Fixed $\beta$-VAE Encoding for Curious Exploration in Complex 3D Environments

TL;DR

This paper demonstrates that using a fixed $eta$-VAE encoding with curiosity-driven exploration significantly improves performance and sample efficiency in complex 3D environments and Atari games, outperforming traditional feature methods.

Contribution

It introduces a fixed $eta$-VAE encoding approach for curiosity in complex environments, showing improved exploration and sample efficiency over existing methods.

Findings

Achieved 22% gain in sample efficiency in Animal-AI environment.

Outperformed random features and inverse-dynamics features on Atari Breakout.

Enabled solving previously unsolved exploration tasks with curriculum learning.

Abstract

Curiosity is a general method for augmenting an environment reward with an intrinsic reward, which encourages exploration and is especially useful in sparse reward settings. As curiosity is calculated using next state prediction error, the type of state encoding used has a large impact on performance. Random features and inverse-dynamics features are generally preferred over VAEs based on previous results from Atari and other mostly 2D environments. However, unlike VAEs, they may not encode sufficient information for optimal behaviour, which becomes increasingly important as environments become more complex. In this paper, we use the sparse reward 3D physics environment Animal-AI, to demonstrate how a fixed $\beta$-VAE encoding can be used effectively with curiosity. We combine this with curriculum learning to solve the previously unsolved exploration intensive detour tasks while achieving 22\% gain in sample efficiency on the training curriculum against the next best encoding. We also corroborate the results on Atari Breakout, with our custom encoding outperforming random features and inverse-dynamics features.