In a system of N sensors, the sensor S-j, j=1,2,..., N, outputs Y-(j)
is an element of [0,1], according to an unknown probability density p(
j)(Y-(j)parallel to X), corresponding to input X is an element of [0,1
]. A training n-sample (X(1),Y-1),(X(2),Y-2),...,(X(n),Y-n) is given w
here Y-i=(Y-i((1)),Y-i((2)),..., Y-i((N))) such that Y-i((j)) is the o
utput of S-j in response to input X(i). The problem is to estimate a f
usion rule f:[0,1](N)-->[0,1], based on the sample, such that the expe
cted square error I(f) =integral[X-f(Y)](2)p(Y parallel to X)p(X)dY((1
)) dY((2))... dY((N))dX is minimized over a family of functions F with
uniformly bounded modulus of smoothness, where Y=(Y-(1), Y-(2),..., Y
-(N)). Let f minimize I(.) over F; f* cannot be computed since the un
derlying densities are unknown. We estimate the sample size sufficient
to ensure that Nadaraya-Watson estimator (f) over cap satisfies P[I((
f) over cap)-I(f)> epsilon]<delta for any epsilon>0 and delta, 0< del
ta<1. (C) 1997 Society of Photo-Optical Instrumentation Engineers.