A numerical method for minimizing a function by invariably moving downwards along a steepest path from the current position. The method does not guarantee to reach a global minimum because, like a mountaineer who tries to reach the lowest point by always going down a steepest slope from any given point and who gets trapped in a local basin separated by hills from lower valleys, the process may get trapped in a local minimum far above the global minimum. It is used in connectionism (1) and parallel distributed processing to minimize the discrepancy between the output of a network model and the desired state for a given input. Also called steepest descent. See also annealing (1), back-propagation algorithm. Compare hill climbing.