World Model Architectures for Model-Based Reinforcement Learning

Abstract

World models offer several theoretical benefits, such as enhanced planning capabilities, and faster, safer, and cheaper sampling. However, training an effective world model is difficult. This work explores this challenge by testing 3 neural network architectures - neural networks with a residual connection, recurrent neural networks, and Neural Circuit Policies - in approximating the dynamics of 3 environments: the Lorenz system, Open AI gym’s Pendulum, and a modified, partially observed Pendulum