In ancient China there were four 'arts' which were considered by educated people as aspirations to master, they were : music, calligraphy, painting and Go (Weiqi/Baduk).
In this new age of COVID in which everyone in the world is encouraged to "stay at home" and Netflix has complained about streaming bandwidth bottle-necking the Internet and even PC gaming platforms like Steam now throttling downloads of gaming updates etc I thought it was good to start a thread about this ancient 3,500+ year old game of Go in hopes that some may take it up as an interest and derive some enjoyment from it.
To start, here is a free "public domain" book around Go that is hosted on the Internet Archive:
To me Go is a flexible and intrinsically rewarding game in that like Chess it stood the test of time, and can be a life long hobby that one could "dive into" and "get lost in" etc..... The rules to Go are simpler than that of Chess, a five year old can learn the basic Go rules in about five minutes, but the depth to Go can be potentially far more vast than Chess and certainly the fact that its 'search space' is orders of magnitudes larger than Chess meant that machines could never 'brute force' it the way that they could with computer Chess.
And while IBM's deepblue conquered Chess back in the 1990's with its Pentium Genius...
(
)
...it wasn't until nearly another two decades later that the researchers at Google applied deep machine learning techniques coupled with other algorithms to create the world's first superhuman Go AI....
Google never released the source code nor the trained neural network to any of its "AlphaGo" versions but a while later Facebook did recreate the project based upon Google's research paper and donated the source code and its trained neural net to the world:
But by this time, the Leela Zero community Go project had already established the first open source public distribution Go project which had already achieved superhuman status... many Go players including myself helped out the project by donating our computer /GPU processing power...
Here is a match of Leela Zero against retired Go Pro (3 Dan) Haylee on Amazon AWS rented servers with 4x V100 GPUs
Even though Leela Zero already achieved superhuman status, even on just a modern desktop PC with average GPU, it was very bad at handicap games especially against top human players or even average players at high handicaps...
Later KataGo comes onto the public scene and finds a way to train the network much more efficiently as well as adapt to variable komi and high handicaps!
Now on decent hardware, such as renting the 8x V100 GPU server on AWS its possible to beat top Go professionals by giving them up to four handicap stones!
This is amazing considering the fact that just as recent as 2014 many in the computer science community thought it would be at least another decade before GO AI was able to beat a top human player at even/fair play.... in fact back then a lot of folks in the Go community thought it would be impossible for computers to ever beat humans at the intuitive game of Go, see Wired article:
Back in 2014, the then top of the line state of the art computer Go AI at the time called CrazyStone had to GET four stones from a professional Go player in order to have any hopes of scoring a win, now the tables have turned and the freely available KataGo that anyone can download from github and run on their desktop can GIVE a Go professional four stones and still defeat the human pro!
And finally for a flashback to the past, more than 20 years again when I was just a little kid our family bought our first Pentium 1 computer running Windows 95 I recall my Dad bought home a "Go" game on a floppy disk that I later very much enjoyed playing... This small 100kb Go program was called "Handtalk" and was even before I had heard of "Many Faces of Go".
handTalk is a purely DOS program but even today I can still play it on my Windows 10 by running it in "DOSBOX" or using Virtualbox and/or VMware Workstation and loading a DOS or Windows 95 VM.
Handtalk is too dated and doesn't support the GTP API, so I handfeed the games manually from it and play against KataGo, it loses again KataGo across the board even when giving it the maximum of nine handicap stones!
lol
In this new age of COVID in which everyone in the world is encouraged to "stay at home" and Netflix has complained about streaming bandwidth bottle-necking the Internet and even PC gaming platforms like Steam now throttling downloads of gaming updates etc I thought it was good to start a thread about this ancient 3,500+ year old game of Go in hopes that some may take it up as an interest and derive some enjoyment from it.
To start, here is a free "public domain" book around Go that is hosted on the Internet Archive:
To me Go is a flexible and intrinsically rewarding game in that like Chess it stood the test of time, and can be a life long hobby that one could "dive into" and "get lost in" etc..... The rules to Go are simpler than that of Chess, a five year old can learn the basic Go rules in about five minutes, but the depth to Go can be potentially far more vast than Chess and certainly the fact that its 'search space' is orders of magnitudes larger than Chess meant that machines could never 'brute force' it the way that they could with computer Chess.
And while IBM's deepblue conquered Chess back in the 1990's with its Pentium Genius...
(
...it wasn't until nearly another two decades later that the researchers at Google applied deep machine learning techniques coupled with other algorithms to create the world's first superhuman Go AI....
Google never released the source code nor the trained neural network to any of its "AlphaGo" versions but a while later Facebook did recreate the project based upon Google's research paper and donated the source code and its trained neural net to the world:
But by this time, the Leela Zero community Go project had already established the first open source public distribution Go project which had already achieved superhuman status... many Go players including myself helped out the project by donating our computer /GPU processing power...
Here is a match of Leela Zero against retired Go Pro (3 Dan) Haylee on Amazon AWS rented servers with 4x V100 GPUs
Even though Leela Zero already achieved superhuman status, even on just a modern desktop PC with average GPU, it was very bad at handicap games especially against top human players or even average players at high handicaps...
Later KataGo comes onto the public scene and finds a way to train the network much more efficiently as well as adapt to variable komi and high handicaps!
Now on decent hardware, such as renting the 8x V100 GPU server on AWS its possible to beat top Go professionals by giving them up to four handicap stones!
This is amazing considering the fact that just as recent as 2014 many in the computer science community thought it would be at least another decade before GO AI was able to beat a top human player at even/fair play.... in fact back then a lot of folks in the Go community thought it would be impossible for computers to ever beat humans at the intuitive game of Go, see Wired article:
Back in 2014, the then top of the line state of the art computer Go AI at the time called CrazyStone had to GET four stones from a professional Go player in order to have any hopes of scoring a win, now the tables have turned and the freely available KataGo that anyone can download from github and run on their desktop can GIVE a Go professional four stones and still defeat the human pro!
And finally for a flashback to the past, more than 20 years again when I was just a little kid our family bought our first Pentium 1 computer running Windows 95 I recall my Dad bought home a "Go" game on a floppy disk that I later very much enjoyed playing... This small 100kb Go program was called "Handtalk" and was even before I had heard of "Many Faces of Go".
handTalk is a purely DOS program but even today I can still play it on my Windows 10 by running it in "DOSBOX" or using Virtualbox and/or VMware Workstation and loading a DOS or Windows 95 VM.
Handtalk is too dated and doesn't support the GTP API, so I handfeed the games manually from it and play against KataGo, it loses again KataGo across the board even when giving it the maximum of nine handicap stones!
lol