[AI Milestone] Meta Unveils CICERO to Achieve Human-level Performance at Diplomacy

In 1997, Garry Kasparov, the 1985 chess world champion, was defeated by Deep Blue—an IBM supercomputer. It wasn’t too long before DeepMind’s AI AlphaGo also successfully defeated Ke Jie, one of the world’s top ‘Go’ players. Cofounder Demis Hassasbis later remarked that Ke Jie had “pushed the AI right to the limit”. 

These are certain breakthroughs that showcase the pinnacle of AI and its capabilities. However, a new breakthrough by Meta has showcased how AI has achieved human-level performance in the gaming industry

Meta’s VP and chief AI scientist Yann LeCun said that it is the first AI system to achieve human-level performance in the popular strategy game ‘Diplomacy’. He explained that Meta’s new model is able to plan a strategy and use language to build relationships with humans and collaborate with them to achieve desired goals. 

Enter scene: CICERO 

Meta AI’s CICERO is the first AI agent to achieve human-level performance in a complex natural language game called ‘Diplomacy’. The AI agent demonstrated this SOTA performance by playing with humans on an online version of the game—webDiplomacy.net—and achieving more than double the average score of any player. It also made it to the top 10% of participants who played more than a single game. 

Check out CICERO’s code here.

The new advancement is made by using a combination of two different aspects of AI. One is strategic reasoning and second is natural language processing (NLP). 

Moreover, strategic reasoning is used in agents such as AlphaGo and Pluribus, whereas natural language processing is used in huge models such as BlenderBot 3, GPT-3, and LaMDA. This way, CICERO will be capable of deducing the need for the support of a particular player—crafting a strategy to win that player’s favour—recognising the risks and opportunities from the other player’s perspective. Such combined integration of these methods enables CICERO to strategise and reason with regards to the players’ motivation. 

The AI agent would use natural language to communicate, achieve shared objectives, coordinate plans and form alliances. Furthermore, with the game of Diplomacy, Meta’s new breakthrough is a leap forward in complexity—a near-impossible grand challenge in AI for decades. 

Check out the complete timeline of AI breakthroughs below

Diplomacy isn’t just a board game

Simply put, Diplomacy is a seven-player board game that combines a few elements from the board game ‘Risk’, the TV show ‘Survivor’ and the card game ‘Poker’. 

Several board games expect players to be the best at moving pieces around the board. However, Diplomacy is different. The game consists of a cooperative component and the only way to win it is through working with opponents to capture as much territory as possible. This coordination is achieved prior to every move in the game, with the help of a natural language negotiation. 

CICERO is made up of a controllable dialogue model that allows it to ground conversations in a set of chosen plans, which then becomes mutually beneficial for both the AI and the other human player. 

The conversations generated by CICERO are in free-form dialogue rooted within the ongoing gameplay. It may even negotiate its tactical plans with the opponent, discuss any strategic dynamics, reassure an ally about its intentions, or simply engage with in-game chit-chat—much like any human player. 

Check out the research paper here

Source link

Leave a Reply