Cribbage

Python library for experimenting with the game of cribbage.

I've forked this heavily from https://github.com/wroberts/cribbage. Thanks very much, Will - your framework made it easy to get started.

Will was interested in training a good computer cribbage player; I'm more interested in learning what heuristics are best that are learnable by a human.

To that end, I'm developing some CribbagePlayer classes here so I can compare their rankings when they play against each other. Currently, they include:

• MaxerCribbagePlayer, aka Max: Uses any convenient techniques, like exhaustive lookahead and tables of expected values, that are easy to do in software.

• HeuristicCribbagePlayer, aka Helen: Uses techniques that I think I could reasonably learn and remember - mostly simple rules that are expressed in integer point values.

• SimpleCribbagePlayer, aka Simon: Will's implementation in its simplest mode, for a straw man baseline.

Currently, Max is the best player, but Helen is better than Simon. Typical stats (from pipenv run compare):

Player       Games won  Percent  Averages for Dealer and Pone                                     Hand counts
Max              17485   59.1%   D crib: 5.0  D hand: 8.0  D play: 3.6  P hand: 8.1  P play: 2.2  4.2+4.4=8.6
Helen            15114   51.1%   D crib: 4.6  D hand: 8.0  D play: 3.6  P hand: 8.0  P play: 2.2  4.2+4.4=8.6
Simon            11801   39.9%   D crib: 4.1  D hand: 8.2  D play: 3.5  P hand: 8.2  P play: 2.0  4.1+4.4=8.5

and with just Max and Helen:

Player       Games won  Percent  Averages for Dealer and Pone                                     Hand counts
Max              10101   54.9%   D crib: 4.9  D hand: 8.0  D play: 3.5  P hand: 8.1  P play: 2.1  4.3+4.4=8.7
Helen             8299   45.1%   D crib: 4.4  D hand: 8.0  D play: 3.5  P hand: 8.0  P play: 2.0  4.2+4.5=8.7

Plans

• Game-centered heuristics

  • Report statistics for point values
    • Predict winner by average point values
      • Use standardized stats
      • Note the score at the beginning of a given phase - when starting a new phase.
      • Starting from where you are, go through the phases in order until reaching the target score.
        • D play and P play simultaneously (if both go over the target, go with the one with higher play value)
          • While playing, use the max of the stat and the points scored so far
        • P hand, D hand, D crib
    • Decide when to play in "risky" or "safe" mode
      • Risky = hoping for max possible points, make the choices with best possible outcomes
      • Safe = make the choices with best expected values
    • Decide when to play in "slow" or "fast" mode
      • When you expect that net point changes will be even, prefer larger or smaller movements, to get in sync with a winning position.

• Improve Max

  • Explore the full decision tree for pegging? It's not very large, and we end up exploring much of it anyway.
  • Replace standard point stats with current averages with this player over 100-200 games

• Add God player, who does full lookahead on each round with knowledge of all cards

  • Can play for min/max, where you assume perfect play in your opponent
  • Or if you predict a loss otherwise, go for the most points if your opponent chooses poorly

• Train Helen continuously, choosing random values for heuristic parameters and keeping the changes that result in better rankings over many games. See if that improves her ranking.

  • Keep track of the number of references to a given parameter? Some may be beneficial, but only in rare circumstances.
    • P6: use 0, or 2?
    • P5 as well.
    • Could tell "compare" to focus on a given parameter, and include only games where that parameter has been used.
  • Compute correlation of weight with games won, maybe piecewise.
  • Add parameters for all the heuristic scores, even the ones from authorities.
  • Parallelize this, so it can run over multiple processors at once.

• Work on determining which heuristics are most powerful/helpful. Similar process, but pick a set of parameters to clear to 0.

• Group the heuristics by degree of power, and make players Anna, Ben, and Clara, where Anna starts with a heuristic set from Helen that's easy to remember, and then each adds some more.

Getting started

pipenv install sets up the environment.

pipenv shell activates the environment, if you want to run other tools or use Python interactively.

pipenv run optimize compiles Will's C function for scoring hands, to improve speed.

pipenv run compare starts an endless tournament to rank the available players.

pipenv run train tries out the heuristics and produces graphs of how weights relate to games won for each.

pipenv run test runs all unit tests.

Links

• https://github.com/wroberts/cribbage
• http://www.cribbageforum.com/index.html