No! Test ai with games.

No! Test ai with games.
GuideMy world is no stranger to Jos é Hernández-Orallo, a computer scientist who is using the game for his own research. His research team at Valencia Polytechnic University designed a benchmark for testing machine intelligence performance, inspired by how he saw his children play in the 3D virtual world. In my world, players usually need to solve problems through interaction rather than shooting monsters.

Artificial intelligence program

Testing ai with games, checking out the three open source 3D testing environments "my world" is a global game that researchers are now using to test AI
In 2014, Microsoft purchased the ownership of "my world". The Research Institute of the technology giant, Microsoft Research Institute, then designed a new version for scientific research, allows computer programs and scientists to explore and customize 3D environments in the game. Later, Microsoft invited some external researchers, including Hernández-Orallo, to download this machine-friendly version of my world. Since July 2015, Microsoft has made it fully open. Now anyone can use it for free. Microsoft hopes to speed up research in the AI field.

Artificial intelligence research in various games has recently become very popular, and many companies are investing in research power in games just like Microsoft. In September December 3, DeepMind opened its own 3D virtual world program DeepMind Lab for all developers to download and customize. The virtual environment designed by Google was initially used to train its own artificial intelligence program. Just two days later, OpenAI released a "meta platform" that allows AI programs to interact with more than a dozen 3D games originally designed for human players. In addition, this environment also contains some Web browsers and mobile apps.

These three tools allow researchers and software developers to easily carry out their own experiments and how to solve the problems they have never seen before, it also helps programs perform self-training in environments similar to real scenarios. "Such a virtual environment will lay the foundation for the development of artificial intelligence," said Pedro Domingos, a machine learning researcher at the University of Washington in Seattle.

DeepMind Lab, which was initially used to train Google's own artificial intelligence program, is now open to all developers

Atari Algorithm

Artificial intelligence is an old player in various video games. However, in the early days, every algorithm used for customs clearance was specially customized. In recent years, people have begun to focus on using machine learning to accumulate experience on their own. In the first half of 2015, DeepMind launched an algorithm that exceeded human players in Atari games. The algorithm scored high scores through continuous trial and error, the designer did not tell the program what the goal of each game is.

The Atari game is just a 2D world. A "First Person" 3D video game like "my world" puts players in a three-dimensional environment, closer to the real world than the former, therefore, it attracted the attention of more researchers.

In my world, players can interact with bricks in the virtual world, build houses, explore routes, and interact with other content in the game world. The developer-oriented version is called Malmo, which allows machine algorithms to be explored in the game like human players. Hernández-Orallo is using Malmo to explore whether a virtual environment can be used to create a benchmark for machine intelligence. Different algorithms can compete with each other to see which one can build bricks into the shape of an object, or compare the time required for finding the path in the same maze, the coverage of this test is much wider than that of the Turing test, the most famous Test Method of machine intelligence.

Another reason that attracts AI scholars in my world is that players can type and communicate with each other in the game. "This can help experts study how AI interacts with humans in the real world," said Katja Hofmann, a scientist at the Microsoft Research Institute. She leads the Malmo development team in Cambridge, England.

About 100 research institutions are currently using the Malmo 3D world, developed by Microsoft Research Institute for AI research

Train robots

"Artificial Intelligence Training in the virtual world is very beneficial to the development of robots," Hofmann said. 「 Because the customization cost of the virtual environment is very low, the customization speed and security are also guaranteed. Virtual Environments can also enable robot researchers to focus on solving the intelligent problem of robots-the mechanical problem is sometimes truly distracting .」

In addition to Hernández-Orallo, Microsoft Research Institute has also worked with many other research institutions to launch a series of Malmo projects. Hofmann believes that the actual number of users is more than this, and there may be 100.

The Malmo platform includes Java modules and code that helps the agent perceive and operate in the Minecraft environment. These two components can run on Windows, Linux, or Mac OS.

Similar to Malmo, DeepMind Lab also allows researchers to create a maze, allowing different algorithms to learn how to find paths and get rewards. DeepMind is trying to add "more natural elements" (such as undulating terrain and plants) to this virtual world. Currently, this 3D environment is open-source, and DeepMind hopes to train smarter training algorithms with the help of other researchers. "Through open source, we can involve everyone and constantly improve this project," said a DeepMind spokesman.

The meta-platform of OpenAI, Universe, goes further than the first two 3D worlds. By providing different types of environments for the same AI program, this platform may solve the most difficult problem in the field: how to create an algorithm that can solve any new problem. The current deep neural network-a computer system created by imitating the structures of brain cells and visual cortex-allows you to quickly learn how to find paths in a 3D maze, however, the same system cannot use the methods it learns to navigate in other maze. 「 You just change the color of the maze and the system will be lost, 」 Hernández-Orallo said. 「 This is the most advanced technology at the moment, which is just a joke .」

It seems that there is still a long way to go in the development of artificial intelligence. At present, Microsoft is trying to make Malmo available in Universe. 「 Everyone will benefit from having a Community platform, 」 said Greg Brockman, co-founder and chief technology officer of OpenAI.

From: http://tech.163.com/16/1216/08/C8D4TLLC00097U80.html

Address: http://www.linuxprobe.com/game-test-3d.html