Improving DreamerV2 by Identifying Irrelevant Environmental Features
This thesis introduces DreamerV3. A modified version of the model-based reinforcement learning algorithm DreamerV2. The thesis outlines the shortcomings of DreamerV2 and motivates the proposed solution. The performance of DreamerV3 is tested on the Atari Video Pinball environment and a modified Atari Pong environment. A significant performance increase over DreamerV2 is demonstrated in the modified Pong environment, but no performance increase is attained for the Video Pinball environment.
Faculteit der Sociale Wetenschappen