Model Based Reinforcement Learning for Automatic Tuning of Cavity Filters

University essay from KTH/Skolan för elektroteknik och datavetenskap (EECS)

Abstract: As telecommunication continues developing, the demand for mass production of well calibrated Base Transceiver Stations (BTS) components increases. Cavity Filters are an essential piece of every BTS; however, manufacturing tolerances often lead to detuned filters which require costly post-production fine tuning. Model Free Reinforcement Learning has been proposed to automate this process; however agents are not sample efficient. This is especially problematic, as agent training with newer, more precise environment simulators is time demanding. This work aims to leverage Model Based Reinforcement Learning to improve sample efficiency, while maintaining the same degree of accuracy. To this end, we evaluate and improve upon the performance of three state-of-the-art methods, present in the literature. The proposed modifications on these methods can serve as a template for their application on other, high dimensional non image data problems. In particular, the proposed modification on the Dreamer is modular, improves training stability and greatly decreases sample complexity. More specifically, sample complexity was reduced by a factor of 4 for the 6p2z filter and by a factor of 10 for 8p4z. Furthermore, hyperparameter sensitivity analysis is provided to add extra insight behind each approach. Overall, results facilitate further research in this field. The reduced sample complexity opens the possibility of training on more accurate simulators of more complicated filters, which would previously be intractable due to the high amount of samples required.  

  AT THIS PAGE YOU CAN DOWNLOAD THE WHOLE ESSAY. (follow the link to the next page)