Expected improvement acquisition function with some trickery
 to overcome underflow when very low.  Specifically, we
 revert to using log(UCB) with v=2 as the acquisition function
 if log(EI/exp(UCB))<-8.

Accepts:
  xi - option for expected improvement (scalar)
  mu-best - best posterior mean evaluated at old data points (scalar)
  gp-predicter - function that returns [mu var]
                 from the weighted gp ensemble posterior given
                 a single point to evaluate x.  See
                 create-weighted-gp-predict-with-derivatives-function*
  x* - new point to evaluate (D length vector)

Returns:
  EI - log value of expected improvement at x*

Calculates the integrated acquisition function from a base
 acquisition function and a weighted sum of GP predictors.
 This is as per Snoek et al, NIPS 2012.

Accepts:
  base-aq-func - Function which takes as inputs a gp-predictor
                 and a point to evaluate, returning a utility of
                 evaluating that point.
  mu-bests - The best of the expected best point previously
             evaluated for each GP, mu_j^+ in the paper.
  gp-predicters - A collection of gp-prediction functions.  Each
                  takes a point to evaluate an returns [mu sig]
  gp-weights - Vector of weights for each gp-predictor.  The
               acquisition function values will be weighted by
               these weights
  x* - The point to evaluate (D length vector)
Returns:
  The integrated acquistion function at x*