Another game falls to an AI player

From The Economist:

Backgammon was an easy win. Chess, harder. Go, harder still. But for some aficionados it is only now that artificial intelligence (ai) can truly say it has joined the game-playing club—for it has proved it can routinely beat humans at Diplomacy.

For those unfamiliar with the game, its board is a map of Europe just before the first world war (except that, for no readily apparent reason, Montenegro is missing). Participants, seven ideally, each take on the role of one of the Great Powers: Austria, England, France, Germany, Italy, Russia and Turkey. Each has armies and navies, and geographically based resources to support them, and can use its forces to capture the territory of neighbours, thus gaining the means to raise more forces while depriving others of the same.

The trick is that, at least at the beginning, players will get nowhere without making agreements to collaborate—yet they are not bound by the game’s rules to keep to these agreements. Only when orders for the movement of troops and vessels, which have to be written down, are revealed, does a player discover who really is a friend, or an enemy.

Cicero, a program devised by a group of Mark Zuckerberg’s employees who dub themselves the Meta Fundamental ai Research Diplomacy Team, proved an adept pupil. As the team describe in Science, when they entered their creation into an online Diplomacy league, in which it played 40 games, it emerged as one of the top 10% of players—and no one rumbled that it was not human.

In all past ai game-playing projects the program has learned by reinforcement. Playing repeatedly against itself or another version of itself, it acts first at random, then more selectively. Eventually, it learns how to achieve the desired goal. Cicero was taught this way, too. But that was only part of its training. Besides having the reasoning to plan a winning strategy, a successful Diplomacy player must also possess the communicative ability to implement it.

The Meta team’s crucial contribution was therefore to augment reinforcement learning with natural-language processing. Large language models, trained on vast amounts of data to predict deleted words, have an uncanny ability to mimic the patterns of real language and say things that humans might. For Cicero, the team started with a pre-trained model with a baseline understanding of language, and fine-tuned this on dialogues from more than 40,000 past games, to teach it Diplomacy-specific patterns of speech.

To play the game, Cicero looks at the board, remembers past moves and makes an educated guess as to what everyone else will want to do next. Then it tries to work out what makes sense for its own move, by choosing different goals, simulating what might happen, and also simulating how all the other players will react to that.

Once it has come up with a move, it must work out what words to say to the others. To that end, the language model spits out possible messages, throws away the bad ideas and anything that is actual gobbledygook, and chooses the ones, appropriate to the recipients concerned, that its experience and algorithms suggest will most persuasively further its agenda.

Cicero, then, can negotiate, convince, co-operate and compete.

Link to the rest at The Economist

PG notes that lawyers frequently negotiate, convince, co-operate and compete. He will also note that the market for legal AI software is booming now.

He understands the state of the legal art hasn’t reached the point where one can buy a software program instead of hiring a lawyer to go to court, but he suspects it’s only a matter of time.

5 thoughts on “Another game falls to an AI player”

  1. The biggest problem with an AI substitute for lawyers is that it will assume that all of the relevant evidence and facts are complete, clearly presented, incontrovertible (not to mention thoroughly honest, or at least not intentionally deceptive), and sufficient to reach a clear result without human inference.

    In reality, not so much.

    The less said about how these assumptions track with law school casebooks and methods, the less y’all will distrust the fabulousness of the legal profession.

    • I haven’t tried any law AI programs, but, given my general amateur-level knowledge of AI in other spheres, I suspect ambiguity may be something a legal AI could deal with and respond with a range of possible solutions.

    • Why would it make that assumption? If it’s learned to play Diplomacy already, it’s got the core of dealing with “not so much” on all that.

  2. Back in college we had a biweekly game of Diplomacy. I am proud to announce that I won more than one time in seven. The idea of an AI being even more successful does not surprise, especially against six human players who don’t know it is an AI. Viewed as a wargame, the mechanics are simple. Yet many players are not very good at the military side. This was my ace in the hole, as I cut my teeth shoving panzers into Stalingrad on hex maps. I would expect an AI to easily master this aspect. On the diplomacy side, the real artistry was in the timing of the stab in the back. I would also expect an AI to handle this easily. This leave the art of the deal. Given the limited scope of discussion, this too should be well within the abilities of a language bot.

    That being said, there are really two versions of Diplomacy, using exactly the same rules: the game played by strangers and the game played by players who know each other. The dynamic is very different: not because friends won’t stab each other in the back, but because they know one another’s style of play. I wonder how an AI would do there.

    • I haven’t played for years, but when we had regular games at work the difference was not so much that we knew each other but that there was a parallel game of “tit for tat” running across the game series. Elegant retaliation for past betrayals was often set above success in the current game , in the expectation that this would improve one’s overall chances in the series.

      I have to agree that most of the players were rather bad at the wargaming side, but I think that I was the only player in our group who had actually commanded miniature figures in action, let alone pushed small pieces of card across a hex map.

Comments are closed.