Games chosen for the 2021 competition are following:

Reconnaisance Blind Chess

Reconnaissance Blind Chess (RBC) (play online) is like Chess except a player cannot directly observe location of its opponent’s pieces. The players can however use a sensing action, a kind of “radar”, that reveals all the pieces under a probed square region of size 3x3. The game has been used in the RBC competition. It is a challenging game, as there is no common knowledge about all piece positions (apart from the starting board). The approximate total number of game states that can be considered in a game of RBC is approximately 10^93 times that of a game of Chess, because each state requires keeping track of not only what the game board actually looks like, but the information that each player has acquired [1]. You can find more detailed analysis of the game.

While there exist strong Chess bots, such as Stockfish, and they have been adopted to work with RBC by identification of the current chessboard, this domain-specific adaptation will not work for the following card game.

Gin Rummy

Gin Rummy (play online) is a classic game, one of the most widely-played two-player card games of the 20th century, and large sums are wagered on it by expert players. Like other well studied games (Backgammon, Chess, Go, Poker, etc…), this makes it a domain where human performance represents an especially strong benchmark. It’s noteworthy that the game is played by some of the most successful hedge fund managers, including Warren Buffett and Howard Marks.

Gin Rummy is a large game with at least 10^80 information states and a large number of states per info-state (there are over 1 billion possible opponent hands off the deal versus 1,225 in heads up Texas hold ‘em). HUNL has ~10^160 info-states, but can be abstracted to a much smaller game through bet bucketing (action abstraction). This technique, along with several others used in Poker research, is not applicable in a straightforward way to Gin.

It is also a difficult game for an algorithm to learn from scratch. Audrunas trained ARMAC on Gin for a week and it failed to learn a good strategy. ARMAC produced promising results in Poker and Montezuma’s Revenge, so that offers some empirical evidence that Gin is difficult and represents a distinct challenge. [2]

Secret game

To drive algorithm development towards generality, we will use a third game, which is secret. It is chosen by a committee member that does not participate in the competition. The game will be announced after the final submissions are posted. The game will use standard OpenSpiel API, just as the previous games.

References

[1] Mastering Reconnaissance Blind Chess with Reinforcement Learning

[2] Conversation with John Schultz.