AI learns to solve Rubik’s Cube without human help
“Artificial intelligence” (AI) is used to describe machines that mimic the human mind like “learning” and “problem-solving”. A major goal in AI is to make systems that can learn how to master various surroundings and goals without human knowledge.
Now, researchers have created an AI system that learned to solve the Rubik’s Cube and other complex puzzles by itself. Without human help, the AI system figured out a strategy that takes half the number of human moves to crack the cube. This system has ‘real-world’ applications because the problems faced with these complex puzzles are common in robotics, healthcare, and the natural sciences (e.g. physics, chemistry, and biology).
It’s alive, it’s alive!!!
Thought-capable artificial beings have appeared in storytelling since ancient times. Millennia before speech recognition and self-driving cars became reality, the stories of the god Hephaestus and his creations—the giant bronze robot Talos and artificial woman Pandora—filled the imaginations of people in ancient Greece. From Frankenstein to The Terminator, artificially thinking beings have persisted in the past couple centuries of science fiction.
Early AI research began in the mid 20th century, with the term ‘artificial intelligence’ coined in 1956. In the 1960s, the US military took interest in AI and began training computers to mimic basic human reasoning. By the 1970s, the Defense Advanced Research Projects Agency (DARPA) used AI to complete street mapping projects.
Now, AI is all around us. It is updating your social media feeds and making decisions on what shows you should watch and podcasts to download. AI is capable of successfully understanding human speech. We are using AI to compete at the highest level in strategic game systems like chess and Go as well as in military simulations.
Domo arigato Mr. Robot
So, what can AI actually do? AI makes it possible for machines to learn from experience and adjust to new information. Many AI examples today—like navigation and speech-to-text capabilities on smartphones—depend heavily on machine learning to perform these human-like tasks.
Machine learning comes in two flavors: with and without human supervision. In supervised learning, the AI system first learns how to make sense of the data with an answer key in hand—there is prior knowledge of what the result should be. In unsupervised learning, the AI system learns without knowing what the solved puzzle looks like—it tries to make sense of the data by finding features and patterns on its own.
The AI that we have today that is very good at perceptual things—recognizing objects and images for self-driving cars, for example. However, currently, AI is not very good at reasoning, processing symbols, and advanced mathematics, all of which are dependent on machine learning. To push the envelope on machine learning, researchers are continually creating AI systems that find solutions to increasingly complex puzzles and problems.
Life is a bunch of puzzles
The Rubik’s Cube is iconic. It earned a place as a permanent exhibit in New York’s Museum of Modern Art and entered the Oxford English Dictionary in 1982. In television and film (or real life for that matter), a person’s ability to solve a Rubik’s Cube quickly is often used as a way of establishing a character’s high intelligence. 350 million have been sold—almost one in twenty people own a Rubik’s Cube worldwide.
Concerning AI, the Rubik’s cube is an interesting machine learning challenge for many reasons. First, it hasn’t been done before. Secondly, it’s a system that has a very large number of possible states that are all kind of random-looking yet there is only one solved state. It’s like trying to randomly drop a golf ball on a gigantic course that is very flat with the goal of finding a tiny hole.
Hasta la vista, baby
So, researchers made an AI system that solves the Rubik’s cube from scratch. They made a machine that plays with the cube, puts it in different configurations, and little by little learns how to solve it by taking the shortest path. In doing so, the system called DeepCubeA learned to solve the Rubik’s cube in nearly half the steps that the most expert humans make.
Senior author Dr. Pierre Baldi says, “Human experts, like those teenagers that solve the cube very rapidly. They take typically on the order of 50 steps. Our system takes on the order of 25 steps—it takes far fewer steps”
There are many reasons for this, says Dr. Baldi, including that human solvers have ergonomic restraints on their wrists that the machine obviously doesn’t have. But, more importantly, he says that there are interesting technical aspects in the solutions the machine is learning compared to what humans are doing. After all, the machine has ‘thought’ of a way to solve the problem better than humans can.
AI vs. 4D chess
Next, Dr. Baldi’s team is trying to solve increasingly complex puzzles—the Rubik’s cube is just one from a family. They also want to figure out how to apply what they learned from DeepCubeA to the real world.
“We’re also thinking about how to use this in robotics or problems in the physical sciences like the Rubik’s cube where everything looks random and there’s only one or very few solutions,” said Dr. Baldi.
Daily, there are teams around the world trying to improve AI by solving different games and problems. However, these are incremental and technical problems. The big questions long term are about how you make machines more intelligent. These problems are based off complex questions that are not solved—they are long term and researchers are making incremental steps.
Dr. Baldi articulates, “An analogy for where we are with AI is that of the Wright Brothers. When they worked on their first flying machines, they could fly fifty yards. Today, we have AI that can fly in very narrow directions for short distances and it’s brittle. We can’t fly long distances in fickle conditions. We are only at the dawn of understanding artificial intelligence.”
Evolution of the machine mind
Dr. Baldi thinks that there are probably many ways to build something that is intelligent. He says that there is no reason to think that the human brain is the only possible solution. For example, there is a deep learning system that can play Go better than humans. And, when human experts play against the system, they definitely feel that the machine ‘thinks’ differently.
“At the root of this is an existential question about the nature of who we are and whether machines can think or be conscious,” Dr. Baldi proposes. “I have always been interested in the brain and how the mind can arise from a physical system.”
As machine learning marches into the future, it will be interesting to see whether it is based off what we find in nature.
Dr. Baldi contemplates, “Airplanes were inspired by birds and their wings. However, there are many ways to fly. For example, helicopters fly in a completely different way. Probably the same thing will happen with AI. The brain still continues to inspire us. And, we are already seeing completely artificial ways of doing things that are available to machines and not humans.”
Soon enough, technological growth will become uncontrollable and irreversible, resulting in unfathomable changes to human civilization.
As the inventor of the cube Erno Rubik said himself, “The Cube is an imitation of life itself—or even an improvement on life.”
Bring on the technological singularity.