An End-to-end Automated AI System for Reinforcement Learning

Abstract

Deep Reinforcement Learning (RL) has emerged as a promising technique for solving sequential decision-making problems under uncertainty in many real-world applications. However, it is known to be highly sensitive to internal hyperparameters and requires significant expert manual effort for tuning an intractably large number of configurations. We present an Automated AI system for RL, combining both open source and proprietary algorithms in a unified framework and enabling rapid algorithm selection and hyperparameter optimization, thus allowing non-experts to benefit from SOTA RL solutions. Our system supports online and offline RL algorithms, achieving optimal solutions for a large set of benchmark RL environments. The system also provides hosted service on IBM API Hub where users can make REST API requests to train RL agents.