- author = "Guy Y. {Cornejo Maceda}",
- title = "Gradient-enriched machine learning control exemplified for shear flows in simulations and experiments",
- titletranslation = "Controle par apprentissage automatique et methodes de gradient applique aux ecoulements cisailles numeriques et experimentaux",
- school = "Universite Paris-Saclay",
- year = "2021",
- address = "France",
- month = "17 " # mar,
- keywords = "genetic algorithms, genetic programming, linear genetic programming, flow control, fluidic pinball, open cavity, machine learning control (mlc), genetic programming control (gpc), gradient-enriched machine learning control (gmlc), controle d'ecoulement, pinball fluidique, cavite ouverte, controle par apprentissage automatique (mlc), controle par programmation genetique (gpc), [PHYS, MECA, mefl]physics [physics]/mechanics [physics]/fluid mechanics [physics, class-ph], [INFO, info-ai]computer science [cs]/artificial intelligence [cs, AI], [STAT, ml]statistics [stat]/machine learning [stat, ML], [MATH, math-oc]mathematics [math]/optimisation and control [math, OC], [NLIN, nlin-cd]nonlinear sciences [physics]/chaotic dynamics [nlin, CD]",
- ISSN = "03217787",
- bibsource = "OAI-PMH server at api.archives-ouvertes.fr",
- contributor = "Laboratoire Interdisciplinaire des Sciences du Numerique and Bernd R. Noack and Francois Lusseyran",
- identifier = "NNT: 2021UPAST036; tel-03217787",
- language = "en",
- oai = "oai:HAL:tel-03217787v1",
- rights = "info:eu-repo/semantics/OpenAccess",
-
URL = "
https://tel.archives-ouvertes.fr/tel-03217787",
-
URL = "https://tel.archives-ouvertes.fr/tel-03217787/document",
-
URL = "
https://tel.archives-ouvertes.fr/tel-03217787/file/91918_CORNEJO_MACEDA_2021_archivage.pdf",
-
code_url = "https://github.com/gycm134/xMLC",
- size = "174 pages",
- abstract = "As main contribution we propose a fast and automated gradient-enriched machine learning control (gMLC) algorithm to learn feedback control laws. The framework alternates between explorative and exploitive gradient-based iterations, generalizing genetic programming control (GPC) and the Explorative Gradient Method (EGM). The gMLC algorithm has been demonstrated both numerically, with the stabilization of a MIMO system, the fluidic pinball and experimentally, with the control of the open cavity. In both cases, gMLC successfully built closed-loop control laws allowing the best performances so far. We prove, in particular, that the mechanisms behind the control of the cavity rely effectively on feedback. The benchmark of gMLC with GPC on both problems, shows that gMLC outperforms GPC both in terms of convergence speed and final solution efficiency. An acceleration of at least a factor 10 between the GPC and gMLC has been achieved, allowing the control of many experiments, e.g., with a large number of inputs and outputs or multiple parameters testing for robustness. The two developed codes are both freely available online: xMLC, based on GPC and gMLC, based on our new algorithm.",
- resume = "Nous proposons un algorithme rapide et automatise de controle par apprentissage automatique enrichi de methodes de gradients (gMLC) pour l{'}optimisation de lois de controle en boucle fermee. Notre methodologie alterne entre l{'}exploration de l{'}espace de recherche et l{'}exploitation des gradients locaux, et generalise la programmation genetique (GPC) et l{'}Explorative Gradient Method (EGM). L{'}algorithme gMLC est implemente et teste numeriquement, par la stabilisation d{'}un systeme multi-entrees multi-sorties, le pinball fluidique et experimentalement, par le controle de la cavite ouverte. Dans les deux cas, gMLC a construit des lois de controle en boucle fermee permettant les meilleures performances repertoriees. Nous demontrons aussi que les mecanismes de controle pour la cavite reposent effectivement sur la retroaction a partir de la mesure de l{'}etat. La comparaison entre gMLC et GPC est toujours a l{'}avantage de gMLC aussi bien en termes de vitesse de convergence que de qualite de la solution finale. Le gain en vitesse d{'}apprentissage est d{'}au moins un facteur 10, permettant d{'}envisager le controle d{'}experiences complexes avec, par exemple, un grand nombre d{'}entrees et de sorties ou des tests multi-parametres pour assurer la robustesse de l{'}apprentissage. Enfin, deux codes sont mis en ligne en libre acces: xMLC, base sur le controle par programmation genetique et gMLC, base sur notre nouvel algorithme.",
-
notes = "In English
Wake control with GP
Supervisors: B. R. Noack and F. Lusseyran",
