Commonsense Go

[djhbrown]

it's taken me a while, but the dawn is slowly beginning to break - not only is AGF a religious fundamentalist organisation, but almost all of my opponents these days have no commonsense whatsoever - which is probably why they play Go, since no-one with any commonsense would do so.

never mind, little things....

here's the latest rewrite of the Introduction, not uploaded yet:

Introduction

Go teachers use a combination of language and particular moves to explain general concepts, from which students can form their own mental images, but they still cannot see clearly what the teacher sees, because a lot of the knowledge an expert has is either tacit (subconscious) or too elaborate to explain in language alone. So they often resort to the Occam's Razor of asking students to choose between 2 or 3 moves (“would you play A, B or C?”).

But where do these choices come from? They come from inside the teacher's head, not from inside the student's head.

Maybe 85% of the thoughts we think are subconscious (Damasio, 20..), so even the most empathetic and open-minded teacher cannot explain why they think what they think, because by definition the subconscious is inaccessible to the conscious.

There is thus much to be gained from endowing a machine with the ability to form and use perceptions that can be explained by visual images and narratives, so that such knowledge can be transmitted to new generations.

In principle, human teachers could be replaced by machines.- but candidate move generation is non-trivial. Contemporary master-level computer go programs such as Alphago and JueYi utilise brute-force kneejerk reaction search, albeit reactions of learned convolutional patterns to reduce the search space. This makes them impressively powerful players - better than the best humans - but their machinations are more alchemy than chemsitry, and there will need to be substantial developments of artificial neural network architectures before they can even come close to assembling a coherent thought, let alone express it.

The parallels between artificial and natural neurons run no deeper than the parallels between any kind of neuron and a transistor (Didales, 2013) - they all perform the same basic computational function of modus ponens - in the neuron's case, moderated by the principle that the louder and more frequently you shout, the more i am inclined to believe you, regardless of whether you have the faintest idea what you are talking about. Neural nets are democracy in action: the blind leading the blind.

But the science of Artificial Intelligence has more to offer than brute force - it offers the rationality of logic. Logic too is predicated upon the principle of modus ponens, which is hardly surprising, for modus ponens is the fundamental computational mechanism upon which all computational operations - such as addition, diagnosis and prognosis - are based.

The great difference between primitive computers such as Alphago -as reflective as her namesake Alf Garnet - and sophisticated thinkers like SHRDLU (Winograd, ) and Swim (Brown, ), is that the latter operate upon conceptual structures that embody aggregate information, not just mere pixels.

"Is this a dagger i see before me?" asks Macbeth. After a few million trials and errors, Alfie could answer yes or no, but she cannot learn to draw a line around the dagger., because she doesn't know in which part of the picture the dagger is. Her convolutions quite literally convolute the real-world structure depicted by an image into a convoluted mess, good for telling A from B, but not where it is.

Swim (= See what i mean) is a software model of Go commonsense, able to explain her thinking in plain English. She is described in the context of several examples:

1. a tactical problem presented by Jennie Shin of Guo Juan's Internet go school.
2. figuring out a defence to Lee Sedol's magic wedge in game 4 of his match with Alphago.
3. providing a rationale for Alphago’s move 37 in game 2 against Fan Hui
4. finding a move for Alphago that combines moyo expansion and reduction
5. finding a moyo invasion for Jue Yi that offers two ways to succeed
6. finding a move for Andrew to grind Nick down yet further
7. finding a move to rescue Kirby from drowning.

icGo is a smart online/offline playing interface / editor / advisor / player based on Swim plus a bevy of bots like Leela that serve as a jury of peers to offer a second opinion.

[djhbrown]

https://papers.ssrn.com/sol3/papers.cfm ... id=3027817

includes some debugged algorithms, some new fightback methods, and Swim's suggested move for Kirby

[djhbrown]

Learning to Swim: Mechanisms are described by which a model of conceptual reasoning about Go can learn new techniques from its own analyses of expert moves and assimilate expert advice.
https://papers.ssrn.com/sol3/papers.cfm ... id=3071677

[djhbrown]

Just when you think you've got it figured out, they come along and make a fool out of you

i should have known that the elegance and simplicity of the colour and shadow map algorithms was too good to be true.

bordbeforeproblem.png (18.21 KiB) Viewed 11880 times

Here, three issues are revealed:
1. E15 shouldn't be a black colour-controlled point, because D15 is in atari, so shouldn't throw out a coloured arm. That's an easy fix.
2. the shape of black's big shadow doesn't look right; it's too square; it should have more of a curved shape, like the grey line for example.
3. D14 is alive, because C16 is dead, so D14 should throw an arrow into black's shadow, something like the white shape. And it should do this, regardless of whether E15 is coloured or not.

Any ideas?

[djhbrown]

talking to yourself is the first sign of madness, but i'm well past that, so more talking to myself won't make any difference.

the size and strength of a cluster's shadow is clearly a function of its thickness. An unconditionally alive cluster is infinitely thick, but although that doesn't mean it casts an infinite shadow, it does mean it casts a bigger one than a weak cluster.

as unconditional life is rare, except towards the end of the game, one has to work with one's best assessment of life and death.

