This chapter is a brief description of how gnubg calculates cubeful equities. The formulae build directly on the work by Rick Janowski Take-Points in Money Games from 1993.
The basic formulae for cubeful equities as derived by Janowski is
E(cubeful) = E(dead) * (1-x) + E(live) * x,
where E(dead) is the dead cube equity (cubeless equity) calculated from the standard formulae. E(live) is the cubeful equity assuming a fully live cube. We'll return to that in the next section. x is the cube efficiency. x=0 gives E(cubeful)=E(dead) as one extreme and x=1 gives E(cubeful)=E(live) as the other extreme. In reality x is somewhere in between, which typical values around 0.6 - 0.8.
Janowski's article doesn't mention cubeful equities, so we use the straightforward generalisation
MWC(cubeful) = MWC(dead) * (1-x) + MWC(live) * x.
as MWC is the entity that is used for match play evaluations.
The live cube equity is the equity assuming that the equity changes continuously, so that doubles and takes occurs exactly at the double point and take point. For gammonless play this is the well-known take point of 20%. Janowski derives the more general formula
TP = (L-0.5)/(W+L+0.5)
where W is the average cubeless value of games ultimately won, and L is the average cubeless value of games ultimately lost. For example, for the following position
GNU Backgammon Position ID: 4HPMwQCMz+AIIQ
Match ID : cAkAAAAAAAAA
+13-14-15-16-17-18------19-20-21-22-23-24-+ O: gnubg
| X O X O | | O X X | 0 points
| X O O | | O |
| X O | | O |
| | | O |
| | | O |
v| |BAR| | (Cube: 1)
| | | X |
| | | X |
| O | | X |
| O X O | | X X | On roll
| O X O | | X X | 0 points
+12-11-10--9--8--7-------6--5--4--3--2--1-+ X: jth
gnubg evaluates
Win W(g) W(bg) L(g) L(bg)
static: 0.454 0.103 0.001 0.106 0.003
and hence W=(0.454 + 0.103 + 0.001)/0.454=1.229 and L=(0.556+0.106+0.003)/0.556) = 1.196. For gammonless positions, e.g., a race, W=1 and L=1.
The live cube equity is now based on piecewise linear interpolation between the points (0%,-L%), (TP,-1), (CP,+1), and (100%,+W): if my winning chance is 0 I lose L points, at my take point I lose 1 point, at my cash point I cash 1 point, and when I have a certain win I win W points:
|
| Figure X Cubeful equities
|
For match play there is no simple formula, since redoubles can only occur a limited number of times.
The live cube take point is generally calculated as
TP(live, n Cube)=TP(dead, n cube) * (1 - TP(live, 2n cube)
So to calculate the live cube take points for a 1-cube at 3-0 to 7 we need the live cube take points for the 4-cube and the 2-cube. For the position above and using Woolsey's match equity table the live cube take point are:
| Cube value | TP for X | TP for O
|
| 4 | 0% | 41%
|
| 2 | 15% | 38.5%
|
| 1 | 24.5% | 27.3%
|
The calculation of these are left as an exercise to the reader.
Ignoring backgammons, the gammon rates for O and X are 0.106/54.6=19% and 0.103/0.454=22%, respectively. If O wins the game his MWC will be
81% * MWC(-3,-7) + 19% * MWC(-2,-7) = 78%
and if X wins his MWC will be
78% * MWC(-4,-6) + 22% * MWC(-4,-5) = 41%.
If O cashes 1 point he has MWC(-3,-7)=76% and if X cashes he has MWC(-4,-6)=36%. Analogous to money game the live cube MWC is calculated as piecewise linear interpolation between (0%,22%), (24.5%,24%), (72.7%,36%), and (100%,41%) (from X's point of view):
| Figure X Fully live cubeful MWC
|
Having established the live cube equities and MWCs we're now in position to calculate the 0-ply cubeful equities.
Let's start with money game: the cubeless equity is -0.097 and the live cube equity can be determined from Figure X as -0.157. Thus, the cubeful equity is -0.138.
