santa-2023

[rose ruin]

Good luck everyone!

[craggy patio]

Is there only one message here?

Sus

[vernal warren]

i think not many people have started working on the competition yet, that's why

[craggy patio]

Why not?

[vernal cedar]

its quite new

[vernal warren]

plus I think many people won't try this challenge till the playground series competition is donr

[vernal cedar]

im trying to use reinforcement learning for this one

but it is very slow and annoying

but from my initial tests

i have brute forced the 2x2x2 cubes and an RL agent solved them in less time

so i hope this generalises...

[sleek kernel]

You'll have to train an agent for each puzzle type no?

[vernal cedar]

my testing is nowhere near complete but i think there are 4 possibilities

1 - 1 agent for each puzzle
2 - one agent for each puzzle type
3 - intuition (ignore impossible actions)
4 - domain specific action space reduction (https://arxiv.org/pdf/2004.00980.pdf)

[rocky spade]

how many time for solve using brute force?

[vernal cedar]

Well it depends on the particular puzzle

For now im working on the 2x2x2 cubes

And it will generalise

[rocky spade]

for the 30 instance of 2x2x2

[vernal cedar]

About 5 minutes

Using a poor quality python program

[rocky spade]

Actually I'm solving the 30 2x2x2 in 7 seconds. I don't think ML can be faster than that

[vernal cedar]

I got the RL agent to do it in 10 miliseconds per move

[rocky spade]

But RL can't prove the quality of the solution, or can?

[vernal cedar]

no

not easily

but my thought process was that for the 100x100x100 cubes it would be very hard

[sleek kernel]

Dom, how long are your episodes?

[vernal cedar]

as brute force scales (N^1−1)!

[vernal cedar]

depends on the problem

for 2x2x2 i have found 80 to be good

[sleek kernel]

You are baselining with 2x2x2?

And you built out the whole game tree?

[vernal cedar]

yes

just because proving for the harder ones seems almost impossible

[sleek kernel]

Well how much memory did the 2x2x2 take? In GB?

[vernal cedar]

brute force or DQN?

[sleek kernel]

Brute force

Did you throw in actual good paths in your replay buffer from the brute forced graph?

[vernal cedar]

well there are 3674160 states

and from each state i found the fastest path to a solution

so quite a bit at peak

like 6gb

[sleek kernel]

Not too bad. I just feel like it won't generalize well if you parametrize the action space to try to handle larger cubes

[vernal cedar]

well yes

but there is some evidence that it could work

thats what experimentation is for

[sleek kernel]

Indeed. It'll be interesting how it works with the 3x3 ones

[vernal cedar]

yeaj

yeah*

also I have worked with RL agents and a changing action space in the past briefly

[rocky spade]

I think that brute force not works on 3x3x3.
I will try localsearch to improva a poor solution

[vernal cedar]

yeah

but when you get to 100x100x100, even local search will be impossible hard to do

no?

[rocky spade]

i dont think so

once you have 1 solution to start

[sleek kernel]

You have too many actions you could do lol

The best idea I think will be to just build out the graphs starting from the solution state

And then sprinkle in known good paths with the random garbage patjs your agent starting from random init state

[rocky spade]

33x33x33 is the max, no?

[thick cave]

how do you represnt the state? If that's not a secrete

[vernal cedar]

yeah mb

[vernal cedar]

i was thinking wreath_100/100

[vernal cedar]

what do you mean?

[thick cave]

e.g. Do you just use one-hot for each position?

I was thinking of somehow representing the puzzel using a graph, and then do GNN. But couldn't think of a good idea

[vernal cedar]

https://arxiv.org/pdf/2210.03907.pdf

[vernal cedar]

I made sure that my representation was rotationally invariant so it is actually slightly less large

[thick cave]

Nice

[vernal cedar]

I am working on a different representation that removes colour symmetries

and reflectional ones too

but i havent used such maths in a long long time

[craggy patio]

Have GNNs ever worked for anything

[craggy patio]

damn solved ~all 2/2 and 3/3 cubes ~optimally and got 24th place

I thought it would be top 10

cause I miscalculated the score

which is the best puzzle type to focus on next?

