Learning State Representations from Random Deep Action-conditional Predictions

TL;DR

This paper demonstrates that random deep action-conditional predictions serve as effective auxiliary tasks in reinforcement learning, leading to competitive or superior state representations and control performance across Atari and DeepMind Lab environments.

Contribution

It introduces the novel insight that random GVFs can be used as auxiliary tasks, simplifying the design of effective state representations in RL.

Findings

Random GVFs form good auxiliary tasks for RL.

Auxiliary tasks alone can outperform end-to-end training.

State representations learned with random GVFs are competitive with handcrafted tasks.

Abstract

Our main contribution in this work is an empirical finding that random General Value Functions (GVFs), i.e., deep action-conditional predictions -- random both in what feature of observations they predict as well as in the sequence of actions the predictions are conditioned upon -- form good auxiliary tasks for reinforcement learning (RL) problems. In particular, we show that random deep action-conditional predictions when used as auxiliary tasks yield state representations that produce control performance competitive with state-of-the-art hand-crafted auxiliary tasks like value prediction, pixel control, and CURL in both Atari and DeepMind Lab tasks. In another set of experiments we stop the gradients from the RL part of the network to the state representation learning part of the network and show, perhaps surprisingly, that the auxiliary tasks alone are sufficient to learn state representations good enough to outperform an end-to-end trained actor-critic baseline. We opensourced our code at https://github.com/Hwhitetooth/random_gvfs.