NVIDIA AI reproduces 'Pac-Man' without a game engine just by watching 50,000 play videos
by
When you see a play movie of a game that you have never played, you may have thought of 'where you play that game' in your mind and feel like playing it somehow. Similarly, by making AI visually learn the game play of the masterpiece action game `` Pac-Man '', an attempt to generate Pac-Man only by simulation by AI without preparing an existing game engine or code, NVIDIA research team announced.
GameGAN
https://nv-tlabs.github.io/gameGAN/
PAC-MAN Recreated with AI by NVIDIA Researchers | NVIDIA Blog
https://blogs.nvidia.com/blog/2020/05/22/gamegan-research-pacman-anniversary/
Pac-Man re-created by Nvidia AI that watched hours of gameplay-Polygon
https://www.polygon.com/2020/5/22/21266829/pac-man-nvidia-ai-game-gan-40th-anniversary
What the AI 'GameGAN' developed by NVIDIA is is explained briefly in the following movie.
NVIDIA GameGAN: Celebrating 40 Years of PAC-MAN with Game-Changing AI-YouTube
On the right is Pac-Man, which GameGAN finally generated. Pac-Man generated by GameGAN can be output as a playable application, not just a demo video.
GameGAN doesn't just play Pacman ...
The big point is to 'play Pac-Man's play video and reproduce Pac-Man'.
GameGAN created playable Pac-Man from scratch by learning and training without using any off-the-shelf game engine.
This study was celebrated on May 22, 2020, when Pacman, released on May 22, 1980, celebrated its 40th anniversary.
The NVIDIA research team made GameGAN learn about 50,000 Pac-Man gameplay movies provided by the
'GameGAN is the first study to use a
The adversarial generation network consists of two networks, the generation network and the identification network. A mechanism that raises the accuracy of AI itself by repeating the cycle 'the identification network judges what the generation network generated and feeds the result back to the generation network.' In other words, the generation network of GameGAN that learned by watching 50,000 play movies generated Pacman as a simulation, and the identification network judges the result, so the accuracy of the simulation by GameGAN becomes higher. is.
GameGAN repeats learning and training, 'Pacman moves in the maze and can not pass through the wall' 'Phantom chases Pacman' 'Pacman dies when touching ghosts' 'Power food item He learned the basic rules of Pacman that 'the ghost turns blue when he gets it, and Pacman can eat ghosts', and he fed it back to the game to generate.
Pac-Man generated by GameGAN can be output as an application, and humans can also play. The accuracy of the finally generated simulation is extremely high, says Koichiro Tsutsumi, Head of Future Development Division of the Bandai Namco Research Institute, `` I can not believe that AI can reproduce the fun of Pacman without a game engine. However, I was surprised at the result of playing it. '
However, since the gameplay skill of GameGAN was too high, GameGAN itself rarely went over the game, and the research team reports that the bias that 'Pacman will not die' was formed in GameGAN. Therefore, in Pac-Man generated by GameGAN, 'Pac-Man operated by the player will die' is not allowed, and if an operation that would normally cause Pac-Man to die is input, AI will avoid that result. It seems that the game itself may be modified.
NVIDIA's research team is also trying to generate DOOM with GameGAN from training data of the reinforcement learning project '
In this research, the research team has set the purpose of 'AI creates an environment for autonomous machines to simulate', and 'carries a warehouse robot that learns how to grasp and move objects, carry food and medicine' Learning to neural networks may replace the creation of simulations needed to develop autonomous machines, such as delivery robots that have to navigate sidewalks in order to do so. '
In addition, NVIDIA says that the Pac-Man demo actually generated by GameGAN will be released after the summer of 2020.
Related Posts:
