A novel artificial neural network approach to constraint satisfaction problems is presented. Based on information-theoretical considerations, it differs from a conventional mean-field approach in the form of the resulting free energy. The method, implemented as an annealing algorithm, is numerically explored on a testbed of K -SAT problems. The performance shows a dramatic improvement over that of a conventional mean-field approach and is comparable to that of a state-of-the-art dedicated heuristic (GSAT+walk). The real strength of the method, however, lies in its generality. With minor modifications, it is applicable to arbitrary types of discrete constraint satisfaction problems.

A feedback neural network approach to communication routing problems is developed, with emphasis on multiple shortest path problems, with several requests for transmissions between distinct start and end nodes. The basic ingredients are a set of Potts neurons for each request, with interactions designed to minimize path lengths and prevent overloading of network arcs. The topological nature of the problem is conveniently handled using a propagator matrix approach. Although the constraints are global, the algorithmic steps are based entirely on local information, facilitating distributed implementations. In the polynomially solvable single-request case, the approach reduces to a fuzzy version of the Bellman-Ford algorithm. The method is evaluated for synthetic problems of varying sizes and load levels, by comparing to exact solutions from a branch-and-bound method, or to approximate solutions from a simple heuristic. With very few exceptions, the Potts approach gives high-quality legal solutions. The computational demand scales merely as the product of the numbers of requests, nodes, and arcs.

A Potts feedback neural network approach for finding good solutions to resource allocation problems with a nonfixed topology is presented. As a target application, the airline crew scheduling problem is chosen. The topological complication is handled by means of a propagator defined in terms of Potts neurons. The approach is tested on artificial random problems tuned to resemble real-world conditions. Very good results are obtained for a variety of problem sizes. The computer time demand for the approach only grows like (number of flights) 3 . A realistic problem typically is solved within minutes, partly due to a prior reduction of the problem size, based on an analysis of the local arrival and departure structure at the single airports.

A strategy for finding approximate solutions to discrete optimization problems with inequality constraints using mean field neural networks is presented. The constraints x ≤ 0 are encoded by x⊖ ( x ) terms in the energy function. A careful treatment of the mean field approximation for the self-coupling parts of the energy is crucial, and results in an essentially parameter-free algorithm. This methodology is extensively tested on the knapsack problem of size up to 10 3 items. The algorithm scales like NM for problems with N items and M constraints. Comparisons are made with an exact branch and bound algorithm when this is computationally possible ( N ≤ 30). The quality of the neural network solutions consistently lies above 95% of the optimal ones at a significantly lower CPU expense. For the larger problem sizes the algorithm is compared with simulated annealing and a modified linear programming approach. For "nonhomogeneous" problems these produce good solutions, whereas for the more difficult "homogeneous" problems the neural approach is a winner with respect to solution quality and/or CPU time consumption. The approach is of course also applicable to other problems of similar structure, like set covering .

In a recent paper (Gislén et al. 1989) a convenient encoding and an efficient mean field algorithm for solving scheduling problems using a Potts neural network was developed and numerically explored on simplified and synthetic problems. In this work the approach is extended to realistic applications both with respect to problem complexity and size. This extension requires among other things the interaction of Potts neurons with different number of components. We analyze the corresponding linearized mean field equations with respect to estimating the phase transition temperature. Also a brief comparison with the linear programming approach is given. Testbeds consisting of generated problems within the Swedish high school system are solved efficiently with high quality solutions as results.

Rotor neurons are introduced to encode states living on the surface of a sphere in D dimensions. Such rotors can be regarded as continuous generalizations of binary (Ising) neurons. The corresponding mean field equations are derived, and phase transition properties based on linearized dynamics are given. The power of this approach is illustrated with an optimization problem—placing N identical charges on a sphere such that the overall repulsive energy is minimized. The rotor approach appears superior to other methods for this problem both with respect to solution quality and computational effort needed.