While artificial intelligence has mastered chess, go, and poker, it has yet to win games reliant on language and comprehending the other person’s intentions. According to Meta, the company’s artificial intelligence group has recently built a model capable of comprehending and acquiring human trust, which has boosted the rankings of the popular game Diplomacy.
Even though the Meta Platforms conglomerate is best recognized for owning and developing social networks such as Facebook, Instagram, and WhatsApp, as well as pioneering virtual reality, Meta is also actively working on artificial intelligence research. The results of this work include artificial intelligence capable of translating from languages that have no written form. Today, Meta introduces another milestone in the evolution of artificial intelligence: Cicero, an AI agent that negotiates, persuades and collaborates with people. It accomplishes this through the popular game Diplomacy.
For decades, games against actual humans have served as not only a training ground for artificial intelligence, but also a means of demonstrating how far technology has progressed. Deep Blue astonished the world by defeating chess grandmaster Garry Kasparov in 1997. In 2015, AlphaGo became the first computer program to defeat a professional at the ancient Chinese board game go. Both chess and go have a set of principles that can be instilled into an artificial intelligence using data from hundreds of games. And with hundreds more, teach it human-level strategy and decision-making.
Diplomacy is a game that has long been considered to be impossible for an artificial intelligence to learn. The game needs the computer to recognize the plans, outlook on the game, and motivations of the other players rather than a mathematical comprehension of the actions done on the board. The AI must then convert this comprehension into natural language signals that persuade the human participants that it is correct. Meta created the Cicero agent by combining two types of artificial intelligence models: a strategic thinking model (similar to AlphaGo and Deep Blue) and a natural language processing model (similar to GPT-3).
The strategic thinking model’s abilities enabled Cicero to make the best decisions for himself, while natural language processing enabled the AI to communicate naturally with the players. It was trained using a BART-type language model with 2.7 billion parameters gathered from various sorts of internet documents, then modified with training data from 40,000 online games involving human players on WebDiplomacy.net.
Cicero was able to become one of the finest players of the browser version of Diplomacy, scoring “more than double the average score” and getting into the top 10 players who have played more than one game as a result of such training.
Although Cicero can only play Diplomacy, the technology used to construct the AI agent has many real-world applications, as Meta claims. Controlling natural language generation through planning and reinforcement learning, according to Meta AI, can help break down communication barriers between humans and AI-powered agents.
References:
https://gizmodo.com/meta-ai-cicero-diplomacy-gaming-1849811840