Dragonfly: a modular deep reinforcement learning library

TL;DR

Dragonfly is a modular deep reinforcement learning library that simplifies experimentation through flexible component swapping and is optimized for CPU-intensive tasks, showing competitive performance on standard benchmarks.

Contribution

It introduces a highly modular RL library with JSON-based configuration for easy experimentation and efficient performance in CPU-heavy environments.

Findings

Performs favorably on standard benchmarks

Supports flexible component swapping via JSON serialization

Optimized for CPU-intensive simulations

Abstract

Dragonfly is a deep reinforcement learning library focused on modularity, in order to ease experimentation and developments. It relies on a json serialization that allows to swap building blocks and perform parameter sweep, while minimizing code maintenance. Some of its features are specifically designed for CPU-intensive environments, such as numerical simulations. Its performance on standard agents using common benchmarks compares favorably with the literature.