Artificial intelligence managed to navigate in a maze
Programs were taught to look for a way out in virtual mazes where the shortest way to the target was blocked by a closed “door”.
The company DeepMind, a Google branch working on the AI research, created a program that could build the optimal routes using an analog of grid neuron cells. These cells constitute a brain network that regulates navigation in all mammals including humans.
The authors of the new algorithm created an artificial analog of grid cells. These brain cells activate when a mammal crosses the line of an imaginary coordinate grid “laid” on the space where the animal is.
In humans deterioration of these neurons is one of the symptoms of Alzheimer’s, where people lose the ability to navigate in space. Scientists assume that grid cells help to look for the shortest routes in familiar surroundings.
In the newest research developers modeled two artificial recurrent neuron networks. In such networks the connection between the elements forms a directional sequence: the program uses its previous steps to plan the next action. One algorithm used artificial grid neuron cells, the other did without them.
Programs were taught to look for a way out in virtual mazes where the shortest way to the target was blocked by a closed “door”. Then algorithms switched to larger mazes of the similar configuration: the program that was using grid cells found the way more efficiently.
When the “door” was opened, the algorithm was able to take in this fact and found the shortest way. The program that was working without the special neurons ignored the opened pathway and looked for the way out longer.
The results of the experiment confirmed the guess of neurobiologists: grid cells indeed take part in the search of the shortest route.