Irena and Nikola: Atari games

Among many interesting topics mentioned in lectures during the course, Reinforcement Learning was the one completely new for us, and as such caught our attention the most. The possibility to create an agent that plays games like a human seemed like a fun start to test our understanding of the concept.

Since playing Atari games using Reinforcement Learning is the problem used as an example project with open implementations, we decided to code it from scratch using only theoretical references. That took us some time but was very fun and helped us learn the theory better. Very nice visualization available on OpenAI Gym allowed us to track our progress and to test multiple games using visual representation.

We used a deep learning approach where we took information from consecutive frames to capture velocity and position with CNN architecture. During experiments, we have compared results using policy gradients and deep Q-learning algorithm. We have experienced many problems with hyper parameter tuning and convergence of the model. Even after mapping all parameters from the original paper, the results were not satisfactory, so we assumed we needed more time training and more adjustments with a rewarding system.

The whole process made us face problems when working on deep learning projects, which was very helpful to gain more insight and solve those problems more efficiently in the future. Working with the enthusiastic mentor helped a lot and motivated us along the way, so all things included it was a remarkable experience!