I built a little program to construct a bearoff database for me and put together an initial cut. It covers bearoffs on all six points in both players' home boards, with each player having up to nine checkers remaining.

I wrote out the data to a text file, with a key representing the board layout and value representing the equity (ie probability of win in this case, since the probability of a gammon is always zero). There are 25M elements in the table and the file takes 557MB, which needs to be loaded into process memory to be used. I looked at a table for six points and ten checkers, but that ended up being too big (closing on 4GB). I'm sure there's a clever way to partially load elements of the table when needed to avoid this kind of process memory bloat, but it's not worth figuring that out now.

The key representing the board layout is very simple. First you start with the player who holds the dice: for each point, starting with #1, add a character representing the number of checkers on the point. "A" represents 0, "B" 1, and so on. I used letters because potentially the number of checkers could be > 9 in the generic case. So for a bearoff db for six points, the key has six elements for each of the player's six points in the home board, and each element is a character. Then the same thing is appended for the opponent's position. For six points that means a twelve-character key.

Now I can build a strategy that toggles between a normal network, a race network, and exact bearoff probabilities in the appropriate cases. I'm going to design it generally so it'll work with potentially many different networks (more than the two listed above).

I wrote out the data to a text file, with a key representing the board layout and value representing the equity (ie probability of win in this case, since the probability of a gammon is always zero). There are 25M elements in the table and the file takes 557MB, which needs to be loaded into process memory to be used. I looked at a table for six points and ten checkers, but that ended up being too big (closing on 4GB). I'm sure there's a clever way to partially load elements of the table when needed to avoid this kind of process memory bloat, but it's not worth figuring that out now.

The key representing the board layout is very simple. First you start with the player who holds the dice: for each point, starting with #1, add a character representing the number of checkers on the point. "A" represents 0, "B" 1, and so on. I used letters because potentially the number of checkers could be > 9 in the generic case. So for a bearoff db for six points, the key has six elements for each of the player's six points in the home board, and each element is a character. Then the same thing is appended for the opponent's position. For six points that means a twelve-character key.

Now I can build a strategy that toggles between a normal network, a race network, and exact bearoff probabilities in the appropriate cases. I'm going to design it generally so it'll work with potentially many different networks (more than the two listed above).