A decentralized algorithm for learning in adaptable networks

A biologically motivated algorithm is described which is ideally suited for learning and adaptation in a network of neuronlike computing elements. Each computing element communicates with its immediate neighbor and uses a simple neighbor-copying rule to improve its performance. A control node communicates with all computing nodes using ´broadcast´ messages. Learning and adaptation occur in the computing nodes and not in the control node. Once trained, the entire network collectively performs subsequent computing tasks. The author describes an implementation of the algorithm, in which each computing element is based on a neuron model and is simulated by a Unix-type process. Each process essentially maps one bit string to another. The implementation is augmented by a programming language interface that allows computational tasks to be solved by the underlying network. Simple computing tasks performed by the implementation are those which are traditionally difficult for conventional computing techniques.<>