Swim vs Alphago: i am as big a fan of Alfredgonefishing as anyone; my nicknames for her are affectionate, like the nicknames English schoolchildren give each other. btw, if you still don't believe me about all this, have a look at my
work history and
books and videos; if that doesn't convince you i'm not barking mad, or am!..., nothing will.
Alphagoat et al perform state-space search which creates a projected move tree, but icGo's Swim really does 'think' and can tell you what she's thinking in plain English. Swim might even give Miss A a run for her money and make you rich beyond Croesus when Google buys your company! (i will probably be dead by then, because 10 years ago i was given 5 years to live by a lung specialist, so i'm already on borrowed time).
Swim's planning methods don't have blind spots like A does, despite her being top of the tree. It will take the pros a while, but if they get the chance to play with her a lot more, they may start to find out where they are (it will require them to play moves that are not overly risky and be willing to sacrifice for positional value, because A has no idea what that is).
Any bot can Monte-Carlo; A's dcnn is better than old Zen's large patterns, which were better than the rest's smaller ones, but Swim looks at the whole board.
We know that A can look ahead more than 50 moves (indeed, all Monte bots read up to 361 moves deep), giving her the edge over human pros in tactical combat - but any bot, including Swim, can do that too. What matters is not how much you read, but what you read, and that's where Swim could have the edge over A.
There's a lot of bulldust being sprayed around about A's inventive moves and new joseki and blah blah spin blah - but the reason those unorthodox yet tactically sound moves have worked so far is because she can look deeper and wider than even Ke Jie, The only sane voice on the subject i have heard so far is Michael Redmond's in his new Redmond's Reviews series - from his various comments and analyses, i think Michael can see what i can, which no-one else seems able or willing to say in public for fear of being ridiculed by the mob because A is top dog.
Alfa is a whole other story, and one that science is just starting to scratch the surface of (and one which intrigues me greatly, because like just about everyone, i never anticipated dcnn+monte could be so powerful at Go. Separately, they aren't all that much better than GnuGo; Zen did get to 6D, for example, but couldn't reach pro level even with parallelisation.
But together, dcnn and monte are astonishing. the DARPA fellow does a pretty good job of explaining dcnn, but that doesn't explain why the combination of dcnn and monte works so well.
To learn how Alfadog thinks, follow the following procedure:
Code:
1. pick a pic of someone or something you're curious about
2. right click, copy link
3. Google search for "Google image search"; click the top link
4. mouse to its search box. left click. paste.
5. Enter
and then you will know as much about them or it as Alfadog does about Go - honestly!
Maybe, that such a mindless technique as autonomic pattern reflexes + throwing dice works so well at Go, tells us more about the mindfulness of Go than it does about the mindfulness of dcnn. Maybe Go isn't so different from chess after all, despite what we used to think 40 years ago.
Be that as it may, Go is still a captivating pasttime and we can read into it all the poetry in the world.
Announcementpnprog, author of
GoReview Partner, has published a python widget called
Gomap that performs stage I of icGo's 3-stage colour-map algorithm.
You give it an sgf of a game record without variations and it will iteratively draw the map on a board image, one step at a time, as you repeatedly press its 'Color' button.
You can step through your game with the 'prev' and 'next' buttons, and redraw the gomap at any move.
'Export' button produces a .ps file of the current frame. You can assemble a sequence of frames into an .mp4 movie and convert it to a .gif
Here is the gomap for the position after move 116 in the Nick vs Andrew game discussed in Mig34:
Attachment:
lastframegm116.png [ 274.71 KiB | Viewed 11927 times ]
There is an animation of the iteration that produces the map, which i made with Simple Screen Recorder, at
https://www.youtube.com/watch?v=BoKi2rw ... S&index=30The blobs are "colour-controlled" points - they are neither territory nor influence - the map is just the first of several icGo stepping-stones towards identifying those.
But even this preliminary map might be helpful to beginners, as it shows how stones are connected or not. You may be surprised to see just how much difference one stone in the right place can make!
Gomap installation instructions:
1. Equip your machine with a python interpreter and the Ktinker package
2. Download Gomap.zip from
GitHub.
3. unzip and run (pun intended) from
within your gomap-master folder.
On Linux, i created the following gomap.sh script which i plug into an icon on a Xubuntu panel, so running gomap is a one-click operation:
Code:
cd /home/d/go/gomap/gomap-master
python gomap.py
To run it under Windows or other o/s, you'll have to ask pnprog