Learning Static Evaluation Functions by Linear Regression
We present a technique for learning the coefficients of a linear static evaluation function for two-person games based on playing experience. This is accomplished by using linear regression to modify the coefficients based on the difference between the static evaluation of a state and the value returned by a mini-max look-ahead search. In an initial experiment, the technique was used to learn relative weights for the different chess pieces.
Unable to display preview. Download preview PDF.