For the match play example at the score 3-0 the cubeless MWC is 29.1% and from Figure X using wins=45.4% we can determine the live cube MWC to be 29.2%. Using a value of x=0.68 we arrive at a cubeful MWC of 29.17%.
The previous section concerned the calculation of 0-ply cubeful equities, so how so gnubg calculate cubeful 2-ply equities? The answer is: by simple recursion:
Equity=0
Loop over 21 dice rolls
Find best move for given roll
Equity = Equity + Evaluate n-1 ply equity for resulting position
End Loop
Equity = Equity/36
Note that evaluating the n-1 ply equity involves a cube decision, since the opponent may double, so gnubg will actually calculate the two n-1 ply equities: (a) assuming no double, and (b) assuming double, take. These two equities are combined with the equity for a pass, and the optimum of these three is added to the resulting equity. For a cubeful 2-ply evaluation gnubg will end up calculating the following cubeful 0-ply equities: centred 1-cube, opponent owns 2-cube, owned 4-cube, and opponent owns 8-cube.
Note that the 2-ply level does not use the cube efficiency, it's not used until at the 0-ply level, but it's possible to calculate an effective one by isolating x in the basic cube formulae:
x(eff) = (E(2-ply cubeful) - E(2-ply dead))/(E(2-ply live)-E(2-ply dead)).
The cube efficiency is obviously an important parameter, unfortunately there haven't been much investigation carried out, so gnubg basically uses the values 0.6-0.7 originally suggested by Rick Janowski:
| Position Class | x (Cube efficiency)
|
| Two-sided (exact) bearoff | n/a
|
| One-sided bearoff | 0.6
|
| Crashed | 0.68
|
| Contact | 0.68
|
| Race | linear interpolation between 0.6 and 0.7
|
For race gnubg uses linear interpolation based on pip count for the player on roll. A pip count of 40 gives x=0.6 and 120 gives x=0.7. If the pip count is below 40 or above 120 values of x=0.6 and x=0.7 are used, respectively.
For the two sided bearoff positions the cubeful money equity is already available from the database, so for money game there is no need to calculate cubeful equities via Janowski's formulae. However, the cubeful equities for money game cannot be used for match play. Instead of using a fixed value of x, say, 0.6, gnubg will calculate an effective value based on the cubeful money equity. The cubeful MWC is calculated as usual, but with the calculated x.
There is obviously room for improvements. For example, holding games should intuitively have a lower cube efficiency, since it's very difficult to double effectively: either it's not good enough or you've lost the market by a mile after rolling a high double or hitting a single shot. Similarly, backgames will often have a low cube efficiency, whereas blitzes have may have a higher cube efficiency.
gnubg's cube decisions are simple based on calculations of cubeful equities. For a double decision gnubg calculates the cubeful equity for "no double" and the cubeful equity for "double, take". Combined with the equity for "double, pass", it's possible to determine the correct cube action.
Figure X shows the relevant cubeful equities for O and X's cube decisions in sample position from earlier.
| Figure X Cubeful equities
|
On 0-ply X will double when the green curve (O owns 2-cube) is above the red curve (centered cube), and O will take as long as the green curve is below 1. Similarly, O will double when the blue curve (X owns 2-cube) is below the red curve (centered cube), and X takes as long as the blue curve is above -1.
Note that gnubg doesn't calculate the take point or double point explicitly. The cube decision is simply made by comparing equities from Figure X.
Janowski has developed two other models for cubeful equities. The first is a generalisation of the one used by gnubg; it introduces two cube efficiencies instead of one. Often you may see that the cube efficiencies are different for the two players, and the "refined general model" as it is named by Janowski, tries to take this into consideration by using different cube efficiency parameters for the two players. For example, the blitzer may have another cube efficiency that the blitzee.
The second model is not published, but redefines the cube efficiency into a value that can be understood more intuitively and calculate easily from rollouts.