If robots have been able, for a long time, to solve a Rubik’s Cube in a very short time, they are always guided by the man and do not have faculty of learning: their software is summed up very often to a memorization of all the possible schemas of the game, which they simply apply according to the context. For DeepCube, it’s very different. Researchers from the California Universe, based in Irvine, have devised an algorithm that can determine how to solve the puzzle in 3D, without human intervention.
“Our algorithm is able to solve 100% randomly mixed cubes, with a median resolution of 30 shots, less than or equal to that of resolvers using human domain knowledge, “ say the scientists. DeepCube is able to learn from its tests and errors, starting from a finished cube and modifying it to determine if a movement is a useful improvement over a given mixed cube. Once the training is complete, DeepCube composes its own search tree as a reference point and compares the learned configurations to solve the mixed cube.
It took DeepCube 44 hours to solve a Rubik’s Cube alone. Some will say that we are far from human prowess, where some experts take only a few seconds to solve a cube. But do not forget that tens or even hundreds of hours of learning are needed to get there.
The goal of the researchers is not to limit DeepCube to the Rubik’s Cube resolution: they wish to use it, eventually, for the “folding of proteins” or the prediction of the tertiary structure of proteins. A challenge with very different issues!