ABSTRACT
A novel scheme for combinatorial optimization is suggested via analog constrained optimization. The constraints are separated into two parts. One is relieved by Lagrange approach, and the other by penalty or barrier function. In comparison with the existing Hopfield type networks, the new scheme has several new features. It is applicable to nonquadratic energy functions, and has reduced needs on externally controlling artificial weighting parameters. It also provides new potential for improving convergence and performance. In addition, it supplies a general combinatorial optimization scheme with its special cases being closely related to statistical physics method, EM algorithm, elastic net and robust PCA.
