No go for com­put­ers

Ben Macin­tyre finds there’s one game at which ar­ti­fi­cial intelligence can’t beat hu­mans

The Weekend Australian - Review - - Feature -

TEN years ago in May, to the dis­may of many chess en­thu­si­asts, the IBM su­per­com­puter pro­gram Deep Blue beat world chess cham­pion Garry Kas­parov: the great­est chess mind alive was el­bowed aside by raw com­put­ing mus­cle. The qual­ity of Deep Blue’s vic­tory is still de­bated, but the mo­ment marked a turn­ing point in the re­la­tion­ship be­tween man and ma­chine.

The com­puter is now dom­i­nant in al­most ev­ery board and card game de­vised by hu­mankind. Com­put­ers can beat us not only at chess but also draughts, Othello ( also known as Re­versi), Scrab­ble, three- di­men­sional noughts and crosses, Mo­nop­oly, and even bridge and poker ( most of the time).

In th­ese games, the com­puter has a blue­print for ‘‘ per­fect play’’: it sim­ply runs the board po­si­tion through a data­bank and chooses the best next move, ev­ery time.

The un­stop­pable march of com­puter power has long been a sta­ple of science fiction, the night­mare evoked in 2001: A Space Odyssey and The Ma­trix , in which ar­ti­fi­cial intelligence seeks to con­trol and su­per­sede hu­mankind’s. If com­put­ers can win at the in­tel­lec­tual chal­lenge of world- class chess, it is as­sumed, then the com­put­erised brain, for good or ill, must be in­evitable.

Yet there is one game in which the com­puter is still no match for hu­mans, a game in which a com­pe­tent teenager can beat the world’s most so­phis­ti­cated com­puter pro­gram with ease, and that is the an­cient Chi­nese board game Go, the old­est game in the world and the only one at which man re­mains the undis­puted cham­pion.

Go orig­i­nated in an­cient China long be­fore there was writ­ing to record it ( the in­ven­tion, so leg­end has it, of an em­peror keen to teach his fool­ish son the virtues of bal­ance and pa­tience). The game in­volves a sim­ple grid board of 19 lines and two play­ers, one with white stones and the other with black.

The ob­ject, sim­ply put, is to stake out a larger ter­ri­tory by tac­ti­cally plac­ing the stones and sur­round­ing the op­po­nent’s forces. Go is a beau­ti­ful game to play and watch: in­tri­cate, com­plex and, for a com­puter, dis­tinctly baf­fling.

The qual­i­ties that mark out the mas­ter Go player are pre­cisely those a com­puter lacks: in­tu­ition, plan­ning, char­ac­ter and pat­tern read­ing. Go is not merely a mat­ter of prob­a­bil­i­ties lead­ing to cer­tain­ties, which is how Deep Blue and sim­i­lar pro­grams work.

At its best, the game re­flects the defin­ing char­ac­ter­is­tics of hu­man intelligence. De­spite the ex­pen­di­ture of mil­lions of dol­lars, the prom­ise of prize money and sev­eral decades of re­search, at­tempts to cre­ate a pro­gram to chal­lenge pro­fes­sional Go play­ers have made lit­tle head­way. Even the most so­phis­ti­cated pro­grams can com­pete with ex­pe­ri­enced play­ers only on a re­duced grid.

Go is seen as a key to un­lock­ing the se­cret of ar­ti­fi­cial intelligence. If com­put­ers can learn the game, some sci­en­tists be­lieve, mankind would be a big step closer to repli­cat­ing hu­man thought pro­cesses, with great sci­en­tific ben­e­fits. Con- versely, so long as we can still beat com­put­ers at this most hu­man game, then the spec­tre of a ma­chine that can com­pre­hend rather than merely com­pute will re­main the stuff of fiction.

From China, Go spread through­out Asia, and it is in­creas­ingly played in the West. Like chess, de­spite its com­plex­ity at the high­est level, Go does not re­quire spe­cial men­tal equip­ment: Al­bert Ein­stein played it, but so does Rod Ste­wart and so do I ( very badly).

Hu­mans are rather good at Go, but com­put­ers are not. Deep Blue won by brute force com­put­ing power, work­ing out the pos­si­ble moves and counter- moves far ahead of play, at a rate of 200 mil­lion po­si­tions a sec­ond. Com­pared with the pos­si­ble per­mu­ta­tions of Go, how­ever, that was a dod­dle.

In chess, each player has on av­er­age 25 to 35 pos­si­ble moves; the av­er­age num­ber of moves in Go is more than 250. It has been cal­cu­lated that there are more dis­tinct games of Go than atoms in the known uni­verse.

In Go, a move early in the game can af­fect the pas­sage of play hun­dreds of moves later. The vast­ness of the pos­si­bil­i­ties of­fers wide scope for in­di­vid­u­al­ity, strat­egy, per­son­al­ity and in­tu­itive spa­tial aware­ness, all of which elude com­put­ers.

If chess is a me­dieval bat­tle, then Go is com­pa­ra­ble to a world war, with un­pre­dictable con­flicts and ter­ri­to­rial shifts across a wide ter­ri­tory. The hard­est task is to tell who is win­ning at any given mo­ment.

A Tai­wanese or­gan­i­sa­tion has of­fered $ 1 mil­lion for the first com­puter pro­gram to de­feat a ju­nior Go cham­pion, yet de­spite some re­cent ad­vances, none has yet reached that skill level.

Com­put­ers crunch num­bers, but hu­mans recog­nise pat­terns and at­tach val­ues to them, some­times sub­con­sciously, in the same way that we recog­nise faces.

To win at Go, a com­puter would have to be able to spot sub­tle, com­plex de­signs and draw the cor­rect in­fer­ences. The ex­pe­ri­enced Go player can tell whether a con­fig­u­ra­tion is alive, mean­ing it has been de­vel­oped in a way that leaves it im­mune to cap­ture, or dead, in the sense that no amount of counter- moves will save it: pre­cisely the sort of fuzzy con­cept that sends com­put­ers into melt­down.

Th­ese are the hall­marks of hu­man intelligence — adap­ta­tion to un­cer­tainty, in­tu­ition, wis­dom, the abil­ity to un­der­stand the thoughts and feel­ings of oth­ers, and a sense of mor­tal­ity — that com­put­ers can­not repli­cate yet.

Alan Tur­ing, the great pi­o­neer of com­puter science, once said a ma­chine could be con­sid­ered in­tel­li­gent if its re­sponses were in­dis­tin­guish­able from those of a hu­man.

Viewed from the level of an an­cient Chi­nese board game, the prospect of a ma­chine that can re­pro­duce the sub­tleties of the hu­man mind still seems as dis­tant as ever. Only when ma­chines sur­pass Go ( and col­lect $ 1 mil­lion) will ar­ti­fi­cial intelligence truly be wor­thy to com­pete with the hu­man kind.

Il­lus­tra­tion: Dave Fol­lett

Newspapers in English

Newspapers from Australia

© PressReader. All rights reserved.