Adapting the Energy Landscape for MFA
Peter Burge, John S. Shawe-Taylor · ePrints Soton (University of Southampton) · 1995
INTRODUCTION Combinatorial optimization problems such as the Traveling Salesman Problem (TSP), Graph Bi-partitioning and Graph Colouring have been used extensively for bench marking new optimization algorithms. The challenge is to develop a general purpose algorithm that converges rapidly and yet still produces a high quality solution. When an optimization problem is mapped onto a neural network as states of nodes connected by edges, we assign the problem a `Cost' or `Energy' function. This energy is a function of the global state of the network at any one instant and the energy function itself models the problem at hand. High energy represents a state that is far from the desired solution. As we iterate through a node updating procedure, we change the states of nodes and attempt to reduce the overall energy, eventually finding a near optimal solution. The solution space for the problem can be thought of as a multi-dimensional energy landscape containing deep valleys or `local