Reinforced-lib: Rapid prototyping of reinforcement learning solutions

Abstract

Reinforcement learning (RL) is emerging as a promising framework for training intelligent agents to solve complex problems. However, developing RL solutions involves a complex process that requires experimenting with different models, agents, and parameter tuning. Existing RL libraries are often unable to meet the requirements set by researchers. In response, we introduce Reinforced-lib, a lightweight Python library designed for rapid development of RL solutions. Our open source library is tailored towards researchers working in areas where machine learning has not yet been applied, with a primary emphasis on user-friendliness. It offers the flexibility to employ both deep reinforcement learning (DRL) and traditional non-neural agents, along with a rich functionality and comprehensive documentation. Built on JAX, a high-performance numerical computing framework, Reinforced-lib grants access to a wide machine learning ecosystem. Moreover, we demonstrate the library’s effectiveness in a specific challenge from the domain of wireless networking and reproduce the results of an existing research paper that employs DRL.

Keywords

• Machine learning
• Optimization
• Reinforcement learning
• Wireless networks