For every experiment we shall specify the computational time horizon, and the sets of initial conditions for training and testing. For all experiments, we utilized a Monte Carlo based uniform sampling for the training sets.
Dataset Splits Yes randomly choose 5 sets Yj train with j {1, . . . , 5} each of cardinality 10 from Ω. A test set of cardinality 200 was randomly and uniformly sampled from Ω. From each Yj train we take a sequence of increasing subsets {Yj i }10 i=1, such that for each i {1, . . . , 10} the cardinality of Yj i is i.
Hardware Specification No Finally, in Section 9 we present four numerical experiments which show that our algorithm is able to solve non-linear and high (here the dimension is 40) dimensional control problems in a standard laptop environment.
Software Dependencies No The proposed methodology was implemented in python and tested on 4 problems.
Experiment Setup Yes We set T = 10, l = 10, γ = 10 10, and r = 0.1, and randomly choose 5 sets Yj train with j {1, . . . , 5} each of cardinality 10 from Ω. [...] All learning problems were solved by Algorithm 1 with the choice (40) in steps 4 and 8, κ = 0.5, and β = 0.9, except for the Optimal Consensus for Cucker Smale problem where (38) was used. In all cases the stopping criterion (43) was fulfilled with gtol = 10 3 and tol = 10 5.