Facebook researchers have been working on developing and teaching a pair of artificial intelligence (AI) powered agents how to negotiate.
Now, according to a recent release from the team, these moved slightly passed their expectations and seem to have developed their own ‘language’ or at least a communication system. After spotting this unexpected consequence, the developers decided to shut down this project.
This Facebook AI project saw to teaching its agents how to bargain. In this test, the AI pair had to become capable of negotiating for itself. The agents were programmed to automatically figure out the best way of splitting the bargained items into groups. This, coupled with the pair each assigning a value to an item, would ensure that there would be no drawback and a deal would be reached each time.
Facebook AI Project Agents Capable of Generating Their Own Contextual Language
Now, their capabilities seem to have evolved and surpassed their developers’ expectation. According to Facebook, the developer team noted that Bob and Alice, the AI agents pair, seem to have developed a ‘language’ of their own. Although seemingly gibberish to us, the team states that the registered conversation is an example of how the AI generated their own contextual language. One that also helped them understand conversations.
Dhruv Batra, a Georgia Tech visiting Facebook AI researcher, is quoted as saying that these AI agents were not programmed to stick to the English grammar and its sentence structure. So, to make up for it, they may have created one of their own.
“Agents will drift off understandable language and invent code words for themselves. Like if I say ‘the’ five times, you interpret that to mean I want five copies of this item. This isn’t so different from the way communities of humans create shorthands,” Batra was reported as saying.
Still, the Facebook team of developers decided to shut down this path. To do so, they re-enforced the need of using the standard English sentence structure as the AIs were negotiating in further tests. The researchers also reportedly added additional training to the negotiating agents. This taught them good versus bad behaviors.
The research team then noted that the AI agents performed ‘extremely well’ on the human-to-agent-tests that followed this reprogramming.
Image Source: Pixabay