Controlling Search Dynamics by Manipulating Energy Landscapes.

摘要

Touretzky and Hinton's DCPS (Distributed Connectionist Production System) is a neural network with complex dynamical properties. Visualization of the energy landscapes of some of its component modules leads to a better intuitive understanding of the model. Three visualization techniques are used in this paper. Analysis of the way energy landscapes change as modules interact during an annealing search suggests ways in which the search dynamics can be controlled, thereby improving the model's performance on difficult match cases. Keywords: Cognition, Psychology, Boltzmann machines, Energy functions, Simulated annealing search. (jes)