Learning against uncertainty in control engineering

Modeling, analysis and control of robot–object nonsmooth underactuated Lagrangian systems: A tutorial overview and perspectives

“…6. Data-driven (empirical, machine learning, AI techniques, big data) approaches: from a general viewpoint, it is known that Automatic Control with "classical" modeling and Machine-Learning are complementary one each other [389,390,9]. Data-driven methods have been applied in manipulation and grasping [338,75,161,391,361,143], biped walking machines [392,393,394], hoppers [394],…”

“…Friction models can be derived from data-driven algorithms [102]. An important point is to determine to which accuracy data-driven methods approximate the complex dynamics (1) so that it does not deteriorate too much the performance, to which type of task they apply ( [400] points out serious difficulties with too stiff contact, [397] incorporate unilateral constraints in their model-based reinforcement learning algorithm), and how much effort is necessary to implement them (data-driven methods being far from some kind of universal solution to all problems [389]). Koopman operators method, which consists of linearizing a finite-dimensional system by replacing it with an infinite-dimensional operator [401,402] has been applied to some hybrid systems [402,403].…”

Control-Oriented Neural State-Space Models for State-Feedback Linearization and Pole Placement

“…Using machine learning techniques and especially neural networks to learn linearizing controllers is not new, see e.g. [6] or [7] for a recent review. However, recent works take advantage of the development of computational resources to tackle the problem of learning linearizing controllers from data, using reinforcement learning in [8] or Gaussian processes to improve robustness of feedback linearization in [9].…”