CHATBOTS are being taught how to drive a hard bargain in a new AI experiment carried out by Facebook's Labs.
According to New Scientist, the research could lead to more effective personal assistants able to negotiate on our behalf, sorting out calendar clashes and the like.
French website Julie Desk is already offering this kind of AI diary management, but now Facebook has jumped on the bandwagon, looking at perhaps getting you a good deal on your next holiday, according to Mike Lewis from the social network's boffin division.
The team trained bots on a database of over 5,000 text conversations between people playing a game where they had to divvy up an inventory of "things". Each "thing" was assigned a value, with the values unique to each player and each item. So, for example, a ball might be worth four to one player, but only two to another.
The object of the game, as in most games, is to score the most points, by acquiring the most objects with the highest personal value.
After learning, the bots were further trained with more matches, some against each other, some against humans. Working in natural language often led to a crappy deal. Working in totally selfish terms often led to a great deal, but often one made in utter gobbledegook.
The trick, therefore, was to find a way of combining techniques to produce something that would allow the bots to communicate with humans in a real world scenario. The result was a good (but not brilliant) negotiator who can work with humans on their terms.
Beyond doing work for you, a bot might be able to give you useful tips when doing a deal that perhaps you don't want to hand over. Say you're negotiating a house price, it could be able to tell you how much of your hand to play and what not to say.
Oliver Lemon at Heriot-Watt University explained that the use of natural language was essential as a user would need to be able to go back to a deal and work out why it did what it did - in other words, justification is important when you're a bot.
Late last year we reported on UCLA students who had created a Judge Rinderbot. µ
Next thing, Tay will be negotating weed prices. µ
Justine Greening and Greg Clark among those affected
A whole new take on 'cord-cutters'
Surely everyone can get a long?
Report also points to an ARM coprocessor for Touch ID