Slashdot Mirror
Machine Figures Out Rubik's Cube Without Human Assistance (technologyreview.com)
An anonymous reader quotes a report from MIT Technology Review: [Stephen McAleer and colleagues from the University of California, Irvine] have pioneered a new kind of deep-learning technique, called "autodidactic iteration," that can teach itself to solve a Rubik's Cube with no human assistance. The trick that McAleer and co have mastered is to find a way for the machine to create its own system of rewards. Here's how it works. Given an unsolved cube, the machine must decide whether a specific move is an improvement on the existing configuration. To do this, it must be able to evaluate the move. Autodidactic iteration does this by starting with the finished cube and working backwards to find a configuration that is similar to the proposed move. This process is not perfect, but deep learning helps the system figure out which moves are generally better than others. Having been trained, the network then uses a standard search tree to hunt for suggested moves for each configuration.
The result is an algorithm that performs remarkably well. "Our algorithm is able to solve 100% of randomly scrambled cubes while achieving a median solve length of 30 moves -- less than or equal to solvers that employ human domain knowledge," say McAleer and co. That's interesting because it has implications for a variety of other tasks that deep learning has struggled with, including puzzles like Sokoban, games like Montezuma's Revenge, and problems like prime number factorization. The paper on the algorithm -- called DeepCube -- is available on Arxiv.
The result is an algorithm that performs remarkably well. "Our algorithm is able to solve 100% of randomly scrambled cubes while achieving a median solve length of 30 moves -- less than or equal to solvers that employ human domain knowledge," say McAleer and co. That's interesting because it has implications for a variety of other tasks that deep learning has struggled with, including puzzles like Sokoban, games like Montezuma's Revenge, and problems like prime number factorization. The paper on the algorithm -- called DeepCube -- is available on Arxiv.
Someone had to tell it what is a solution. If you give it a solved cube, that's assistance. Is it really that hard not to inflate headlines?
Games are easy for "AI" because games have strict rules that a modeler can account for/predict.
This algorithm was able to figure out how to solve Rubik's Cube with no help from humans other than humans providing the (simulated) cubes, describing what the solution looks like, and designing an algorithm specific to solving Rubik's Cube?
Color me less than impressed.
it has implications for a variety of other tasks that deep learning has struggled with, including... problems like prime number factorization
If it could help with finding the prime factorization of large semi-prime numbers – ie two or more prime numbers that multiplied together result in a target original number - then that would be quite useful.
*cough* cryptography
Games are easy for "AI" because games have strict rules
Just because the rules are strict (or even simple) does not mean that the game is easy. You can achieve arbitrary complexity by iterating the rules a large number of times. For example, the rules of Go are strict, the question whether a given board position is winning for white is hard. The rules of a programming language are strict. Writing a Linux kernel is hard. The rules of math are strict. Providing a proof for Fermat's last theorem is hard. The rules of physics and soccer are strict. Making a robot that can beat a human at the game is hard.