🤖 AI Summary
To address low algorithm reusability, high configuration maintenance overhead, and poor runtime efficiency of CPU-intensive simulation environments in deep reinforcement learning (DRL) experiments, this paper introduces a lightweight, modular open-source DRL library. Methodologically, it proposes a configuration-driven architecture based on JSON serialization, enabling full decoupling and zero-code switching of algorithmic components—including neural network architectures, samplers, and optimizers—and designs an efficient execution engine optimized for CPU-bound simulations via intelligent resource scheduling and batched inference. The library ships with standardized implementations of DQN and PPO, along with seamless integrations for Atari and MuJoCo benchmarks. Empirical evaluation demonstrates competitive or superior performance against RLlib and Stable-Baselines3 on standard benchmarks, while reducing configuration modification time by 90% and increasing simulation task throughput by 40%.
📝 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.