This post is also available in: heעברית (Hebrew)

Artificial Intelligence (AI) researchers at Carnegie Mellon University created an AI based on artificial nerve networks, that is capable of defeating computerized and human adversaries at the Doom computer game. AI plays the game similarly to the way humans do: It receives as an input only the pixels on screen, and from there it can identify the location and situation of the character, navigate the map, shoot adversaries and evade their attacks, and even collect objects such as ammunition boxes and new weapons, and use them.

The system received positive feedback for attaining objects, rapid movement and adversary killing, and negative feedback when it sustained damage or died. The system that was finally accepted outperformed computerized and human actors alike.

Such systems can serve for the development and testing of effective strategies for taking control over building in urban terrain, especially when the buildings structure is known in advance and can be fed into the computer, according to techcrunch.com.

Moreover, it is also possible to feed into the system the locations and expected behavior patterns of guards and even civilians, and produce offers for the best action suitable for the soldiers who are supposed to participate in an operation on this site – for example, to release a hostage.

The developers claim that their model exploits game feature information such as the presence of enemies or items, during the training phase. The model is trained to simultaneously learn features along with minimizing a Q-learning objective, which is shown to dramatically improve the training speed and performance of the autonomous agent. The system’s architecture is also modularized to allow different models to be independently trained for different phases of the game.