According to Meta, an AI that has been taught to play a board game, and therefore negotiate with human players and infer their motivations, could have applications in enterprise chatbots.
Meta trained an AI agent to play a board game involving interaction with other players to persuade them to support his strategies, only to eventually betray them. In a blog post published by the company, which owns Facebook, Instagram and WhatsApp, Meta claims that its AI named Cicero may have extensive applications in the near future, including the development of more advanced virtual assistants through the use of combination of technologies such as natural language processing (NLP) and strategic reasoning.
In a research paper published in the academic journal Science, Meta reports that his AI Cicero achieved human-level performance at the historical simulation and negotiation game Diplomacy (or Diplo). In an online competition where she played 40 games against 82 humans, the AI Cicero ranked in the top 10% of participants who played more than one game. Diplo opposes seven players who each control the military forces of seven “great powers”, the objective being to manage to control more than half of the European continent extended to the Maghreb and the north of the Middle East. Each turn begins with a negotiation between players to gain support for their plans and a coordinated movement of the military forces of players who have made alliances. Without the support of other players, many of these campaigns will fail.
“The game challenged the AI agent, because in order to win, he had to figure out whether his opponents were bluffing or strategizing to win the game. The AI had to show a certain level of empathy during the game in order to collaborate with other players, which was not the case until now in games like chess where the AI against human adversaries. Since they’ve been around, AI agents have gotten better at strategy games: In 1997 IBM’s Deep Blue software beat world chess champion Gary Kasparov, and in 2016 DeepMind’s AlphaGo beat world chess champion Gary Kasparov. beat top Go player Lee Sedol. Facebook has also developed another AI engine capable of surpassing humans in poker.
Cicero relies 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 come up with its own strategy, the natural language processing engine generates messages and analyzes responses in conversations with other players to negotiate and achieve an agreement,” the researchers explain. To help the AI agent generate relevant conversations, the researchers started with a pre-trained 2.7 billion-parameter natural language generation model on texts from the Internet and refined it. with conversations between human gamers in over 40,000 game sessions sourced from webDiplomacy.net.
“We have developed techniques to automatically annotate messages in the training data with the corresponding expected in-game moves, so that at the time of inference we can control the generation of dialogue to discuss the specific actions desired. for the agent and its chat partners,” the researchers said in a more detailed blog post. Meta has made Cicero’s code freely available so that other researchers can exploit the capabilities of its AI agent. In addition, the company has created a portal to generate research proposals in the field of human-artificial intelligence cooperation through NLP using diplomacy as the core concept.
Long term plans
Big tech companies, like Microsoft, Google, and Amazon, are racing against time to develop more advanced freelance virtual assistants to support a variety of business use cases, from call centers to sales agents. AIs that can analyze feelings and teach a person new skills. According to a report by Fortune Business Insights, the global natural language processing (NLP) market, which includes such assistants, is expected to grow from $26.4 billion in 2022 to $161.8 billion in 2029. what they stated in a blog post, Meta researchers seem to think that Cicero’s success in Diplo exceeds the capabilities of other virtual assistants available today. “For example, today’s AI assistants can perform simple question-and-answer tasks, like telling the weather, but what if they could sustain a conversation long enough to teach the other person a new skill? »
In their note, the researchers refer to tools like Google Duplex, Amazon Alexa, Microsoft’s Xiaoice, and Apple’s Siri. But Cicero isn’t cut out for long conversations either, as his reasoning is strictly short-term. As Meta researchers explain in their Science article, “Strategically, Cicero reasons about dialogue solely in terms of player actions for the current turn. He didn’t model how his dialogue might affect the relationship with other players over an entire game.”