[craggy patio]

Holy number 1 got sub 200k

[sand scroll]

At this point I am sure that people started to use ML models, or supercomputers 😄

[craggy patio]

well I'd expect everyone is using ML models

but yeah this is a very compute heavy contest, I'd expect the top spots have a decent number of large gpus

[craggy patio]

I think this server is now my own personal journal

solving the 4/4s is haard

my ubuntu keeps killing my program because it's taking too much memory

or maybe it's killing it for some other reason, my GPU's memory is only 50% full 🤔

yeah dmesg says it's out of memory

weird

[fathom hazel]

I'm still bruteforcing 2x2x2

[sand scroll]

I wrote something on Efficient Cube deep learning algorithm on Discussions and published one code on Codes.

It could be a starting point for a DL approach if you are not using it already.

Sorry for the code being too messy

[craggy patio]

I don't see it?

[fathom hazel]

Is it possible that the solution of the cubes is more than the theoretical maximum because the cube must be "aligned" in the correct direction? Or that shouldn't play any role? I am just bothered by the fact that a solution of my brute force approach on 2x2x2 has 2 of the same moves one after the other (e.g. '-f0', '-f0') and I'm tryin to explain that.

[fathom hazel]

wait, I think I got something wrong. Initially I thought both 90 and 180 degrees movements were allowed, but apparently they are not. That probably means that the maximum number of moves should be higher than I first expected. And having 2 of the same move is totally fine

[sand scroll]

the name of the discussion and the code is "Efficient Cube"

[craggy patio]

Max is 26

But also I think you're right about the orientation

But I didn't see any cubes over the maximum

[fathom hazel]

26? For the 2x2x2 cube? How so?

Isn't it 14 as it says on the internet?

[wraith vapor]

I also posted about efficientcube with TPU a while ago https://www.kaggle.com/discussions/general/405935

[craggy patio]

Sorry, I thought you were talking about the 333

[fathom hazel]

so on the 2x2x2 max is 14?

[uneven kindle]

hey! I made an interactive 3D viz with plotly of any NxN rubik's cube:
https://www.kaggle.com/code/edomingo/rubik-s-cube-3d-interactive-viz-plotly
let me know if this could be useful 🙂

[frank kayak]

Hi! i made a sollution that i want to submit but it needs 24hours (estimation)
i didn't make any optimisation and all calculations are on cpu whith no multi threading
Is there a processing time limitation on the challenge ? (first time i really try kaggle challenges and didn't find the information for this one)

[broken skiff]

You only need submit .csv file. There is no limitation.

[ebon heron]

It can take as long as you like on your own hardware and you can simply submit the resulting csv file, but if you choose to run it in Kaggle, they have a 12-hour limit on CPU jobs.

[sacred birch]

yes, 14 (according to google/chatgpt)

[vernal cedar]

Ughhh reinforcement learning is so annoying... Time for evolutionary networks i guess

[sleek kernel]

Did you grab a stablebaselines model? Or did you write your own?

[loud vapor]

actually an interesting problem to tackle

given the sparse reward structure

i wonder if anyone will come up with something worth publishing

[sleek kernel]

I'm thinking that move history does matter right? So we can penalize an agent for reversing a move, or doing the same one 4 times. Does that sound about right?

[loud vapor]

Possibly, maybe not moves but revisiting states. But thats seems implicit in the -1 per step formulation

[fathom hazel]

Yes but, would be useful to find if the fact that the orientation of the solution needs to be specific, causes any increase in the total number of moves.

Otherwise I get that the max moves are indeed 14 in a normal scenario

[forest cipher]

Hi everyone! I am looking for a teammate for this competition. I am currently rank ~30 (changes a lot) and I would love to discuss possible approaches to the problems.

[ebon heron]

@forest cipher Happy to send a team request for your consideration, currently looking to fill up our team slots with complementary ML and algorithm skillsets.

[steel ravine]

the leaderboard is crazy. 74321???????

[old arch]

now is 72407 😅

[split void]

Has anyone know the exact deadline time?

Ah I found - 11:59 PM UTC

[dense kindle]

In about 2 days 19 hours from now