I think the standard for cube decisions is to follow a Tesauro (2002) approach of approximating cube decisions with a function that takes cubeless probabilities and some measure of volatility (eg std dev of equity over the next roll or two). I assume this is what GNUbg does?
Has anyone looked at approximating cube decisions with a neural network? ie a different net than the cubeless evaluation nets - one that takes the cubeless probabilities as inputs, plus volatility, plus whatever other inputs you might want to include. Outputs would be "double" (you double if output value exceeds a threshold - only valid with the cube in the middle), "redouble" (only valid if you own the cube), and "take/pass" (what you do when offered the cube). Since (in principle) a neural net can approximate any nonlinear fn, it should be able to approximate the Tesauro one. But it allows for a more accurate approximation (again, in principle), plus the flexibility to add new inputs that improves the approximation. It sounds kind of plausible to me, but I don't see much mention of this approach when I google around, so I suspect it's harder than it initially sounds. _______________________________________________ Bug-gnubg mailing list [email protected] https://lists.gnu.org/mailman/listinfo/bug-gnubg
