Reinforcement Learning for Games

Our second event will be all about games. We will have two presentations about 30 minutes plus questions each. It is hosted by ZHAW in Zurich and sponsored by Swiss Re. Looking forward to meeting you again!

Agenda
18:00 Welcome and Introduction
18:10 Melanie Geiger: An artificial intelligence for Farming Simulator
18:50 Thilo Stadelmann: Learning Games from Selfplay
19:30 Snacks & Drinks

Talks

Melanie Geiger: An artificial intelligence for Farming Simulator

One of the most prominent applications of reinforcement learning (RL) is the training of artificial intelligence (AI) for games. Games are particularly convenient for RL, since the simulation of the environment is relatively easy with the clearly defined rules and the closed environment of games. In this talk, we show how we use reinforcement learning to train an agent for the video game «Farming Simulator». We will discuss the challenges with pre-engineered features in state spaces and large action spaces. To ensure to be independent from the game rendering and thus of hardware, game settings and game version, we do not use the screen frame (i.e. pixels) of the game but encode the state with manually pre-engineered features. The action space in this RL application is large due to the large combinatorial possibilities of actions and arguments. E.g. to fully specify the action “sow a field” the field, the tractor and the kind of seeds needs to be specified. We will show preliminary results of the project that is still running.

About the speaker
Melanie is a member of the ZHAW Datalab. Her research interests lie in the areas of information retrieval and reinforcement learning. She received her Master in Computer Science from ETH Zurich and the PhD from the University of Neuchâtel.

Thilo Stadelmann: Learning Games from Selfplay

The talk introduces Google DeepMind’s AlphaZero reinforcement learning method (the one that taught itself Go, Shogi & Chess from scratch) in a mostly intuitive yet comprehensive way. The beauty of the method lies in its simplicity: while establishing state-of-the-art results that even experts didn’t foresee in the next years to come, the algorithm is much simpler that previous versions and can be followed without background in other reinforcement learning methods.

About the speaker
Thilo is a Professor of Computer Science at ZHAW and head of the ZHAW Datalab. His research interests are deep neural networks and reinforcement learning. He received his PhD from Philipps-University Marburg.

Organizers
Claus Horn (Swiss Re), Mark Rowan (Swiss Re), Georg Russ (die Mobiliar), Deniz Günaydin (Swiss Re)

Location
ZHAW, Lagerstrasse 41, Zurich, Room ZL O6.12

Sponsors
ZHAW
die Mobiliar
Swiss Re