What is a good way to begin a search for parameter estimates in a GLM?

Question

Following up on my answer here, I am wondering

• What is the reasoning for initialize expression of the family objects in glm in R (see ?family). glm fits generalized linear models and the initialize argument is a way to provide initial estimates of the parameters to get an iterative solution started.
• Is it "good" way that glm.fit start the first iteration be selecting the mean values for each observation? Is there a "better" way?

As an example, binomial()$initialize selects mustart <- (weights * y + 0.5)/(weights + 1) where weights (usually) are the number of either cases or controls when you use glm with a one dimensional outcome variable.