Output functions for probabilistic logic nodes

PLN nets consist of RAM-based nodes which can learn any function of their binary inputs; they require only global error signals during training, and they have been shown to solve problems significantly faster than nets learning by error back-propagation. Output functions for PLNs may be probabilistic, linear or sigmoidal in nature; this paper deals with designing an output function which yields fastest convergence. Experiments with several small problems support the values derived. Choice of an appropriate output function is suggested to be highly problem-dependent, but heuristics for this selection are outlined.