The new system of artificial intelligence (AI) DeepCube over the 44 hours had taught himself how to assemble a Rubik’s cube.
Before AI could play chess and Go. However, the system is reinforcement learning, which was used in these cases are not suited to solve 3D puzzles like the Rubik’s cube.
In the case of chess or Go AI is easy enough to estimate your move as successful or unsuccessful and based on this, draw conclusions for future games. But in the case of Rubik’s cube AI could quickly determine the consequences of the decision, has not received a conditional “reward” that encourages learning, and, consequently, not studied.
When you create DeepCube team of programmers led by Stephen Makaira used a special system. Thanks to her, after each turn the AI “jump” to the already assembled cube and thus determines the “strength” moves. When the system collects enough data, it uses the classic method of “decision tree”, checking every movement to learn, thanks to what’s the quickest way to solve the puzzle.
© 2018, paradox. All rights reserved.