Libratus takes machine learning from board games to the bargaining table

A poker program beat four top professionals by $1.77 million in chips. Hidden information now yields to computing, and negotiation is next.
Libratus, a program from Carnegie Mellon University, finished 20 days of heads-up no-limit Texas hold'em against four top professionals $1,766,250 ahead in chips. The match, 120,000 hands at a Pittsburgh casino, ended on 30 January with a margin its developers call statistically significant.
Poker matters because the cards are hidden. Chess and Go show both players the whole board; a negotiation, an auction or an attack on a network does not, and those are the settings where companies spend real money on strategy. A system that bluffs, reads bluffs and adjusts when an opponent finds a weakness moves machine learning toward the bargaining table, where the prize is measured in procurement savings and auction prices rather than in clicks.
The method runs on computing, not on human records. Libratus learned no hands from people. It computed its base strategy in about 15 million core hours before the match, used some 600 of the 846 nodes of a public research supercomputer during play, solved the end of each hand in real time and patched the holes the professionals found while the match went on. On 6 January a University of Alberta team posted a separate program, DeepStack, which beat professionals over 44,000 hands with self-play and deep learning. Two labs reaching the same result within a month marks a technique, not a stunt.
The people who can build such systems sit in a handful of university labs, and the commercial path runs through their spin-outs. Tuomas Sandholm, who leads the Libratus work, already applies the same ideas to automated negotiation through his company Optimized Markets, and names auctions, cybersecurity, finance and medical treatment planning as further fields. Value in this corner of machine learning sits with a small group of researchers, and it goes to whoever backs them at the spin-out stage, before the largest technology companies hire them.
Founders looking for a product should start where the rules of an interaction can be written down: procurement auctions, advertising bids, price negotiations with suppliers, defense against an attacker who adapts. Each has a counterparty who hides information and an outcome measured in money, the setting Libratus was built for. Fanaura looks for teams that pair this research with access to a real market's transaction data, the combination that turns a laboratory result into revenue.
Two-player poker is still a game, with fixed rules, two seats and a score kept in chips, and that is the main reason for caution. Real negotiations have many parties, shifting rules and payoffs no one can fully measure, and the computing bill grows with the size of the game. Two results would answer it: a system that holds up with more than two players, and a first commercial deployment in an auction or a negotiation that shows a measurable gain in price.
Machines have learned to bluff. The open question is who will pay them to negotiate.
Building an AI-native company? Apply in five minutes. →
More from Fanaura
Building machine learning for auctions, bidding or negotiation? Send us the project.
Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.
Get the news by email
One short note each weekday. No marketing, and you can unsubscribe at any time.