DispatchGym: Grab’s reinforcement learning research framework

This post introduces DispatchGym, a new research framework for Reinforcement Learning (RL) studies applied to Grab's dispatch system. It provides a suite of modules including an RL algorithm, a dispatching process simulation, and an interface connecting them via the Gymnasium API. The framework is designed for reliability, cost-effectiveness, and user empowerment, allowing data scientists to independently develop and test RL concepts. Key features include a modular architecture, Python implementation accelerated by Numba for simulation, and a CLI for seamless deployment to Spark jobs. DispatchGym was applied to develop a contextual bandit strategy for auto-adaptive tuning of dispatch hyperparameters, experimenting with various bandit models and action sampling strategies. The post highlights the importance of aligning reward functions with business goals and sensitivity to control levers in RL applications.