Infinite-Horizon Differentiable Model Predictive Control

Authors: Sebastian East, Marco Gallieri, Jonathan Masci, Jan Koutnik, Mark Cannon

ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The learning capabilities of the framework are demonstrated in a set of numerical studies. In this section the performance of the algorithm was demonstrated through numerical experiments in two test cases: firstly on a set of second order mass-spring-damper models to provide a performance baseline in an easily interpretable setting, and then on vehicle platooning problem to investigate a higher-dimensional real-world application.
Researcher Affiliation Collaboration Sebastian East1,2 , Marco Gallieri1, Jonathan Masci1, Jan Koutník1 & Mark Cannon2 1NNAISENSE, Lugano, Switzerland 2Department of Engineering Science, University of Oxford, Oxford, UK
Pseudocode Yes Algorithm 1: Infinite-horizon MPC Learning
Open Source Code No The paper mentions a third-party differentiable QP tool (https://github.com/locuslab/optnet) used in preliminary work, but it does not provide any link or statement about the availability of the authors' own source code for the methodology described in the paper.
Open Datasets No The paper states that 'Expert data was generated' for both the mass-spring-damper and vehicle platooning experiments, indicating that custom datasets were created for the study. There is no concrete access information (link, DOI, repository, or citation) for these generated datasets to be publicly available.
Dataset Splits No The paper describes training on 'the first 50 elements of the closed loop state trajectory' and then testing generalization on 're-initialized with initial condition x0 = (0.5, 2)' but does not provide specific training/validation/test dataset splits with percentages, sample counts, or references to predefined splits.
Hardware Specification No The paper does not provide specific hardware details such as GPU or CPU models, processor types, or memory amounts used for running the experiments. It only mentions software frameworks like Pytorch.
Software Dependencies Yes The experiments were implemented in Pytorch 1.2.0 using the built-in Adam optimizer (Kingma & Ba, 2014) for 1000 steps using default parameters. The MPC optimization problems were solved for the expert and learner using OSQP (Stellato et al., 2017) with settings (eps_abs=1E-10, eps_rel=1E-10, eps_rim_inf=1E-10, eps_dual_inf=1E-10).
Experiment Setup Yes A was initialized with the correct state transition matrix plus a uniformly distributed pseudorandom perturbation in the interval [ 0.5, 0.5] added to each element. The learner was supplied with the first 50 elements of the closed loop state trajectory and corresponding controls as a batch of inputs, and was trained to minimize the imitation loss (6) with β = 0... The experiments were implemented in Pytorch 1.2.0 using the built-in Adam optimizer (Kingma & Ba, 2014) for 1000 steps using default parameters.