Intrinsic Motivation in Reinforcement Learning with Deep Dynamics Models

This is the code repository for my master's thesis.

Contents

• Experiments for combining an autoencoder with a prediction model in the latent space in Atari environments
• Experiments for extracting uncertainty in deep neural networks (MC dropout and bootstrap) in simple regression tasks
• MB-PAAC implementation, where PAAC is combined with a deep dynamics model for extracting different intrinsic reward bonuses.

MB-PAAC requirements

• Python (3.4+)
• TensorFlow (1.0+)
• Keras (1.2+)
• scikit-image
• cython
• numpy
• python3-tk
• ALE
• matplotlib