Could this mean the definite next step in gaming development?
Nvidia, who we all associate with graphics cards, has been carrying out some serious research in the AI department (that’s Artifical Intelligence for us noobs). One of their latest projects has been to test the capability of an AI system to recreate the basic visual setup and rules by only “watching” a game in action. The lab-rat of choice was good old Pac-Man, which resulted in a near to perfect replica of the game.
The system was not given any pre-coded data or images to build on, but only fed basic control inputs and visual data which it then created frame-by-frame. The end product is a playable copy of the original that Nvidia will be releasing for online play at a later date.
Within the AI system used for the recreation of Pac-Man, is a program called gameGAN (Generative Adversarial Network); which works in two halves: first replicating the input data whilst the second compares it to the original. If the replication does not match the original, the data is rejected, re-tweaked and then resubmitted. The project with Pac-Man had little human input in terms of game-play learning. Instead, an AI-agent was used in order to produce around 50 000 episodes of game-play data. This in turn revealed a snag in programming, as the agent was so good at the game, it hardly ever died. This saw difficulty in the AI learning the concept of dying – instead of the title character doing his famous “wa-wa” on his capture, the ghosts would instead follow him around like “baby ducks following a parent” (via source).
While not perfect, the results are very impressive and have placed a benchmark in the world of game design with all its possibilities. The speculation is that this AI system could be used to design new levels or even mesh levels together, giving designers more power to tweak and perfect games.