SEED, a research division of Electronic Arts, has unveiled a system that will help improve the artificial intelligence of opponents in large 3D games, including multiplayer shooters.
Their goal is to create an AI that will not execute prepared scenarios, but react to the player's actions based on the same information that a living person would have.
The developers have created a special space where the machine can learn. This is a kind of polygon with primitive virtual opponents, where the "agent" under the control of a neural network is given the task of defending a key area.
The "agent" has a limited supply of health and ammunition, which he replenishes with the help of bonuses. At the same time, he learns to navigate the terrain himself - due to information about what is in his field of vision and a radar operating at short distances.
So far, researchers have succeeded in making the machine behave like a player in a similar situation.
For example, when the AI runs out of ammunition, it instantly changes tactics and looks for a place to replenish ammunition. In a three-dimensional city, he moves without colliding with obstacles and peeking around the corners.
According to SEED staff, the standard way of creating AI, where each combination of actions is considered a separate action by the program, becomes impractical. In modern games, the researchers note, the number of such combinations can exceed a million.
Google has already conducted similar experiments with artificial intelligence in three dimensions, but EA is going to develop these ideas.
Promotional video: