in IMIS, if BFGS optimization fails, sample some new points rather than using the same starting points over and over...

[jeffeaton]

Or look at alternate approaches to approximating the hessian.

No benefit from optimization step if it doesn't produce a covariance to sample from.