Include prior info in penalty with Bayesian estimation.

This can improve a bit optimization routines when parameter go beyond prior bounds during line search algorithms or when numerical gradient is computed.