CICERO: Meta AI's Open-Source Model Plays Diplomacy, Strategizes With Players
Meta AI has released a new model designed to play the game Diplomacy. This model shows human-level performance and communication skills within the game.
Created on November 22|Last edited on November 22
Comment
Some of the most recognizable AI models today are those which played games: DeepBlue and AlphaGo stand as pillars in the progress of machine learning. However, beyond the rigid structure of Chess and Go, games that heavily rely on human communication have been considered the final boss of game-playing AI models.
In recent years, AI models for human communication have rapidly developed to a point where it can be difficult to tell AI from humans. With this constantly evolving tech, building models to play games with a focus on human communication becomes more attainable.
With that in mind, Meta AI researchers have developed Cicero, a model built to play Diplomacy, a game with player communication at its core.
In the game, players must strategize together, make bluffs, or even lie and backstab their way to accumulating territory on the map, all accomplished by talking one-on-one amongst themselves.
Cicero takes the strategy and logical reasoning of classic game-playing AI models and adds a new layer of natural language processing for fruitful communication. It can naturally form alliances with others, suggest actions, set goals, and everything else a human player would do exactly the same way they would do it.
How Cicero works
Cicero has a lot going on under the hood, but essentially, it is made up of two main models: one for strategic reasoning, and the other for natural language processing. Read through the blog post for more in-depth information on Cicero's pipeline.
Strategic reasoning
To start a turn, it uses its strategic reasoning brain to come up with an initial prediction based on board state and dialogue history, which it then develops into an intent for it and its partners.
Interpreting dialogue is tricky, and depending on the training data, could be easily exploited. To make sure that Cicero's decisions are reasonable and well thought out, a planning algorithm called piKL was created to very robustly account for everything going on in the game so that it can weigh decisions properly, avoiding scenarios where players could manipulate Cicero's inherent flaws as an AI system.
Natural language processing
Cicero uses the intent, board state, and dialogue history with its natural language processing brain to create candidate messages, which it then filters through to decide the perfect message to send to another player.
Without careful filtering, it was observed that Cicero would sometimes unintentionally leak information from conversations it had with other players. To fix this, as well as avoid nonsense output and keep things focused on playing the game properly, they had to implement a variety of output filters.
Cicero's performance, and using it for yourself
Over the past few months, Cicero played live games with real players on webDiplomacy.net, a site for playing the game online. In the 40 games it played, Cicero earned more than twice the average score of its opponents and even ranked in the top 10% of those who had played more than one game previously.
Cicero is found to not make emotionally-motivated decisions, such as those of revenge or pity, which a human player could often succumb to. Additionally, Cicero is a very honest player, which is considered high-level play unlike playing with lies and deception, which might be the first instinct for new players.
Logs for the games that Cicero played are available to look through on the GitHub repository, where you can also find instructions for downloading the various components of Cicero and running it for yourself.
Also available on GitHub is Diplodocus, a separate model build to play a variation of Diplomacy called Gunboat Diplomacy, which eliminates the language portion, leaving just the actions (which can, in a way, still facilitate communication).
Find out more
Learn more at Cicero's web page by clicking here.
Add a comment
Tags: ML News
Iterate on AI agents and models faster. Try Weights & Biases today.