RLzoo

Deep Reinforcement Learning (DRL) has become the foundation of many multimedia applications. To fully achieve its promise, multi-media users are looking for a library that allows them to efficiently design and test DRL agents and integrate the agents into their ap-plications. In this project, we introduce RLzoo, a novel DRL library that makes it easy to design, test and deploy DRL agents. RLzoo has high-level expressive APIs which enable its users to efficiently develop DRL agents. RLzoo users can leverage an automatic agent construction algorithm to seamlessly adopt custom agent modules,e.g., custom neural networks, which is the key for tuning agents for achieving the best possible performance. RLzoo users can access a large number of pre-implemented DRL environments and algorithms, making it a comprehensive DRL platform. On this platform,users can further easily manage and tune DRL agents through an interactive training terminal. Evaluation results show that: com-pared to existing DRL libraries, RLzoo not only achieves a high degree of abstraction in its API design. It also provides numerous useful DRL algorithms and environments which are not available in other libraries.