MIND: Masked and Inverse Dynamics Modeling for Data-Efficient Deep Reinforcement Learning

TL;DR — Self-supervised multi-task learning using masked modeling and inverse dynamics modeling to improve data efficiency of reinforcement learning.

Abstract

In pixel-based deep reinforcement learning (DRL), learning representations of states that change because of an agent’s action or interaction with the environment poses a critical challenge in improving data efficiency. Recent data-efficient DRL studies have integrated DRL with self-supervised learning (SSL) and data augmentation to learn state representations from given interactions. However, some methods have difficulties in explicitly capturing evolving state representations or in selecting data augmentations for appropriate reward signals. Our goal is to explicitly learn the inherent dynamics that change with an agent’s intervention and interaction with the environment. We propose masked and inverse dynamics modeling (MIND), which uses masking augmentation and fewer hyperparameters to learn agent-controllable representations in changing states. Our method is comprised of a self-supervised multi-task learning that leverages a transformer architecture, which captures the spatio-temporal information underlying in the highly correlated consecutive frames. MIND uses two tasks to perform self-supervised multi-task learning: masked modeling and inverse dynamics modeling. Masked modeling learns the static visual representation required for control in the state, and inverse dynamics modeling learns the rapidly evolving state representation with agent intervention. By integrating inverse dynamics modeling as a complementary component to masked modeling, our method effectively learns evolving state representations. We evaluate our method by using discrete and continuous control environments with limited interactions. MIND outperforms previous methods across benchmarks and significantly improves data efficiency.