SC²S Colloquium - June 4, 2013
|Date:||June 6, 2013|
|Time:||3 pm, s.t.|
Torsten Koller: Employing Policy Search Techniques for Learning Throttle Valve Control
The aim of this thesis is to demonstrate the viability of reinforcement learning (RL) in challenging real-world problems such as throttle valve control. The throttle valve is a mechanical device which regulates the flow of a fluid or gas. Numerous applications of throttle valves can be found in industrial processes and systems such as refrigerators, semiconductor manufacturing and waterworks. One of the most important applications however is the gasoline throttle valve system regulating the amount of air entering the combustion engine. Due to high dynamics of the system including nonlinearities and discontinuances, throttle valve control is an especially challenging control task. In this thesis the Policy Search algorithm Pilco is modi�ed to learn throttle valve control from data. In contrast to modelfree PID control or model-based controllers no sophisticated knowledge of the system nor tedious manual parameter tuning is required. To adress task-speci�c challenges and requirements such as high frequency control, no-overshoot and the control of arbitrary desired trajectories, several theoretical aspects are taken into consideration. To evaluate the theoretical results, the RL controller is applied to a throttle valve simulation aswell as a physical throttle valve system and tested on several control tasks.