How to build a decent backgammon bot:
- Build a backgammon framework. This should define things like: strategies, which take boards and return values (larger values are "better boards"); games, which take a strategy for each player and can evolve until the end of the game; matches, which hold games and represent tournament play; and doubling strategies, which take a strategy & board & doubling cube details and determine when to double and when to take vs pass.
- Build a strategy that uses neural networks to estimate game probabilities: probability of any win; probability of any gammon win; probability of backgammon win; probability of any gammon loss; and probability of backgammon loss. Inputs should be the standard Tesauro inputs, plus a few extended inputs that encode more complex game features. One exception: you don't need an input noting whose turn it is; just always evaluate the network from the perspective of the current player. You can use different neural networks for different game phases: at least contact and race, and perhaps others. Use a standard neural network package like FANN, or roll your own if you like.
- Build one-sided bearoff databases for cubeless play to handle end games. You can use two-sided databases if you like, but one-sided databases perform almost as well and are much smaller (or equivalently, for the same database size you can cover a much larger fraction of the end game).
- Train your neural networks for cubeless play. Start training with TD training on self-play; once that's converged, switch to supervised learning based on the GNUbg training databases. Benchmark the performance of your bot with the GNUbg benchmark databases, possibly also with self-play against standard players like PubEval.
- Implement the Janowski doubling strategy for cubeful money play. Use a constant cube life index of 0.70, or try to come up with more state-specific cube life indexes.
- Get a match equity table. Either build your own or use one of the standard published tables.
- Implement the Janowski doubling strategy for match play, hooked into your match equity table. Use the same constant cube life index of 0.70, or try to do better with a state-dependent cube life index.
- Implement multiple-ply strategies, taking the regular neural net-based strategies and digging into those.
And then you're done! At that point you'll have a fairly strong player which you can wrap a UI around or hook into online backgammon servers like FIBS.
