AI Weekly #2: Deep Reinforcement Learning-Machine solves the Rubik’s Cube
Solving complex problems with minimal human interference has been sought after, especially in the age where a lot of work is digitalized. Companies are turning to AI and Deep Learning for managing workflow, but let’s step outside the work area for a while.
In recent years, deep reinforcement learning algorithms coupled with self-play have attained superior abilities in games such as Go, Chess and Shogi without using knowledge of human data. The goal of reinforcement learning is to create smart agents that respond respond to its surroundings through the analysis of their own interactions.
The Rubik’s cube solved without human assistance.
To nearly everyone, solving a Rubik’s cube single handedly within a short time period is pretty near impossible. Although some records have been made, there has never been a digitalized method to solving a Rubik’s cube. But now, scientists from the University of California, Irvine (UCI) have constructed a method called Autodidactic Iteration. In short, it analyzes a proposed move to solve the cube, working backwards from a finished one, at the same time finding a way that the machine can reward itself from making moves.
What sets this apart from automating chess moves, for example, is that in chess, each move can be calculated and awarded, in part due to the larger space required to play it. The reward process can help the machine discern good moves from bad moves. In contrast, the Rubik’s cube, being more complex, it is more difficult to determine whether a move is closer to solving the puzzle. Not being easily discouraged, the team from UCI developed an algorithm that is able to solve 100% of mexed-up cubes in around 30 moves – almost on par with those using human knowledge. (source: https://arxiv.org/pdf/1805.07470.pdf )
Why this matters
So in what way is this knowledge valuable, you might ask?
In this day and age, it is crucial to make progress in teaching machines to perform a wide range of functions with as little human assistance as possible, preferably none.
Without a doubt, making constant progress within Artificial Intelligence requires the development of learning algorithms where human supervision is limited.
By doing to, this opens a wide door for opportunities in solving different, complex mathematical problems or clarifying physical or biological phenomena, among others.
This signifies a serious step in the evolution of the possibilities of AI, one that will most certainly bring about beneficial changes at large.