Meta has trained an AI agent to play a board game that involves chatting with other players to persuade them to support their strategies, and then betraying them.

The company, which owns Facebook, Instagram and WhatsApp, says its Cicero AI could have widespread applications in the near future, including the development of smarter virtual assistants with the combined use of technologies such as natural language processing (NLP) and strategic reasoning, according to a blog post published by the company.

In a research article in the academic journal Science, Meta said that his Cicero AI achieved human-level performance in the strategy game Diplomacy in an online league where he played 40 matches against 82 humans, ranking in the top 10% of all players. participants who played more than one game.

Diplomacy pits seven players against each other for control of a map of Europe. Each round begins with the players negotiating with each other to gain support for 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 for the AI ​​agent, Meta said, because winning required him to understand if his opponents were bluffing or strategizing in a certain way to win the game. The AI ​​needed to extend a certain level of empathy while playing to form collaborations with other players, which the AIs did not need to do when playing games like chess against human opponents.

AI agents have gotten better at strategy games over the years: In 1997, IBM's Deep Blue software beat world chess champion Gary Kasparov, and in 2016, DeepMind's AlphaGo beat the best Go player, Lee Sedol. Facebook has also developed another artificial intelligence engine that can outperform humans at poker.

strategic reasoning

Cicero is based on two main technological components: strategic reasoning and natural language processing (NLP). While the strategic reasoning engine predicts the moves of other players and uses this information to form its own strategy, the natural language processing engine generates messages and analyzes the responses in conversations with other players to negotiate and strike a deal, the players explained. researchers.

To help the AI ​​agent generate relevant conversations, the researchers started with a 2700 billion parameter natural language generation model pre-trained on Internet text and refined it with conversations between human players in more than 40 webDiplomacy games. net.

“We have developed techniques to automatically annotate messages in training data with the corresponding planned moves in the game, so that at the time of inference we can control the generation of dialogue to discuss the specific actions desired for the agent and their chat partners. the researchers said. he said he in a more detailed report. blog post.

Meta has opened up Cicero's code for other researchers to build on the AI ​​agent's capabilities.

In addition, the company has created a portal to invite research proposals in the field of human-AI cooperation through NLP using diplomacy as a central concept.

long term projects

Big tech companies like Microsoft, Google, Amazon are competing with each other to develop smarter standalone virtual assistants to support a variety of business use cases, ranging from call centers to AI agents that can perform sentiment analysis and teach an individual new habilities. The global market for natural language processing (NLP), which includes these assistants, is expected to grow from €26,4 billion in 2022 to €161,8 billion by 2029, according to a Fortune Business Insights report.

The Meta researchers seemed to suggest that Cicero's success in diplomacy superseded the abilities of other virtual assistants available today, stating in a blog post: "For example, today's AI assistants can perform simple question-and-answer tasks. , as indicated by the weather report. but what if they could have a long-term conversation with the goal of teaching you a new skill? »

It's an exploration of tools like Google Duplex, Amazon Alexa, Microsoft's Xiaoice, and Apple's Siri. But Cicero is not prepared for long-term conversations either, since his reasoning is strictly short-term. As the Meta researchers put it in the Science article, “Strategically, Cicero reasoned about the dialogue solely in terms of the player's actions for the current round. He did not model how his dialogue could affect the relationship with other players throughout the game ».

Copyright © 2022 IDG Communications, Inc.

Share This