[Prior Research Team Lee Jung-woo]
For a long time, games (Go, chess, Atari games, etc.) have been used to verify the performance of reinforcement learning algorithms. With the development of algorithms, there are many appearances in the field of reinforcement learning who want to solve various problems that exist in reality and create services, just like other fields of image and natural language.
In this article, we will look at what fields are being covered through the recently released three reinforcement learning environments.
1. DeepMind – AndroidEnv
AndroidEnv is an environment for reinforcement learning on Android that was released by DeepMind. You can access basic apps that can operate in the Android environment, and you can experiment to see if you can learn how to use the same way as humans do in the modern times when smartphones are used a lot. The example task also provides some games, so you can learn how to play and act like a human using touch and drag gestures.
Because of the importance of the environment, reinforcement learning is applied to the platform devices most commonly used today, and what real problems can be solved and applied can be studied. Also, if algorithms are developed from a service point of view, it is expected that various reinforcement learning applications will emerge on the Android platform.
2. Facebook Research – CompilerGym
CompilerGym is a new kind of reinforcement learning environment never seen before. As mentioned earlier, there were many environments for robot control in games or simulation situations. The Facebook research team seems to hope that reinforcement learning's ability to find the optimal policy can be applied well to fields that require optimization in the computer field. Compiler is a language translation program that is responsible for changing a programming language written in high level into a low level language suitable for computer systems. This compiler also has a process of optimization using several APIs, and reinforcement learning finds policies to reduce compilation time in the items of Compiler optimization.
If reinforcement learning is proven to be helpful through CompilerGym, it is expected that it can be used to increase the efficiency of many kinds of computer systems.
3. Facebook Research – Habitat Lab
Habitat Lab is an environment for experimentation in the field of Embodied AI. Embodied AI interacts with virtual worlds and other virtual robots, and is the field of solving AI problems in virtual robots. Habitat 1.0 was previously released, but this time it has been updated to Habitat 2.0. In addition to simply moving the robot, it transmits questions and commands in natural language to perform commands, or finds a location that can directly serve as the basis for the correct answer to find the correct answer.
We hope that the Habitat Lab environment will create a robot AI that can communicate well and help people in the real world.
In this article, we introduced three recently released environments for reinforcement learning. It would be good to see how the development of reinforcement learning using open environments will directly affect our lives.