Facebook created a chatbot that learned to negotiate like humans, on its own

Facebook has published the research and the code for the agents as well as the data source have been open sourced.

The Facebook Artificial Intelligence Research (FAIR) team has developed a chatbot that has learned to haggle like humans. The chatbot was trained to get the best possible outcome from a negotiation, through multi-issue bargaining. Two agents were presented with three items, books, hats and basketballs. Each agent was assigned a particular value for each of the items. For example, one agent may have higher points for getting books, and another may have higher points assigned for getting hats.

Just like real world negotiations, neither agent knew the points allocated to the items of the other agent. Then, the two were made to come up with a way to distribute the items among themselves. Additionally, both the bots would get zero points if there was no result after 10 rounds of dialogue. This meant that the bots were forced to negotiate a deal. When interacting with humans, the bots came up with negotiation tactics that actually worked.

The bots could hold longer conversations with the humans, and not settle unless there was a good deal. The humans chose to walk away faster than the bots. Another sneaky thing the agents learned to do was pretend to be interested in buying one item, and eventually settle for another item. This is a tactic used by humans to negotiate, and the AI learnt to to do this on its own. Although the agents were likely to repeatedly use phrases from the training data set, they did demonstrate the possibility of coming up with novel sentences for negotiations.


The bots did all of this by predicting the possible outcomes of various responses, right to the end of the conversation, and the resolution of a beneficial deal. A vast number of responses that were not likely to result in a good deal down the line were discarded right away. The dialogue model is rolled out till the end of the conversation, and the chatbots choose only the phrases with the maximum expected reward.

The aim of the research was to create chatbots that can engage more meaningfully with humans. Key steps to building a personalised digital assistant include building the capabilities to reason, converse and negotiate. Facebook has published the research and the code for the agents as well as the data source, all of which have been open sourced.

also see