Google DeepMind has created a new AI capable of playing computer games well. What is he capable of?

Google DeepMind has created a new AI capable of playing computer games well. What is he capable of?

Habré has written more than once about artificial intelligence, which surpasses humans in various board and computer games. But these are specially trained agents who specialize in a specific game. And is it possible to develop a system capable of interacting with the three-dimensional environment of any game without extensive preliminary preparation? Google thinks so, and has backed it up with action. She created an agent capable of this. What are the possibilities of the new development?

The Google DeepMind team and its new project

Representatives of DeepMind created the SIMA agent, which is an abbreviation of four words: Scalable, Instructable, Multiworld Agent. He is not trained to win, but he can understand what is happening on the screen and recognize natural language. At the same time, the agent is able to navigate well in different games.

At the moment, the project is at the stage of proof of concept, it cannot be considered completed. But the results are encouraging and allow us to say that it is possible to bring the new product to the level of an ordinary gamer, at least not now, but in the future. In general, the goal of “catching up and overtaking” a person is not worth it, most likely, researchers want to provide a basis for creating AI platforms of various levels. People will be able to interact with them in various areas, not only gaming.

Here is what the authors of the project say:

“Learning to play even one video game is a technological feat, but if you train an agent to follow instructions in different game environments, you can create a system that will prove useful to humans in the real world as well.”

According to the same developers, a 3D game is one of the best environments for learning AI. The team chose nine games that allow for open interaction with the world and where at the same time there is no excessive level of violence. However, there is the exploration of outer space, the study of the game world, and simply the ability to do whatever you want. Google worked with eight game developers, including Hello Games, Embracer, Tuxedo Labs, Coffee Stain, and others, to train and test SIMA in games like No Man’s Sky, Teardown, Valheim, and Goat Simulator 3.

The agent was not given access to the game’s internal kitchen or even the API. No, instead, the system receives data from the screen and controls its actions with a keyboard and mouse – everything that a person has done in the last 50 years of the development of the gaming world. Moreover, the agent works in real time without hundreds of thousands and millions of hours of training, as in other cases.

This increases the complexity of this particular project, but allows the AI ​​agent to integrate into the new game with minimal settings and without special training. Scientists are now studying the behavior of the agent and watching how the AI ​​system uses the experience gained in one game when moving to another environment. According to the team, this could be a key step towards the creation of general artificial intelligence.

For the initial minimal training, SIMA uses video of a person playing games, accompanied by a description of what is happening in the footage. The videos show actions that can be performed within 10 seconds. Also, the agent integrates with pre-trained models such as SPARC and Phenaki, which avoids the need to spend a lot of time “understanding” speech or visual data.

What happened as a result?

DeepMind researchers trained and tested the system on 1,500 different tasks, which can tentatively be divided into nine different skill categories — from movement “go forward” and navigation “return to the ship” to gathering resources “find raspberries” and managing objects “cut potatoes”. The effectiveness of the model was then evaluated both from a computer point of view, i.e. whether the task was completed or not, and from a human point of view – an ordinary user was asked to watch a video of the agent’s gameplay and say whether SIMA was successful.

SIMA managed to find commonalities in different games. For example, you can usually move forward by pressing the W key. In addition, the model is well oriented in the instructions that are given using text and voice. An example is “jumping on something”, it can be a car in one game or a box in another.

In the course of testing, the agent was able to perform some basic tasks – for example, approach the spacecraft, even when the target object was not visible. But success still varies depending on the team. Thus, the model showed 75% success in solving tasks related to driving a car, compared to 40% success in performing tasks related to walking.

However, the result is still good, at least SIMA outperformed agents trained for specific titles in all nine games. Accordingly, it showed wider universal capabilities.

There is still a lot of work to do

Yes, you should not think that the successes that the researchers are talking about are evidence that SIMA can go through any game or at least a separate episode, as a person does. So far it’s just about some commands and actions. For example, in No Man’s Sky, the model succeeded in only 34% of the tasks tested, compared to 60% for a human. However, the tasks were difficult, which explains the low percentage of success even for the agent and the person.

Most of the problems are related to the fact that the agent does not always understand how to achieve success. For example, SIMA knows what “chop a tree” is, but how to target a specific plant specified by the user is not yet clear. In addition, one more example can be given. This is shooting at enemies. The agent successfully uses the weapon to defeat the target, but as soon as it disappears from the screen (for example, an NPC walks behind the character), SIMA immediately forgets what to do. A person is perfectly aware that the enemy must be pursued, as they say, to the victorious end. Or a computer opponent will win, not a human gamer.

The project team is currently working to ensure that future versions of the model can perform a wide range of tasks, not only narrow ones like cutting down trees, but also finding resources and building a camp, which already requires strategic planning. Well, the last stage is the use of game-trained agents in the real world to solve various tasks.

Related posts