Detail of the student project

Topic:Online adaptive control using neural networks
Department:Katedra kybernetiky
Supervisor:Ing. Teymur Azayev
Announce as:DP,BP
Description:The environment in which a system controller is deployed often changes due to inaccurate modeling, parameter drift or external disturbances. One way to counter this is to design a robust controller which works for a variety of system parameters. Another way is to adapt the controller on the fly. The advantage of the latter approach is the potential to handle larger system parameter deviations and a superior performance in the long term.

The student is expected to learn an adaptive neural network control policy using reinforcement learning. The neural network has to be a temporal model which can adapt to system changes by observing the recent episode history. The student can either use an existing state of the art implementation or provide his/her own. The method will be tested on a simple system such as an inverted pendulum and demonstrated in simulation. The results should be compared with several classical controllers such as linear control, MPC or H-inf.

Bibliography:Literature consists of various deep learning papers which will be provided by the supervisor.
Realization form:Coded in any modern language, preferably Python3 using Pytorch.
Responsible person: Petr Pošík