lobotommy wrote:
The point of the SWIM thing proposed by djhbrown is to provide explanation for us, humans, in natural language, using our heuristics for description of situation on the board. Good enough for me.
:) here we go... actually, as it happens, as far as i am concerned, that's not
my point of the Swim thing at all and never was! - and i've said so before in this thread too....
i started my journey on this topic back in 1971, when i set out to see if i could program a computer to learn a language. i figured that children begin by learning single word associations, and then move on to Chomsky type 3, then 1, skipping 2 because people dont think that way, except some Bishop in 18something or other, who caused all the trouble for schoolkids today because he misunderstood Artistotle. Specifically, a Sentence in language (as opposed to a logical statement) is NOT composed of a Subject and a Predicate (and, as Boole showed, neither is a logical statement, so Aristotle was wrong too!!).
type 0 sounds meaningless to me as far as human language is concerned; there has to be some kind of syntax - but let me tell you this: the English grammar you learned in school is completely wrong!! and i proved it, following in the footsteps of Chomsky who proved it long before when i was still in short pants - >>>>>>>>> but, thanks to a student at Shenzhen University in 2001, who asked me to explain prepositions, i discovered the syntactic form of Chomsky's hypothesised "universal grammar", which will make me famous long after i'm dead because no-one knows about it yet (not even Chomsky), despite it being there for all to see in black and white [https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2205530]
<<<<<<<<<< back in 1971, i started on the linear language of Contract Bridge-bidding, but that was too tough (for me to program a computer to learn properly), so i ended up producing a program that could learn
any language
Or, to be more precise (and restricted), any
linear language.
Such as the language of differential diagnosis of liver disease. I was disappointed when it only performed at 85% accuracy, but when i expressed this disappointment to a Professor of Internal Medicine, he said "Oh, that's as much as we get, anyway".
My program also learned to bid at Contract Bridge, but only as well as a weak amateur. I privately concluded from this that Bridge was more complicated than medicine
But i was still dissatisfied, because i realised that although my (nameless) concept-learning program could learn "how" to do something, it couldn't learn "why" it should do it.
[Reasoning About Games. Proc European Conference on Artificial Intelligence, Hamburg, 1978. Reprinted in AISB Newsletter 32, 14-17, 1978.]
I saw this as a significant defect, and started a new journey, which was quickly interrupted by a need to make a living, and didn't come back to it for 37 years, not until 2015, after having been ethnically cleansed and put out to grass, and having solved the riddle of who invented Christmas, and why, quicker than i would have imagined possible (it turned out to be a much straighter lineage than i had expected, just by lots of Googling, and with nothing better to do, because i was by then a rentier capitalist, living (modestly but comfortably) off the interest alone, and without the avarice of Morgan, found my old interest in Go resurface.
By then i had started to enjoy making movies, and no longer needed peer review to keep my job, as i didn't have one, so could freely indulge my sense of humour and love of exploration at the same time.
My quest was - and still is - and i am finally getting around to the point - to find a way for a machine to
understand Go, not as an end in itself, but as a stepping-stone to understanding understanding in general (and yes, i did proofread that and do mean two "understanding" words; the first is the act (relation) and the second is the object (concept). C -> RC )
First came HaLY, then HoLY, then CG, then Swim, and now icGo.
Moreover - and this point is key - i do
not try to make it think like a human - on the contrary, i try to make it think about the fundamental nature of Go, based upon the objective and mind-independent nature of the semeotic relationships between stones and the rules of the game. Whenever something Swim thinks happens to coincide with a Go proverb (such as Andrew's 5-space jump, for example), i celebrate the fact, but my idea of AI concurs with what i believe to be the view of its prophets McCarthy and Minsky, who say that their objective for AI is not to replicate
human intelligence, but to discover the essence of intelligence in general.
As it happens, that's what Demis says he's up to as well, but frankly i am unconvinced that he and all the other neural net nutters are going in the right direction - i agree with Chomsky. btw, the epithet "nutters" is not deprecatory, it merely means people who are single-mindedly fascinated by something. Eg, Einstein and Feynman were both particle nutters; Debussy was an impressionist nutter; i (and, i like to think, Feynman) am a hierarchy nutter, etc, etc.
The aim of icGo is twofold:
1. to help people
2. to beat the pants off Alphago!
Now look here, let's get serious for a moment; i am as big a fan of Alfie as anyone, but i see a hole in her, the very same hole that was in my concept learning program all those years ago.
She knows "how", but she doesn't know "why" - and because of that, she can drive herself off a cliff, as she did in game 4.
DM found a patch for that, and have improved her by using "anti-Alpha", so she is further up the beanstalk than this time last year, which is why Ke Jie is toast.
But she still can't see eyes!!
So, it's not that Swim can explain herself, it's that she
can see eyes and etc.
that might give her the edge over A's dcnn - one evidence for which is that dcnn without Monte is almost as hopeless across the board as i am. btw, GnuGo and all that lot tried to see eyes, but they couldn't see potential eyes well enough to avoid having needles stuck in them and being cut into pieces and being dumped on their backsides by even weak players like me.
and Swim can use googleplexes of cpus just as easily as the doggie in the windows, so A's comparative advantage is nothing!
Even without them, even with no Monte at all, CG found a better move than Alfie - i'm convinced J13 works, because it's kikashi, and then Swim can play Myungwan Kim's move at L10, which Swim found all by herself, with no help from Kim, who himself didn't notice it until Haylee suggested it in the middle of a different sequence she was exploring, when they were commenting on the game, before Lee made his move. That's why they were so disappointed when Lee made the wedge, because unlike all the other commentators - and Alfie! - they knew it didn't work (shouldn't have worked). Had Swim been watching over Lee's shoulder, she would have whispered in his ear not to do that silly thing that Demis even today still calls a godlike move because it beat his own baby up.
See what i mean?
Of course, every proud father sees his own child through rose-tinted glasses (until it becomes a teenager) so my view of Swim's prowess potential may well be exaggerated, but i challenge you all to this:
My God exists unless you can prove otherwise! - which you can easily do by programming her and showing that she falls over her own feet.
PS it feels good to get things off one's chest, doesn't it