I'm looking at minimizing an n-dimensional likelihood function using Powell's method. Some parameter values are between 0 and 1, and others are in the thousands.

Powell's method uses Bracketmethod::bracket to initially bracket the minimum. In Linemethod::linmin, the initial bracketing values are hard-coded to be 0 and 1, which isn't appropriate for all parameter values. I've experimented with changing these values, and it definitely changes what minimum is found.

Has anyone looked at a way of coming up with more appropriate initial bracketing values based on the initial starting point?

The initial bracketing values are 0 and 1 times what's in xi. You can scale those to control the actual stepsize it starts with, if you use the second form of Powell::minimize and adjust the columns of ximat appropriately.