Researchers at Open AI have taught an artificial intelligence to play Minecraft. The neural network was trained with the help of gameplay videos. As a result, the neural network was able to find the necessary amount of resources and make a diamond pickaxe.

To train the neural network, engineers developed a method of video pre-training (Video PreTraining, VPT), which allows you to use a large amount of uncompartmentalized data. At the same artificial intelligence used emulation of a standard mouse and keyboard for gameplay, which, according to experts, will help in the future to teach the neural network to use the computer without additional connecting interfaces.

For the first stage of training, engineers used marked-up videos of Minecraft gameplay for a total of 2,000 hours. For markup, they used data about the keys that users pressed during the game. At the output the researchers had a neural network that could independently process videos, guess keystrokes and record them. With the help of this neural network they automatically marked 70 thousand hours of gameplay from open sources.
The scheme of neural network training
The scheme of neural network training

As a result, the neural network learned to perform not only basic operations in Minecraft, but quite complex ones requiring consistent decision-making. For example, a neural network is able to extract resources and make items from them, run, swim, bypass obstacles, hunt animals for food and eat food, replenishing the hunger scale. The artificial intelligence also learned how to put blocks under the character by bouncing, allowing it to climb to higher ground.

After that, the researchers decided to fine-tune the neural network, and to do this, they asked the users participating in the project to create a new world in the game, collect all the necessary resources to start and make the basic necessities from them. At the same time, the users recorded the gameplay, and the obtained data were used for training. At the output, the neural network learned how to properly start the game and no longer wandered aimlessly, but sought to make a workbench with which to make in-game items. Also, some users were building basic shelters while recording data, and the artificial intelligence picked up on this skill.

Next, engineers used reinforcement learning to tune the neural network, which eventually allowed the artificial intelligence to make a diamond pickaxe on its own. The entire journey to get the tool was broken down into steps. The neural network was rewarded for each step taken. In the end, the artificial intelligence was able to complete the task.

Open AI engineers believe that the method can be used to train neural networks quickly and well using videos. In doing so, the video pre-learning method gives a stable base that can be refined and tweaked using other available methods. So far, the researchers have tested the performance of neural networks only in Minecraft, but they believe that this method will help to teach the artificial intelligence to use other computer programs with the keyboard and mouse.

Scientists have published a detailed study and posted the source code of the project. Also, the Open AI development team announced a MineRL NeurIPS contest, whose participants can use the neural network to solve more complex problems in Minecraft.

Автор: Иван Алексеев (e-mail: info@smolensk-auto.ru) Добавлено: 01.07.2022 16:12

Просмотров: 180