MODIFIED HERAULT-JUTTEN ALGORITHMS FOR BLIND SEPARATION OF SOURCES

We present several modifications of blind separation adaptive algorith ms which have significant advantages over the well-known Herault-Jutte n learning algorithm in handling ill-conditioned signals. In particula r, the proposed algorithms are more stable and converge to the correct solutions in cases where previous algorithms did not. The problem is the classical one in which several independent source signals s(j)(t)( j = 1, 2, ..., n) are linearly combined via unknown mixing coefficient s (parameters) a(ij) to form observations x(i)(t) = Sigma(j=1)(n) a(ij )s(j)(t), i = 1, 2, ..., n. The synaptic weights w(ij) of a linear sys tem (often referred to as a single-layer feedforward neural network) m ust be adapted to combine the observations x(i)(t) to form optimal est imates of the source signals S-p(t) = y(p)(t) = Sigma(i=1)(n), W(pi)x( i)(t). The optimal weights correspond to the statistical independence of the output signals y,(t) and they simultaneously ensure self-normal ization of these signals. Starting from the modified Herault-Jutten re cursive neural network model, we have derived a family of on-line adap tive learning algorithms for feedback (fully recurrent) and feedforwar d architectures. The validity and high performance of the proposed neu ral network are illustrated by simulation. (C) 1997 Academic Press.