Meta trained an AI agent to play a boardgame that involved chatting with other players to convince them to support their strategy – and then betraying them.
The company that owns Facebook, Instagram and WhatsApp said its Cicero AI could have wider applications in the near future, including developing smarter virtual assistants with the combined use of technologies such as natural language processing (NLP) and strategic reasoning. Blog posts published by companies.
In a research article in the academic journal Science, Meta said its Cicero AI achieved human-level performance in the strategy boardgame Diplomacy in an online league where it played 40 games against 82 humans, who ranked in the top 10% of participants. a game
Diplomacy pits seven players against each other for control of the map of Europe. Each turn begins with the players negotiating with each other to support their plans and ends with them simultaneously trying to execute their moves. Without the support of other players, many of these moves will fail.
The game posed a challenge to the AI agent, Meta said, in order to win it had to understand whether its opponents were bluffing or strategizing in a certain way to win the game. While the AI needed to develop a certain level of empathy while playing the game to form cooperation with other players, there was no need for the AI to do so when playing chess against a human opponent.
AI agents have been getting better at strategy games over the years: In 1997, IBM’s Deep Blue software beat world chess champion Garry Kasparov, and in 2016, DeepMind’s AlphaGo beat top Go player Lee Seidl. Facebook has also developed an AI engine that can beat humans in poker.
Cicero is built on two main technology components: strategic reasoning and natural language processing (NLP). While the strategic logic engine predicts the moves of other players and uses that information to formulate a strategy of its own, the natural language processing engine generates messages and analyzes responses to negotiate with other players and reach agreements, the researchers explained.
To help the AI agent generate relevant conversations, the researchers started with a 2.7 billion-parameter natural language generation model pre-trained on text from the Internet and fine-tuned it with conversations between human players in more than 40,000 games from webDiplomacy.net.
“We developed techniques to automatically annotate messages in training data with similar planned steps in the game, so that while guessing we can control dialogue generation to negotiate specific desired actions for the agent and its conversation partners,” the researchers elaborated. Blog post.
Meta AI has open-sourced the code for Cicero to other researchers based on the agent’s capabilities.
In addition, the company has created a portal to invite proposals on research in the area of human-AI collaboration through NLP using diplomacy as a core concept.
Long term planning
Big tech companies like Microsoft, Google, Amazon are competing against each other to create smart independent virtual assistants that can conduct sentiment analysis and teach new skills to support a variety of business use cases, from call centers to AI agents. . A separate. According to a report by Fortune Business Insights, the global natural language processing (NLP) market, which includes such assistants, is projected to grow from $26.4 billion in 2022 to $161.8 billion by 2029.
Mater researchers suggest that Cicero’s success in diplomacy surpasses the capabilities of other virtual assistants available today, saying in a blog post, “For example, current AI assistants can complete simple question-and-answer tasks, such as telling you the weather — but if they teach you a new skill Can you have a long-term conversation about goals?
It’s a dig at tools like Google Duplex, Amazon Alexa, Microsoft’s Jiaoice, and Apple’s Siri. But Cicero is not up to long-term conversations, because his argument is strictly short-term. As the Mater researchers put it in the Science paper, “From a strategic perspective, Cicero argued about dialogue purely in terms of players’ actions for the current turn. It did not model how its dialogue might affect relationships with other players over the long-term course of a game.”
Copyright © 2022 IDG Communications, Inc.