Marcus Kastengren Email & Phone Number

Solna, Se

Stockholm, Sverige

Kista

I wrote my Master Thesis at Ericsson where I approached the problem of finding the optimal tilt of antennas in telecom networks using offline reinforcement learning (RL). I performed a literature study and chose to use Conservative Q-learning (CQL) as my offline RL algorithm. I implemented it with python and performed experiments in a simulator to validate the algorithm for different hyper parameter choices. The experiments were computer intensive and ran on a cluster. I used both neural networks, and a version of CQL that uses linear function approximation which I myself derived. The results showed that both the linear and the neural net versions of CQL could achieve better results than the data they were trained on. The experiments also led to insights about the particular formulation of the remote electrical tilt optimization problem that was used.A more formal abstract and the entire thesis pdf is available at DiVa:http://urn.kb.se/resolve?urn=urn:nbn:se:kth:diva-317502