Computers learn 'regret'
Feb 29,2008 00:00 by Bend_Weekly_News_Sources

Sci­en­tists in Italy have de­vel­oped com­put­er pro­grams that mim­ic hu­man decision-making by us­ing sim­u­lat­ed “re­gret” to im­prove per­for­mance. The mod­els do a bet­ter job than oth­ers in pre­dict­ing some as­pects of hu­man decision-making, the re­search­ers report.

The stu­dy’s bas­ic as­sump­tion was that peo­ple mod­i­fy their be­hav­ior dur­ing stra­te­gic games by look­ing back­ward to what might have been their best move, once they know what the oth­er play­ers’ move was. 

Da­vide Mar­chiori of the Uni­ver­s­ity of Tren­to and Mas­si­mo War­glien of Ca’ Fos­cari Uni­ver­s­ity in Ven­ice built math­e­mat­i­cal mod­els based on bi­o­log­i­cal neu­ral net­works. These use sim­u­lat­ed net­works of “brain cells” to ar­rive at de­ci­sions and learn by tri­al and er­ror.

In­tro­duc­ing an ap­proxima­t­ion of re­gret al­lowed the mod­els to pre­dict hu­man be­hav­ior more pre­cisely than con­ven­tion­al eco­nom­ic learn­ing the­o­ries, the re­search­ers said. Their find­ings ap­pear in the Feb. 22 is­sue of the re­search jour­nal Sci­ence.

“Re­gret refers to the dif­fer­ence be­tween out­comes at­tained and the best out­comes that might have been at­tained,” wrote Mi­chael D. Co­hen of the Uni­ver­s­ity of Mich­i­gan, Ann Ar­bor, in a com­men­tary in the jour­nal. “This is an im­por­tant step in the de­vel­op­ment of a work­a­ble new syn­the­sis,” added Co­hen, who was­n’t in­volved in the stu­dy. The work has ap­plica­t­ions in de­vel­op­ment of eco­nom­ic the­o­ries that pre­dict hu­man be­hav­ior, he added.

The mod­el’s pre­dictions, he con­tin­ued, aren’t based on “con­ven­tion­al, forward-look­ing ex­pecta­t­ions of gain, the no­tion so long at the heart of eco­nom­ic the­o­riz­ing.” Rath­er, its pre­dictions rely on “propens­i­ties that de­vel­op through a back­ward-look­ing learn­ing pro­cess that is driv­en by re­gret.”

Courtesy Science and World Science staff