Reinforcement Learning (RL) is one of the crucial areas of machine learning and has been used in the past to create astounding results such as AlphaGo and Dota 2. It typically refers to goal-oriented algorithms that learn how to attain complex objectives with superhuman performance. These algorithms are penalised when they make wrong decisions and rewarded for making the right ones, making it one of the most reliable algorithms to work with.
There are many RL platforms available today that allow users to develop and compare algorithms. In this article, we list 10 such popular RL platforms that every developer must know to speed up his/her research in the field.
Sign up for your weekly dose of what's up in emerging technology.
(The list is in alphabetical order)
It is a suite of RL environments that illustrate various safety properties of intelligent agents. The environment is implemented in pycolab, a highly-customizable gridworld game engine that allows recognising AI safety problems into robustness and specification problems, depending on whether the performance function corresponds to the observed reward function.
Amazon SageMaker RL is built on Amazon SageMaker, adding pre-packaged RL toolkits and making it easy to integrate any simulation environment. It allows focussing completely on RL problem as the training and prediction infrastructure is fully managed. It also allows creating a custom environment using other RL libraries such as TensorForce or StableBaselines. It has the first party simulators for AWS RoboMaker and Amazon Sumerian, Open AI Gym environments and open-source simulation environments, customer-developed simulation environments and more.
DeepMind Control Suite serves as a performance benchmark for RL agents. It comes with a set of continuous control tasks with a standardised structure and interpretable rewards. In the DeepMind Control Suite, the tasks are written in Python and is powered by MuJoCo physics engine, making them easy to use and modify. It also includes benchmarks for several learning algorithms and the control suite publicly available on GitHub.
4| DeepMind Lab
Deepmind Lab by Google is an integrated agent-environment platform for general artificial intelligence research. It was built to accommodate extensive research done at DeepMind and is based on an open-source engine ioquake3. Deepmind Lab can be used to study how autonomous artificial agents learn complex tasks. It also has simple and flexible API which enables creative task-designs and novel AI-designs to be explored and quickly iterated upon.
This RL framework by Google is an open-source research framework for fast prototyping of reinforcement learning algorithms. It is TensorFlow-based research framework that enables fast prototyping of reinforcement learning algorithms. Dopamine allows for easy experimentation and flexible development while being reproducible and compact and reliable.
This RL platform by Facebook is touted to be the first open-source reinforcement learning platform for large-scale products and services. The platform was developed to bridge the gap between reinforcement learning’s growing impact in research and its traditionally narrow range of uses in production. Facebook has been constantly working on improving the platform’s ability to adapt RL’s decision-based approach to large-scale applications. It includes workflows for simulated environments as well as a distributed platform for preprocessing, training, and exporting models in production.
7| OpenAI Gym
It is a toolkit that allows developers to both develop and compare reinforcement learning algorithms. There are essentially two parts to OpenAI gym — the open-source library and the service that includes their API. It contains a variety of environments and examples for testing reinforcement algorithms. It is one of the most used and preferred RL platforms, where reinforcement learning is done at a proprietary level. Since it is open-source, it increases the chances of achieving safe artificial intelligence that is fairly available to all.
This research initiative by Microsoft research aims at building AI agents to do complex tasks. It is a sophisticated AI experimentation platform built on top of Minecraft, designed to support fundamental research in artificial intelligence. It sets out to address the core research challenges by integrating reinforcement learning, cognitive science, and many ideas from artificial intelligence.
PLE is a reinforcement learning environment that allows a quick start to RL in Python. It mimicks the Arcade Learning Environment interface, allowing practitioners to focus on the design of models and experiments instead of environment design. It has been tested with Python 2.7.6 and hopes to eventually build an expensive library of games.
It is a DeepMind’s Python component of the StarCraft II Learning Environment (SC2LE), and is exposed as a Python RL environment. The collaboration between DeepMind and Blizzard develops StarCraft II into a rich environment for reinforcement learning research. PySC2 provides an interface for RL agents to interact with StarCraft 2 to get observations and send actions.