Meta has taught an AI agent how to play a boardgame in which it must converse with other players in order to convince them to support its strategies, and then turn on those players after the conversation.
According to a blog post published by the company, which also owns Facebook, Instagram, and WhatsApp, the company believes that its Cicero artificial intelligence may have widespread applications in the near future. One of these applications may be the development of smarter virtual assistants through the combined use of technologies such as natural language processing (NLP) and strategic reasoning.
Meta stated that its Cicero artificial intelligence achieved human-level performance at the strategy boardgame Diplomacy in an online league where it played 40 games against 82 humans and ranked in the top 10% of participants who played more than one game. The research article was published in the academic journal Science.
In the board game Diplomacy, seven players compete against one another to see who can control a map of Europe. At the beginning of each turn, the players confer with one another to garner support for their respective strategies, and at the end of the session, they attempt to carry out their respective manoeuvres simultaneously. Many of these actions won’t be successful unless they get the backing of other players.
According to Meta, the AI agent faced difficulty playing the game since in order to win, it needed to determine whether or not its opponents were bluffing or were planning to win the game by employing a certain strategy. In order to successfully build alliances with other players, the AI had to demonstrate a certain degree of empathy while playing the game. This is something that AIs have not been required to demonstrate while competing against human players in games such as chess.
The ability of AI agents to play strategy games has significantly improved throughout the years: In 1997, IBM’s Deep Blue software prevailed over the reigning world champion of chess, Gary Kasparov. In 2016, DeepMind’s AlphaGo software prevailed over the reigning world champion of Go, Lee Sedol. In addition, Facebook has built yet another AI system that is superior to human players in poker.
Logic based on a strategy
Strategic reasoning and natural language processing are Cicero’s two primary foundational technologies (NLP). The researchers explained that the natural language processing engine generates messages and analyses responses in conversations with other players in order to negotiate and come to an agreement. While the strategic reasoning engine forecasts the moves of other players and uses that information to form its own strategy, the natural language processing engine also engages in conversations with those players.
Researchers began with a natural language generation model that had been pre-trained on text taken from the internet and then fine-tuned it using conversations that took place between human players in over 40,000 games hosted on webDiplomacy.net. This was done in order to assist the AI agent in generating conversations that were pertinent to the situation.
In a more in-depth blog post, the researchers said, “We developed techniques to automatically annotate messages in the training data with corresponding planned moves in the game, so that at inference time we can control dialogue generation to discuss specific desired actions for the agent and its conversation partners.” “We developed techniques to automatically annotate messages in the training data with corresponding planned moves in the game,” the blog post began.
Meta has made the source code for Cicero available to the research community so that other scientists can expand upon the capabilities of the AI agent.
In addition, the company has established a platform that serves as an online forum for the solicitation of research proposals in the field of human-AI collaboration via NLP with the concept of diplomacy serving as its central tenet.
Long-term plans
Large technology companies such as Microsoft, Google, and Amazon are competing against one another to develop smarter, more independent virtual assistants to support a wide variety of business use cases. These use cases range from call centres to AI agents that can conduct sentiment analysis and teach new skills to an individual. According to a report by Fortune Business Insights, the global market for natural language processing (NLP), which includes such assistants, is projected to grow from $26.4 billion in 2022 to $161.8 billion by 2029. This growth will take place from a current value of $26.4 billion.
Researchers at Meta appeared to imply that the accomplishments of Cicero in the field of diplomacy surpass those of other virtual assistants that are currently on the market. They wrote in a blog post: “For example, current AI assistants can complete simple question-answer tasks, like telling you the weather — but what if they could hold a long-term conversation with the goal of teaching you a new skill?” They seemed to imply that Cicero’s success in the field of diplomacy surpasse
This is a jab at services such as Google Duplex, Alexa from Amazon, Xiaoice from Microsoft, and Siri from Apple. Cicero, on the other hand, is not equipped to engage in lengthy discussions due to the fact that its reasoning is only applicable to the immediate future. “From a strategic perspective, Cicero reasoned about dialogue purely in terms of players’ actions for the current turn,” the researchers from Meta said in the paper that was published in Science. It did not model how its dialogue could potentially affect the relationship with other players over the course of a game’s long-term progression.
