• Welcome to the Speedsolving.com, home of the web's largest puzzle community!
    You are currently viewing our forum as a guest which gives you limited access to join discussions and access our other features.

    Registration is fast, simple and absolutely free so please, join our community of 40,000+ people from around the world today!

    If you are already a member, simply login to hide this message and begin participating in the community!

DeepCube (Using General AI for getting cube solutions)

abunickabhi

Member
Joined
Jan 9, 2014
Messages
6,687
Location
Yo
WCA
2013GHOD01
YouTube
Visit Channel

wiredknight

Member
Joined
May 14, 2019
Messages
9
There is a new paper on this topic, published 15 July 2019 in Nature Machine Intelligence by Agostinelli, McAleer, Shmakov, and Baldi
"Solving the Rubik's cube with deep reinforcement learning and search". It describes DeepCubeA, which I gather is an upgrade to DeepCube. The project is described for the general public here:

https://www.futurity.org/rubiks-cube-artificial-intelligence-2115172/

The title of the futurity article is, IMHO, misleading "A.I. taught itself to solve Rubik's Cube in under a second". When I first read that, I thought, "meh - MIT has a system that solves a PHYSICAL cube in less than a second" and almost didn't read the article. I think futurity meant that longest it takes on any single scramble is less than a second (again, meh) because, Dr. Baldi states that DeepCubeA taught itself in two days with no direct human help. But even that isn't so impressive, because DeepCube taught itself in 44 hours as reported in 2018.

So what's the big deal? Apparently it is the fact that DeepCubeA finds the shortest path about 60% of the time taking ".... about 20 moves, most of the time solving it in the minimum number of steps". DeepCube was hitting about 30 moves.

So I guess it is closing in on the fabled God's Algorithm.

I was curious about the test cases used and I had a hint that they could be found on codeocean.com under "Learning to Solve the Rubik's Cube", but my software skills were not up to the task of navigating the codeocean project. How about yours?

What I would like to see are statistics on DeepCubeA solving a large number of scrambles across a range distances, with a focus on something like the d20 list ala Kociemba at cube20.org.
 
Top