Facebook AI Research Is A Game-Changer
For decades, computer programmers have been trying to beat multiplayer games by finding reliable patterns in data.
Researchers at Facebook and Carnegie Mellon University published a whitepaper in Science Journal in July that flips this switch. Their software embraces randomness, and it is reliably beating humans at games.
This is a big step forward for data science. It shows where the world is headed.
By now, investors should know that machine learning and artificial intelligence have taken the enterprise world by storm. Companies like Alphabet, Amazon and Netflix built huge platforms in internet search, ecommerce and media distribution by using software algorithms to understand what their customers wanted, sometimes even before they knew it.
There have been big strides in programs mastering board games like checkers and chess. A super AI called AlphaGo, a prototype from Google Mind, even conquered Go, an extremely complex Chinese strategy game. It trounced Lee Sedol, the best Go player in the world, in a widely publicized match.
Many software developers assumed Go was unassailable given the state of AI development.
But AlphaGo caught Sedol off-guard. Developed by playing against itself millions of times, the software learned to be unpredictable. In the end, Sedol couldn’t determine why moves were being made.
Pluribus, the Facebook and CMU collaboration, leans hard on being unpredictable. It makes sense.
Whereas checkers, chess and even Go are limited to two players, poker traditionally involves more, creating a significant challenge for an AI. Exploiting the weaknesses of a single player would not be enough. Pluribus was designed to take on five world-class opponents simultaneously.
With six players, six hands of cards, the bets, and the sheer multitude of possible outcomes, the game would have overwhelmed even the most robust supercomputer. Pluribus needed a reliable edge.
The edge researchers used was counterfactual regret minimization.
The big idea: During training, as Pluribus played against itself, the AI noted the outcome of games given all of the variables. Then, the software systematically played game after game, altering a variable each time.
Pluribus learned that some moves, even crazy ones, were winners in defined circumstances. Flummoxed players, for example, would fold. And that is the key.
We humans fidget nervously; we behave predictably when we are anxious. Professional poker players can spot these tells a mile away. Pluribus does not really play to its opponents’ weaknesses. It plays to its inherent strength — it has no tell. It is completely unpredictable.
The proof is in the gameplay.
