ScaRLib: A Framework for Cooperative Many Agent Deep Reinforcement Learning in Scala
Abstract
Multi Agent Reinforcement Learning (MARL) is an emerging field in machine learning where multiple agents learn, simultaneously and in a shared environment, how to optimise a global or local reward signal. MARL has gained significant interest in recent years due to its successful applications in various domains, such as robotics, IoT, and traffic control. Cooperative Many Agent Reinforcement Learning (CMARL) is a relevant subclass of MARL, where thousands of agents work together to achieve a common coordination goal.In this paper, we introduce ScaRLib, a Scala framework relying on state-of-the-art deep learning libraries to support the development of CMARL systems. The framework supports the specification of centralised training and decentralised execution, and it is designed to be easily extensible, allowing to add new algorithms, new types of environments, and new coordination toolchains.This paper describes the main structure and features of ScaRLib and includes basic demonstrations that showcase binding with one such toolchain: ScaFi programming framework and Alchemist simulator can be exploited to enable learning of field-based coordination policies for large-scale systems.