This paper summarizes the commonly used open source environment platforms for validation of reinforcement learning algorithms.

Once we design a reinforcement learning algorithm, how do we verify that the algorithm is good or bad? Just like data sets, we need a recognized platform to measure this algorithm. Such a platform, the most basic need to have simulation and rendering.

OpenAI Gym and Universe

The most famous names in the industry are Gym and Universe. OpenAI Gym is used to evaluate and compare the performance of reinforcement learning algorithms. Its interface supports algorithms in any framework, such as TensorFlow, Theano, Keras, etc.

Gym library collects and solves many problems in the testing process of the environment, which can make your reinforcement learning algorithm Work well. It also includes a game interface that helps you write more versatile algorithms. Now the mainstream of the industry is this environment, if there is no understanding of the need to configure the students, you can refer to a detailed installation and use of the tutorial I wrote before: intensive learning environment -Gym installation to use the introduction

MuJoCo is often seen at major summits for continuous control of physics engines, but this is for a fee.

RoboSchool

As an alternative to MuJoCo’s implementation, OpenAI developed Roboschool based on Bullet’s physics engine. It provides an OpenAI Gym-like interface for simulating robot control. It currently contains 12 environments. In addition to traditional mujoco-like scenarios, there are interactive control and multi-agent control scenarios.

  • GitHub link: github.com/openai/robo…
  • Official website: openai.com/blog/robosc…

In the Gym, there are manipulator for grasping and dexterous hand of mechanical arm, robot arm environment Robotics:

  • GitHub link: github.com/openai/gym
  • Gym.openai.com/

OpenAI Universe is an extension of OpenAI Gym, which can provide a more complex environment with a taste of real-time strategy games and certain requirements on decision time.

The release of Universe is also a boon for the industry. Any program can be converted into a Gym environment. Universe works on a VNC remote desk by automatically launching the program, so it does not require special access to the program’s internals, source code, or BOT APIS. Environment is packaged Docker image, installation and use will be more convenient.

  • GitHub link: github.com/openai/univ…
  • Official website: openai.com/blog/univer…

DeepMind Lab

DeepMind Lab is a first-person 3D gaming platform that presents scenes with rich sci-fi visuals. The actions available allow an agent to look around and move in 3D. Example tasks include collecting fruit, navigating a maze, traversing dangerous passages without falling off a cliff, using a launching pad to move from platform to platform, playing with a laser pointer, and quickly learning and memorizing randomly generated environments. DeepMind Lab has become a major research platform within DeepMind, which has both richer visual effects and more natural physical effects.

  • Official website: deepmind.com/research/pu…
  • GitHub link: github.com/deepmind/la…

Project Malmo

Malmo, led by Katja Hofmann, a Researcher at Microsoft Cambridge, is an artificial intelligence experiment and research platform developed by Microsoft based on Minecraft. The platform was also used in the first Collaborative AI Challenge.

Malmo is an open source platform for multi-agent reinforcement learning algorithms, requiring agents to cooperate with each other to achieve the ultimate collaborative AI. And researchers can speed up the clocks in Minecraft to speed up experiments.

  • GitHub link: github.com/crowdAI/mar…
  • Collaborative AI Challenge official website: www.microsoft.com/en-us/resea…

ViZDoom

ViZDoom provides the environment for playing doom games with AI (a first-person shooter). It provides a multi-agent competitive game environment for testing the quality of the algorithm.

  • GitHub link: github.com/mwydmuch/Vi…

My main research topics include reinforcement learning, computer vision, deep learning, machine learning, and so on. I want to share my notes and experiences in the process of learning. Look forward to your attention, welcome to learn and exchange progress together!