Teach­ing ma­chines the art of war

Wangaratta Chronicle - North East Regional Extra - - Northeast Jobs - WITH CHRIS FEBVRE,

THE war be­tween man and ma­chine is com­ing.

In a timely co­in­ci­dence, James Cameron’s clas­sic film ‘Ter­mi­na­tor 2: Judg­ment Day’ has been re-re­leased to the­atres in 3D just as Bliz­zard En­ter­tain­ment and Google’s Deep­Mind have an­nounced the re­lease of their ap­pli­ca­tion pro­gram­ming in­ter­face (API) that will give ar­ti­fi­cial in­tel­li­gence (AI) re­searchers to abil­ity to cre­ate ma­chine learn­ing bots ca­pa­ble of tack­ling Bliz­zard’s pop­u­lar es­ports ti­tle - Star­craft 2.

Ear­lier this year, Deep­Mind’s AI pro­gram Al­phaGo eas­ily de­feated world cham­pion ‘Go’ player Lee Sedol, dis­play­ing the al­ready im­pres­sive abil­ity of ar­ti­fi­cial in­tel­li­gence to master the com­plex strat­egy game in­vented in an­cient China.

Now Deep­Mind have their sights set on a far, far greater chal­lenge, which is an ex­cit­ing prospect, even if it does in­voke trou­bling com­par­isons to Ter­mi­na­tor 2’s om­nipresent vil­lain ‘Skynet’.

For those un­fa­mil­iar with Bliz­zard’s hit video game se­ries, Star­craft 2 is a sci­ence fic­tion mil­i­tary strat­egy game.

And just like Go, it in­volves at least two op­pos­ing play­ers vy­ing for con­trol of ter­ri­tory in or­der to de­feat their op­po­nent - but that is where the sim­i­lar­i­ties end.

For the pur­poses of ar­ti­fi­cial in­tel­li­gence train­ing, Go is a so-called ‘per­fect in­for­ma­tion’ game, which means you can see all your op­po­nents moves, and those moves are fi­nite, even if the num­ber of po­ten­tial moves is one to the power of 170.

With the right hard­ware, it’s pos­si­ble to teach a com­puter all the pos­si­ble moves given enough time.

Star­craft 2, on the other hand, is in­fin­itely more com­plex.

For one thing, ‘moves’ hap­pen si­mul­ta­ne­ously in real time.

For an­other, a player can’t see their op­po­nent’s moves un­less they move a scout into their ter­ri­tory, and even then, in­ter­pre­ta­tion of those moves is re­quired.

Lastly, the game pieces, or ‘units’, each have spe­cific strengths, weak­nesses and abil­i­ties, and those strengths and weak­nesses are aug­mented by the ter­rain they are oc­cu­py­ing com­pared to op­pos­ing units.

In essence, Star­craft 2 con­sists of al­most all the com­plex­ity a com­puter would face if it were a real mil­i­tary strate­gist com­mand­ing forces in a real bat­tle.

The sheer amount of pos­si­bil­i­ties, moves, and coun­ter­moves is dizzy­ing to com­pre­hend for the pur­pose of cre­at­ing a com­puter pro­gram ca­pa­ble of learn­ing how to play.

Which is why Deep­Mind and other AI re­searchers aren’t ex­pect­ing that a com­puter will be able to de­feat a hu­man player for at least five years.

It also means that re­searchers can’t sim­ply place the AI in a real, full-fledged game of Star­craft from the out­set and ex­pect it to learn ef­fi­ciently, if at all.

In­stead, just like a hu­man player might be­gin to learn to play, Deep­Mind and Bliz­zard have bro­ken the game down into mini-games in­volv­ing in­di­vid­ual tasks and as­pects of a full match.

To be­gin with, the AI bots are sim­ply be­ing tasked with col­lect­ing re­sources on the map re­quired to build units, or to move a unit to a par­tic­u­lar point on the map.

So the good news is we’re a while away from need­ing to worry about Google in­ad­ver­tently unleashing an army of Ter­mi­na­tors on us all.

And the other good news is that if ar­ti­fi­cial in­tel­li­gence be­comes good enough to de­feat hu­man play­ers in Star­craft 2, it will mean we’ve reached a point in ma­chine learn­ing that will open up ex­cit­ing pos­si­bil­i­ties and ap­pli­ca­tions be­yond merely play­ing strat­egy games.

In the words of Star­craft 2 char­ac­ter Ty­chus Find­ley, ‘Hell, it’s about time.’

JUDGE­MENT DAY: Deep­Mind and Bliz­zard En­ter­tain­ment has re­leased pro­gram­ming tools to teach ma­chine learn­ing bots to play Star­craft 2.

Newspapers in English

Newspapers from Australia

© PressReader. All rights reserved.