Google soft­ware beats Board-Game champ in 3 straight matches

The Pak Banker - - COMPANIES/BOSS -

NEW YORK: Chalk up an­other win for ar­ti­fi­cial in­tel­li­gence. Google Deep­Mind's AI sys­tem won its match against a top-ranked player of Go, as ma­chine-learn­ing soft­ware mas­tered the in­tri­ca­cies of the 2,500-year-old strat­egy board game. The pro­gram scored its third vic­tory against Lee Sedol Satur­day, win­ning a five-match tour­na­ment. The South Korean is con­sid­ered the world's best player of Go in the past decade.

"It is a huge land­mark for Go, a long-an­tic­i­pated mo­ment in the West and -- to a greater de­gree than I imag­ined -- a shock to many peo­ple in Asia," said An­drew Okun, pres­i­dent of the Amer­i­can Go As­so­ci­a­tion.

Deep­Mind's suc­cess at Go has as­tounded ex­perts, who thought it would be five to 10 years be­fore AI could beat topranked pro­fes­sional play­ers of the game. While the rules are sim­ple -- play­ers bat­tle for ter­ri­tory by plac­ing white or black stones on a 19-by-19 grid of squares -- Go is much more com­plex than chess, by an or­der of 10 fol­lowed by 99 ze­ros.

The vic­tory po­si­tions Google as a leader in the next gen­er­a­tion of su­per-smart com­put­ing. The search gi­ant al­ready uses AI in a range of prod­ucts -- au­to­mat­i­cally writ­ing emails, rec­om­mend­ing YouTube videos, help­ing cars drive them­selves. The next wave of AI tech­nolo­gies will use tech­niques akin to those de­vel­oped by Deep­Mind, though the com­pany hasn't yet dis­closed any par­tic­u­lar prod­ucts.

Google re­vealed the Al­phaGo game-play­ing soft­ware in a sci­ence jour­nal Na­ture in Jan­uary. At that time, Google said its sys­tem had al­ready beaten pro­fes­sional Euro­pean Go player Fan Hui in matches held at its Lon­don of­fice in Oc­to­ber. Since then, the soft­ware has been play­ing thou­sands upon thou­sands of games against it­self, with Google run­ning as many as a hun­dred sep­a­rate ver­sions of the pro­gram in par­al­lel at any one time. That let the soft­ware to ac­quire ex­pe­ri­ence and knowl­edge at a rate faster than a hu­man ever could.

"I'm some­what shocked," Lee said af­ter his first loss. "I am quite speech­less," he said af­ter his se­cond loss.

De­vel­op­ment of the soft­ware does not fin­ish with this vic­tory. Deep­Mind's co-founder, Demis Hass­abis, has said the com­pany wants to de­vise a ver­sion of the al­go­rithm that re­quires less knowl­edge of the game. Deep­Mind also plans to test out the pe­cu­liar in­tel­li­gence dis­played by its soft­ware by tweak­ing the Go board, per­haps by re­mov­ing cer­tain points on the grid or chang­ing how they're con­nected, to see how Al­phaGo re­acts.

There are al­ready signs that Al­phaGo has come up with strate­gies that pro­fes­sional Go play­ers haven't pre­vi­ously con­sid­ered. Dur­ing its se­cond match, the pro­gram made a move that the game ex­perts found hard to un­der­stand. "It's play­ing moves that are def­i­nitely not usual moves," said Michael Red­mond, the game com­men­ta­tor and also a highly ranked Go player, spec­u­lat­ing it was "com­ing up with the moves on its own." All five games will be played to de­ter­mine the fi­nal match score and learn more from Lee, ac­cord­ing to a post on Google's Asi­apa­cific blog. The next game will be on Sun­day and the fi­nal on Tues­day, March 15n ac­cord­ing to the post.

Newspapers in English

Newspapers from Pakistan

© PressReader. All rights reserved.