thesis on eval function learning in Arimaa
Posted: Fri Dec 04, 2015 3:46 pm
http://arimaa.com/arimaa/papers/ThomasH ... Arimaa.pdf
Basically this used supervised learning against a database of strong server games (something like the MMTO method used in Shogi). The eval function is modelled as a linear or quadradic function. But even in the linear case it appears that the weights ultimately control a polynomal function (not sure I am understanding that on a quick read).
The author used either a linear programming solver (Mosek) or a quadratic programming solver (Gurobi) to do the optimization. These are expensive commercial solvers (but free for academic use).
--Jon
Basically this used supervised learning against a database of strong server games (something like the MMTO method used in Shogi). The eval function is modelled as a linear or quadradic function. But even in the linear case it appears that the weights ultimately control a polynomal function (not sure I am understanding that on a quick read).
The author used either a linear programming solver (Mosek) or a quadratic programming solver (Gurobi) to do the optimization. These are expensive commercial solvers (but free for academic use).
--Jon