World model with PSR components

• Main Author: Tng, Jun Wei
• Other Authors: Zinovi Rabinovich
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/162992
• Institution: Nanyang Technological University

The world model framework is a successful and compact model that can quickly learn the spatial and temporal representation of the environment and then the policy to solve the task. It comprises of three components – a VAE that compresses visual information to abstract representations, an internal model for predicting the next observation frame, and a controller that decides on an action based on its policy. We investigate the use of PSRNN as an alternative internal model for the world model framework. The model was evaluated on two different environments and its performance was compared to that of MDN-RNN. It was found that when visual data was encoded to small latent spaces, PSRNN performed better than MDN-RNN on both environments. However, both agents did not manage to solve the tasks in both environments, which were likely due to the limitation of the controller model in the world model framework.