BMRM
----

Configuration options description:
1.  double  EPSILON_INSENSITIVE.epsilon
    - The hyper-parameter i.e. epsilon, for the use in the loss
      function EPSILON_INSENSITIVE only.

2.  string  WTA_MULTICLASS.marginScalingFunctionType
    - The type of margin scaling functions. Applicable to
      WTA_MULTICLASS loss only. Two possible choices:
      a. SLACK_RESCALING
         - See [ref 1] for further description.
      b. MARGIN_RESCALING  
         - See [ref 2] for further description.

3.  string  WTA_MULTICLASS.labelLossType
    - The type of label loss function. Applicable to WTA_MULTICLASS
      loss only. Three possible choices:
       a. ZERO_ONE
          - Return 1 if actual label and predicted label differ, 0
            otherwise.
       b. ZERO_ONE_HUNDRED
          - Return 100 if actual label and predicted label differ, 0
            otherwise.
       c. SQUARED_DIFFERENCE
          - Return the square of the absolute difference of actual and
            predicted label.

4.  int    NDCG.truncation
    - The maximum number of top documents will be considered in the
      evaluation of loss. Applicable to NDCG_RANK loss only. Valid
      values must be > 1.

5.  double  QUANTILE_REGRESSION.tau
    - The hyper-parameter i.e. tau, for the use in the loss
      function QUANTILE_REGRESSION only.

6.  double  NOVELTY_DETECTION.rho
    - The hyper-parameter i.e. rho, for the use in the loss
      function NOVELTY_DETECTION only.

7.  double  HUBER_HINGE.h
    - The hyper-parameter i.e. h, for the use in the loss
      function HUBER_HINGE only.

References:
1. C. H. Teo, S. V. N. Vishwanathan, A. J. Smola, Q. V. Le
   Bundle Method for Regularized Risk Minimization
   submitted to JMLR, 2009.


