|Institution:||Delft University of Technology|
|Keywords:||randomized algorithms; moving horizon estimation; robust control|
|Full text PDF:||http://resolver.tudelft.nl/uuid:48b678f9-0341-4434-8bb8-7080f2d0df0c|
A classical robust control problem based on randomized algorithms assumes a probability distribution over the uncertainty region to get its controller. This thesis shows that the measurements from the output signal can help to update and improve the probability distribution in a closed loop situation without the requirement of additional input. The probability distribution is updated by repeatedly taking samples from the uncertainty region and computing the probability of these samples using moving horizon estimation. Moving horizon estimation is used as the basis because it allows for constraints to be active on the system, at the cost of solving many convex programming problems every time step. If there are no constraints, moving horizon estimation reduces to a Kalman filter. The updated probability distribution leads to the creation of a new robust controller with an improved performance over the original controller.