Ro­bots that learn and share new lan­guage

New Straits Times - - Bots - Daily Mail

Re­searchers at OpenAI, which is the ar­ti­fi­cial in­tel­li­gence lab cre­ated by Tesla founder Elon Musk, have taught ro­bots to use lan­guage to com­plete sim­ple tasks. RO­BOTS have com­mu­ni­cated with each other us­ing a shared lan­guage that they made up as they went along in an amaz­ing new ex­per­i­ment.

Ex­perts at OpenAI, an ar­ti­fi­cial in­tel­li­gence lab cre­ated by Tesla founder Elon Musk based in San Fran­cisco, have taught ro­bots to start their own lan­guage so they could help each other com­plete sim­ple tasks.

In the fu­ture, re­searchers hope that as the lan­guage be­comes more com­plex, they will have a “trans­la­tor bot” that will be able to con­vert robot chat into English.

The team cre­ated a world where ro­bots learnt a lan­guage that en­abled them to com­plete sim­ple tasks such as mov­ing them­selves from A to B.

Ex­perts cre­ated a two-di­men­sional world to nav­i­gate and the ro­bots — or agents — had to find their way around.

The ro­bots re­lied on a tech­nique called re­in­force­ment learn­ing which in­volved them us­ing dif­fer­ent sounds.

They taught agents to cre­ate lan­guage by drop­ping them into a set of sim­ple worlds, giv­ing them the ability to com­mu­ni­cate, and then giv­ing them goals that can be best achieved by com­mu­ni­cat­ing with other agent.

“If they achieve a goal, then they get re­warded,” said the team.

“With re­in­force­ment learn­ing, they de­vel­oped a shared lan­guage to help them achieve their goals”, they said.

For ex­am­ple, a speaker would start to form an as­so­ci­a­tion be­tween the word “house” and the im­age of a house.

As they build up their vo­cab­u­lary, they can start us­ing sen­tences that al­low them to con­vey ideas to one an­other.

As the lan­guage be­comes more com­plex, ex­perts hope they will be able to trans­late what they are say­ing into English.

The team gave the ro­bots two types of ac­tions, one en­vi­ron­ment ac­tion such as look­ing at some­thing; and com­mu­ni­ca­tion ac­tions, such as broad­cast­ing a word to one of the other agents.

Ro­bots were able to mod­ify their mes­sages to im­prove the ef­fec­tive­ness of com­mu­ni­ca­tion.

For ex­am­ple, if an agent re­alises it could have per­formed some­thing bet­ter had it had sup­ple­men­tary in­for­ma­tion, he will com­mu­ni­cate this in­for­ma­tion to the next agent so its in­struc­tions are as clear as pos­si­ble.

The team found that agents were able to evolve lan­guages to fit the sit­u­a­tion and able to ac­com­plish tasks by com­mu­ni­cat­ing with one an­other.

“We hope that this re­search into grow­ing a lan­guage will let us de­velop ma­chines that have their own lan­guage tied to their own lived ex­pe­ri­ence,” the re­searchers wrote.

“We think that if we slowly increase the com­plex­ity of their en­vi­ron­ment, and the range of ac­tions the agents them­selves are al­lowed to take, it’s pos­si­ble they’ll cre­ate an ex­pres­sive lan­guage which con­tains con­cepts be­yond the ba­sic verbs and nouns.”

Newspapers in English

Newspapers from Malaysia

© PressReader. All rights reserved.