ShareDeepmind, an AI company backed by Google, has created a system called Player of Games which is expected to perform well in both complete information games and incomplete games. This style of AI is thought to have significant potential for solving many real-world problems in the coming years.
Tasks such as planning the most efficient route around traffic, customer relations and even contract negotiations can benefit from this type of AI.
Jack of All Trades
The release paper published by Cornell University contains the following:“...a general-purpose algorithm that unifies previous approaches, combining guided search, self-play learning, and game-theoretic reasoning.”
“Player of Games is the first algorithm to achieve strong empirical performance in large perfect and imperfect information games -- an important step towards truly general algorithms for arbitrary environments.“
Exciting stuff, and a change in direction from the more normal strategy of building a program for only one specific goal.
Player of Games recently squared up against Slumbot, reputed to be the strongest, publicly available NL Hold’em AI program and won the battle comfortably but nowhere even close to the hammering Facebook’s ReBel gave it.
Player of Games also went head-to-head against Stockfish, one of the strongest chess engines available off the shelf and didn’t disgrace itself, proving stronger in many configurations. Against AlphaZero in a match of go it only won 0.5% of the games but was thought to be performing almost at the level of a professional human player.
Techie Details
This new creation starts its learning by using a standard tree search strategy with typical machine learning processes. The trick that makes it great at both perfect and imperfect information challenges is to factor in game theoretic reasoning and growing-tree counterfactual regret minimization (GT-CFR) when planning which action to take.DeepMind senior research scientist Martin Schmid said:
“This is a step towards generality — Player of Games is able to play both perfect and imperfect information games, while trading away some strength in performance. AlphaZero is stronger than Player of Games in perfect information games, but [it’s] not designed for imperfect information games.”
Generalised AI has to be one of the most exciting advances we will witness over the next few decades as random tasks we previously thought impossible to perform perfectly will have close to perfect strategies revealed to us.
