University of California-developed AI taught itself to solve a Rubik's Cube in one second
Researchers at the University of California developed an artificial intelligence that taught itself to solve a 3x3x3 Rubik’s Cube in just over a second. While not the fastest machine to solve a Rubik’s Cube (that title is currently held by a robot that solved one in 0.8 seconds), it is an interesting look into how machine learning tackles logic problems and puzzles.
The AI, dubbed DeepCubeA, ran through 10 billion different start states of the Rubik’s Cube and was instructed to solve the Cube in 30 moves or fewer. The AI did relatively well, solving 1000 tested iterations. Perhaps more impressively (from a machine learning view), DeepCubeA found the shortest solution path 60.3% of the time. Keep in mind that this was achieved solely by the algorithm running within DeepCubeA; outside of creating the original AI, there was no outside assistance given to the machine.
The research team also tested DeepCubeA on other algorithmically-based logic puzzles with one goal state. Notably, the team ran DeepCubeA through multiple iterations of sliding tile puzzles with 15 tiles, 24 tiles, 35 tiles, and 48 tiles; a 7x7 game of Lights Out; and the box-pushing game Sokoban. DeepCubeA performed well in most tests, beating other algorithmic systems in each puzzle.
The full article detailing the research methodology was published in a recent issue of Nature Machine Intelligence. It is an interesting look into how far AI has come. Perhaps we will soon see an artificial intelligence that can find optimal solutions to simple real-world problems, laying the groundwork for complex learning algorithms that could control everyday systems. We’ll leave it to you to decide whether that is a dream or dystopian future