01/01/1998 09:00 AM

Computer Science

This work investigates some refinements of the exponentiated gradient algorithm, a recent mixture-solution method. The EG_n algorithm uses a relative-entropy distance function as a penalty term inside a gradient-assent framework to learn the values of the parameters of the log likelihood of a given data sample. These parameters include a mixture vector and the mean vectors and covariance matrices of the axis-parallel or spherical Gaussian distributions that comprise the mixture model.