Swim sees obvious life and death (lad) in a simple yet logical static way, and that's good enough to start with, although clearly it is desirable to continuously reassess lad through dynamic analysis.

Operations Research (OR) substitutes data for knowledge, and quantity for quality. This is appropriate for computing the best elevation and direction in which to point a cannon, but with the notable exceptions of OR machines like Alfie, AI might be more I if it goes the route of qualitative reasoning - that's what evolution came up with, and evolution has had plenty of time to think about it, so it's probably a good idea.

The magic number 7 plus or minus 2 might suggest that 7 degrees of thickness is appropriate, but i feel that 4 is probably sufficient for practical purposes:

dead, weak, strong, and alive.

A dead cluster casts no shadow, even if, like a narcissistic poster, it can leave a bad taste in the mouth.

An alive cluster surely casts a shadow at least as big as itself, and a merely strong one a smaller shadow. This could be approximated by varying the conditions under which the shadow of an alive or strong group propagates itself: 3 for a strong cluster, but just 2 for an alive one.

A weak cluster is little better than a dead one, so the cautious player would not be too optimistic about a weak group casting any shadow at all.

Obvious lad has already been defined, so it only remains to differentiate weak from strong [group strength is already defined, but this is about cluster strength].

One eye is a bit better than none, but one swallow does not a summer make, so maybe dangosity (stones/size) is a sounder way of assessing non-alive strength qualitatively: heavy clusters are weaker than light ones.

Dangosity is a quantitative measure, but it can be sieved through a threshold to generate a qualitative value.

0.5 (= as much eyespace as stones) sounds about right.

[djhbrown]

whereas Hofstadter's snowflake theory of light works for light, Go shadows are different from light shadows, because their strength falls off with distance, a bit like the way gravity does.

So recursive shadow propagation should take that into account.

If the inverse square law is good enough for Newton, maybe it's good enough for Swim too.

However, a Go board isn't a continuum like space, but a quantised chequerboard (without the Harlequin colours), so square roots of square steps doesn't make much sense in that context.

An alive cluster has much greater gravity than its first fringe of shadows, so the lowered threshold of 2 should only apply to the first iteration; and 3 to subsequent ones. This may yield shadows with a more curved than square perimeter, so they would at least look prettier.

[djhbrown]

A long time ago, in a faraway place, i was sitting on the bank of a lake in France trying to undo a seized half-hitch. Unless i undid it, i wouldn't be able to move the boom up the mast, and unless i did that, i wouldn't be able to ride the borrowed windsurfer. And we had gone there especially to windsurf.

If you have ever done any sailing, you may know just how seized a knot can become, and there i was without my trusty rusty stainless steel marlin spike. It took ages and ages of wrenching this way and that, to tease the threads of the rope a fraction of a millimetre at a time.

After what felt like forever, it came apart. The friend of a friend sitting with me said "You are tetu". I didn't know what that word meant, but her tone and eyes were positive, so feeling rather pleased with myself i figured "tetu" meant "tenacious".

Later, i asked my friend what "tetu" meant. "Stubborn", she replied !

I must be stubborn, to keep on trying to explain the bleeding obvious to what feels like a seized-up brick wall of religious fanatics that their bovine god isn't the be all and end all - but there you Go - what else is there to do? And besides, i have to acknowledge that despite all my efforts i still haven't explained it adequately, because if i had, there wouldn't be all this negativity.


Swim vs Alfie

In 19nn, alien intelligences conquered Amsterdam [1] creating a flurry of excitement that spawned a forest of clones and progeny, one of whose grandchildren [2] has conquered the world, leaving old sages like Ke Jie nursing their bruises whilst the new turks dance in the spotlight of an adoring goggling mass media to the delight of Wall Street.

For sure, it's game over for Gomans, but it's only game just beginning for AIGo, for even as the New Emperor shines in the sun, the seed of a contender lurks in the dusty corner of an old backroom [3]. Few that have glimpsed it have been able to see the wood for the trees, for to those goggling at the blinding light of saturation press, the upstart just looks an old idea in new clothes that never got anywhere when it was first tried.

But Swim is not just old; it's positively geriatric, dating back about 2000 years to a bloke called Aristotle who wrote about syllogisms.

The acid test of science is experiment. Alfie has demonstrated beyond doubt that of all the bots that virtually exist, she is simply the best, in chess as well as Go. She might even be good enough to tell the difference between a cat and a child, to decide whether it's worth the risk of crashing the car by swerving violently, something that a London driver (who in their right mind drives a car in London??) studying a Go board couldn't do because he isn't looking where he is Going, which is downhill all the way, because that's the Second Law of Thermodynamics.

What kind of BorD are you?
https://www.lifein19x19.com/viewtopic.p ... 23#p227223

1. Monte-Carlo Alien Intelligence Conquers Amsterdam. https://sites.google.com/site/djhbrown2 ... yStone.doc
2. https://www.nature.com/articles/nature24270
3. https://papers.ssrn.com/sol3/papers.cfm ... id=3071677