Are you badass enough to beat a deep convolutional neural network at Go? Time to find out!
https://chrisc36.github.io/deep-go/
>>43623086
Are you kidding? I can't beat my pet cat at Go.
>>43623227
According to the paper the network is ranked at approximately 4-5 kyu
I've been playing this network for the past 42 minutes and now I fucking hate Go with a passion. This was my first time playing.
Non toroidal board
Go programs are apparently improving by 1 dan/year.
So from a very good amateur now, to master in a decade, and then superhuman.
It may not be truly 'solved' but computers win again...
>>43623587
>to master in a decade
Maybe even less than that. The authors say they can still tweak and change the kind of network they use to improve it.
I think we'll have a master ranked go program in 3-5 years.
I don't even know how to fucking play this game, much less how to beat an AI at it.
>>43623686
https://en.wikipedia.org/wiki/Rules_of_go
>>43623086
Will playing people who are really bad at the game (like most of us, I'd imagine) make it worse at playing people who aren't?
>>43624039
I don't think it's in active training, so no.
>>43624112
I should actually follow links before posting.
>>43623686
>>43623723
Literally easiest game in the world, anon.
(Rule-wise)
>>43624395
Alter igo variant is easier
its go with those changes
no pass
all groups without adjacent empty squares are removed on the same time.
if you make a move and the stone you added to the board is the only stone removed, you lose
>>43623086
Sometimes it tries to do illegal moves (I've only seen attempted suicides as of now, but it probably ignores ko in the right conditions). I think they should tweak it so that illegal moves get a weight of 0 and the choice is only among the remaining ones.
>>43624768
I've read now the part of the paper that talks about this. They actually accounted for ko, but they seem to have forgotten about suicides.
>>43623086
neat little game!
>>43624816
Ko is accounted for, superko is ignored - you can actually cheese the game into a perpetual "gain" (of exactly 1) for the human player by getting the first capture in a superko loop, inversely the computer can and will cheese you by being the first one to start the loop, which is why it's fairly happy to do that because it always maintains a lead and cannot get bored.
bad living/dead calculation in the centre. should never have started the chain.
lacks vision when picking opportunities. needs more heuristics to value edges/corners more than the mid, predictions on whether it can actually win fights, etc.
>>43625250
the AI doesn't know about ladders, it looks like.