Practical Neural Network Recipies in C++

Optimizing the

Decision Threshold

399

if (xnew > 1.0) // Bound to legal domain

xnew = 1.0 ;

if (xnew < 0.0)

xnew = 0.0 ;

ynew = parzen_crit ( nO , n1 , sO , s1 , wtO , wt1 , sigma , xnew ) ;

if (fabs(ynew) < bestcrit) { // Keep track of best

bestthresh = xnew ; // as secant method may

bestcrit = fabs(ynew) ; // not necessarily converge!

}

}

return bestthresh ;

}

Little has been said regarding the choice of σ, the window-

width parameter. Larger values result in loss of detail but more

immunity to sampling error. On page 379, where the subject was

confidences, it was suggested that σ be somewhere around 0.02 to 0.05.

This is good when importan ...