This project uses Counterfactual Regret Minimization to solve the game of Setback (aka Auction Pitch).

Setback's move tree is not as large as that of some other card games (e.g. Bridge, Hearts), but it is still too large for vanilla CFR on commodity hardware. The approach used here is to create a simpler abstract game that is essentially identical to Setback, but whose move tree is small enough for vanilla CFR. Defining such an abstraction is much more of an art than a science, relying heavily on personal intuition for what information is most salient when playing Setback.

Keys are only created for infosets with more than one legal action. Each key is a string with the layouts described below.

These keys are used when training the baseline (score-insensitive) model.

Suit strength is determined as follows:

• Convert the rank of each card in the suit to an integer by subtracting one from its face value. (E.g. Two → 1, Three → 2, ..., Ten → 9, Jack → 10, ..., Ace → 13).
• Sum the values.

The second-strongest suit is included iff the difference between its strength and the strength of the strongest suit is less than 2.

These keys are used when bootstrapping a score-sensitive model from a baseline model. Unlike the baseline keys, these keys are formatted using / characters, rather than by absolute position within the string, as follows:

Score-need/High-bid/High-bidder/Num-bids/Hand

Where:

• Score-need: Representation of the current score, consisting of two characters: Them-need and Us-need.
  • Them-need: The number of points the other team needs to win, or x if the other team needs more than four points (i.e. can't win on this deal).
  • Us-need: If the other team is ahead and needs four points or less then !, else ..
• High-bid: The current high bid as an integer: 0, 2, 3, or 4.
• High-bidder: Index of the current high bidder, relative to the dealer: 0, 1, 2, 3, or -1 if there is no current high bidder.
• Num-bids: The number of bids made so far in this auction (0 - 3).
• Hand: The bidder's hand, as expressed in positions 1-12 of the baseline auction key.

These keys are used by the auction winner to establish trump with the first card played.

A normal playout key is a string of up to 24 characters, laid out as follows:

Trailing dots (.) are trimmed from the baseline keys to save space. (There are a lot of them, especially the playout keys.)

1. Run the TrainBaseline project to generate a baseline strategy profile (Baseline.strategy) that is optimized to play a hand of Setback without regard to the score of the game. Warning: This requires a machine with at least 16GB of RAM, and will take several weeks/months. I recommend at least 10 million CFR iterations.
2. Copy Baseline.strategy into the TrainBootstrap project and then run the project to bootstrap a strategy profile (Bootstrap.strategy) that is optimized for score-aware bidding (e.g. by bidding more aggressively if the opposing team is close to winning the game). This requires less time and RAM. I recommend at least 20 million CFR iterations.
3. Copy Baseline.strategy and Bootstrap.strategy into the LoadDatabase project, and then run the project to create a SQLite database that can be used to play Setback (via the DatabasePlayer module).
4. Copy the SQLite database Setback.db into the Setback.Cfrm.PlayGui project, and then run the project to play against the trained model (Windows only).