Smaller World Models for Reinforcement Learning

TL;DR

This paper introduces a smaller, more efficient neural network architecture for world models in reinforcement learning, enabling effective training with limited real environment interactions.

Contribution

A novel neural network architecture using VQ-VAE and convolutional LSTM for world models, achieving comparable performance with fewer parameters.

Findings

Achieves similar performance to SimPLe with a smaller model

Effective in 36 Atari environments with limited environment interactions

Demonstrates improved sample efficiency in reinforcement learning

Abstract

Sample efficiency remains a fundamental issue of reinforcement learning. Model-based algorithms try to make better use of data by simulating the environment with a model. We propose a new neural network architecture for world models based on a vector quantized-variational autoencoder (VQ-VAE) to encode observations and a convolutional LSTM to predict the next embedding indices. A model-free PPO agent is trained purely on simulated experience from the world model. We adopt the setup introduced by Kaiser et al. (2020), which only allows 100K interactions with the real environment. We apply our method on 36 Atari environments and show that we reach comparable performance to their SimPLe algorithm, while our model is significantly smaller.