Relying on deep learning, AI has been able to create games like Bao Ke Meng.
▲ Real "tailor-made."
Recently, concepts such as "AI painting" and "AI composition" are popular. This kind of behavior of automatically generating works according to short words and descriptions provided is undoubtedly the dream of many people. What if similar concepts spread to games? Could it be possible to generate games automatically based on a given description?
A deep learning scientist from the University of Washington recently tweeted about his new work: given images and data, he can make computers generate games on their own.
The researcher, named "Ollin Boer Bohan," released an AI-generated game screen. Although the screen is slightly blurred and the map will deform strangely in some places, it is believed that most players can still find such a painting style with a strong "Pokémon" shadow at a glance.
▲ "Pokémon Pearl Diamond" game screen
Whether it is the room in the video or the grass, woods, high and low ground settings, they are very similar to the "Pokémon Pearl Diamond" released on the NDS platform in 2006. The author also said on his personal homepage that the learning data provided to AI is a large number of "Pearl Diamond" game images.
However, it should be pointed out that what AI generated was not the video above, but a game that could be operated.
Through the keys on the keyboard, players can operate the characters in the screen to move "up, down, left and right" in four directions. By pressing the direction keys twice quickly, they can also trigger the sprint action of the characters. This is all the functions that this game can achieve at present.
Although it looks very simple, the map will also appear strange tear screen, but the author said that he did not write a line of code for the game, all the functions and images are generated by AI through deep learning algorithms such as neural networks.
The phrase "no game code" can be simply understood as: the input operation instructions actually have nothing to do with the screen. After pressing a certain button, it actually gives only the image that AI guesses will appear, which also explains why the map in the video always shows strange changes.
However, even if it was such an effect, it would not be easy to achieve. The author said that in order for AI to understand the desired game effect, in addition to a large number of reference videos, it is necessary to manually mark some key frame images of operations to establish a database. After a lot of experiments, the effect in the video is finally completed.
Although the function realized at present is relatively simple, if the accuracy of this "imitation" can be improved, the author thinks that a complete game system can be simulated theoretically. Over time, perhaps "AI games" is the best solution to "electronic impotence."
This article comes from Weixin Official Accounts: Game Research Society (ID: ysaag), author: Yingrou Satellite
