We have proposed a neural network called the Lagrange programming neural network with polarized high-order connections (LPPH) for solving the satisfiability problem (SAT) of propositional calculus. The LPPH has gradient descent dynamics for variables and gradient ascent dynamics for Lagrange multipliers, which represent the weights of the clauses of the SAT. Each weight wr increases according to the degree of unsatisfaction of clause Cr. This causes changes in the energy landscape of the Lagrangian function, on which the values of the variables change in the gradient descent direction. It was proved that the LPPH is not trapped by any point that is not a solution of the SAT. Experimental results showed that the LPPH can find solutions faster than existing methods. In the LPPH dynamics, a function hr(x) calculates the degree of unsatisfaction of clause Cr via multiplication. However, this definition of hr(x) has a disadvantage when the number of literals in a clause is large. In the present paper, we propose a new definition of hr(x) in order to overcome this disadvantage using the "min" operator. In addition, we extend the LPPH to solve the constraint satisfaction problem (CSP). Our neural network can update all neurons simultaneously to solve the CSP. In contrast, conventional discrete methods for solving the CSP must update variables sequentially. This is advantageous for VLSI implementation.
The copyright of the original papers published on this site belongs to IEICE. Unauthorized use of the original or translated papers is prohibited. See IEICE Provisions on Copyright for details.
Copy
Takahiro NAKANO, Masahiro NAGAMATU, "A Continuous Valued Neural Network with a New Evaluation Function of Degree of Unsatisfaction for Solving CSP" in IEICE TRANSACTIONS on Information,
vol. E89-D, no. 4, pp. 1555-1562, April 2006, doi: 10.1093/ietisy/e89-d.4.1555.
Abstract: We have proposed a neural network called the Lagrange programming neural network with polarized high-order connections (LPPH) for solving the satisfiability problem (SAT) of propositional calculus. The LPPH has gradient descent dynamics for variables and gradient ascent dynamics for Lagrange multipliers, which represent the weights of the clauses of the SAT. Each weight wr increases according to the degree of unsatisfaction of clause Cr. This causes changes in the energy landscape of the Lagrangian function, on which the values of the variables change in the gradient descent direction. It was proved that the LPPH is not trapped by any point that is not a solution of the SAT. Experimental results showed that the LPPH can find solutions faster than existing methods. In the LPPH dynamics, a function hr(x) calculates the degree of unsatisfaction of clause Cr via multiplication. However, this definition of hr(x) has a disadvantage when the number of literals in a clause is large. In the present paper, we propose a new definition of hr(x) in order to overcome this disadvantage using the "min" operator. In addition, we extend the LPPH to solve the constraint satisfaction problem (CSP). Our neural network can update all neurons simultaneously to solve the CSP. In contrast, conventional discrete methods for solving the CSP must update variables sequentially. This is advantageous for VLSI implementation.
URL: https://global.ieice.org/en_transactions/information/10.1093/ietisy/e89-d.4.1555/_p
Copy
@ARTICLE{e89-d_4_1555,
author={Takahiro NAKANO, Masahiro NAGAMATU, },
journal={IEICE TRANSACTIONS on Information},
title={A Continuous Valued Neural Network with a New Evaluation Function of Degree of Unsatisfaction for Solving CSP},
year={2006},
volume={E89-D},
number={4},
pages={1555-1562},
abstract={We have proposed a neural network called the Lagrange programming neural network with polarized high-order connections (LPPH) for solving the satisfiability problem (SAT) of propositional calculus. The LPPH has gradient descent dynamics for variables and gradient ascent dynamics for Lagrange multipliers, which represent the weights of the clauses of the SAT. Each weight wr increases according to the degree of unsatisfaction of clause Cr. This causes changes in the energy landscape of the Lagrangian function, on which the values of the variables change in the gradient descent direction. It was proved that the LPPH is not trapped by any point that is not a solution of the SAT. Experimental results showed that the LPPH can find solutions faster than existing methods. In the LPPH dynamics, a function hr(x) calculates the degree of unsatisfaction of clause Cr via multiplication. However, this definition of hr(x) has a disadvantage when the number of literals in a clause is large. In the present paper, we propose a new definition of hr(x) in order to overcome this disadvantage using the "min" operator. In addition, we extend the LPPH to solve the constraint satisfaction problem (CSP). Our neural network can update all neurons simultaneously to solve the CSP. In contrast, conventional discrete methods for solving the CSP must update variables sequentially. This is advantageous for VLSI implementation.},
keywords={},
doi={10.1093/ietisy/e89-d.4.1555},
ISSN={1745-1361},
month={April},}
Copy
TY - JOUR
TI - A Continuous Valued Neural Network with a New Evaluation Function of Degree of Unsatisfaction for Solving CSP
T2 - IEICE TRANSACTIONS on Information
SP - 1555
EP - 1562
AU - Takahiro NAKANO
AU - Masahiro NAGAMATU
PY - 2006
DO - 10.1093/ietisy/e89-d.4.1555
JO - IEICE TRANSACTIONS on Information
SN - 1745-1361
VL - E89-D
IS - 4
JA - IEICE TRANSACTIONS on Information
Y1 - April 2006
AB - We have proposed a neural network called the Lagrange programming neural network with polarized high-order connections (LPPH) for solving the satisfiability problem (SAT) of propositional calculus. The LPPH has gradient descent dynamics for variables and gradient ascent dynamics for Lagrange multipliers, which represent the weights of the clauses of the SAT. Each weight wr increases according to the degree of unsatisfaction of clause Cr. This causes changes in the energy landscape of the Lagrangian function, on which the values of the variables change in the gradient descent direction. It was proved that the LPPH is not trapped by any point that is not a solution of the SAT. Experimental results showed that the LPPH can find solutions faster than existing methods. In the LPPH dynamics, a function hr(x) calculates the degree of unsatisfaction of clause Cr via multiplication. However, this definition of hr(x) has a disadvantage when the number of literals in a clause is large. In the present paper, we propose a new definition of hr(x) in order to overcome this disadvantage using the "min" operator. In addition, we extend the LPPH to solve the constraint satisfaction problem (CSP). Our neural network can update all neurons simultaneously to solve the CSP. In contrast, conventional discrete methods for solving the CSP must update variables sequentially. This is advantageous for VLSI implementation.
ER -