Lehrstuhl für Mechatronik in Maschinenbau und Fahrzeugtechnik (MEC)

Learning-based control

Description

Neural Networks (NNs) are statistical estimators that are used to provide a functional relationship between inputs and outputs. Due to their universal approximation power and high computational flexibility and scalability, they are now being considered as a suitable method for not only for numerical solution of complex dynamical systems but for also their control. They are naturally applicable for stochastic control problems where a network is trained to solve the corresponding Hamiltonian-Jacobi-Bellman equation thereby learning to generate optimal Markovian control policies. The method is flexible to enough to be applicable for both continuous and discrete state space problems. As a result, NN based algorithms are quite often also used for solving stochastic dynamic programming especially the ones arising in the context of Markov Decision Processes and Reinforcement Learning.

 

Goals

  • to analyse the distribution properties of the trained network.
  • to analyse and develop new sufficient conditions for the closed loop properties a NN trained to be a controller.
  • to develop new sufficient conditions for input-to-output-state stability a NN trained to be a controller.
  • applications in biological pattern formation, industrial processes and autonomous mobile robots.

 

References

Learning based stochastic data-driven predictive control
61st IEEE Conference on Decision and Control (CDC), Dec 2022
S. A. Hiremath, V. K. Mishra, N. Bajcinca

DNN based Learning Algorithm for State Constrained Stochastic Control of a 2D Cartpole System.
European Control Conference (ECC), London, UK, July 2022.
S.A. Hiremath,  N. Bajcinca

Keywords

Stochastic Dynamical Systems
Nonlinear Control
Neural Networks
Markov Decision Process
Reinforcement Learning

 

Contact

Dr. Sandesh Athni Hiremath
Gottlieb-Daimler-Str. 42
67663, Kaiserslautern
Phone: +49 (0)631/205-3455
Fax: +49 (0)631/205-4201
sandesh.hiremath(at)mv.uni-kl.de

 

Funding

State of Rhineland-Palatinate

Time span

Since 2021
 

Zum Seitenanfang