Neural Networks are one of the biggest tools that the AI researchers use. In summary, they enable natural learning of an artificial algorithm, in order to achieve an assigned goal.
In a recent experiment done by French researchers, the goal was to have robots get closer to a food supply, which meant earning points, and keep away from a poisonous food supply, which meant loosing points. The robots were also emitting light and were able to see light emitted by other robots.
The experiment included 1000 robots divided into 10 groups. It began by assigning each group of robots a 264 bit binary “genome” code which determined the mechanism of robots achieving the goal. After the first round, the most successful genomes were selected. And the experiment was run again. This process was repeated hundreds of times.
Eventually, the robots’ success for finding the food source increased. In addition to that, because of the limited amount of the “food” source, some robots learned to conceal their lights when they reached the food source, in order to keep other robots away. This was not a given rule in the algorithm, which the robots learned themselves.
Assigning costs or points to different situations in order to earn points is actually nothing new. One famous example is chess software. However, combining this with the natural development ability of neural networks seems to produce interesting results, and is likely to find applications in a variety of areas in artificial intelligence and robotics.