Solving Temporal Constraints Using Neural Networks.
Gabriela Ribas Klein, Malek Mouhoub · International Conference on Artificial Intelligence · 2002
There was a resurgent in research of neural nets during the late 70’s and 80’s due to advances made in learning algorithms for feed-forward and feedback networks. These advances, coupled with better computer technology, made it possible for practical applications of such networks to be made. In 1985, John Hopfield and David Tank first attempted using neural nets as an approximation method to solve optimization problems, mainly the Traveling Salesman Problem. Since then, there has been wide spread interest in applying neural nets to solve different types of optimization problems. In this paper we will mainly focus on using the Hopfield model to solve the Maximal Temporal Constraint Satisfaction Problem (MTCSP). An MTCSP is an optimization problem that consists of looking for a solution that satisfies the maximal number of temporal constraints. This can be the case of over constrained problems involving time constraints and where a complete solution does not exist, or those problems such as real time applications where a solution is needed by a given deadline. The quality of the solution (number of satisfied constraints) depends here on the time allocated for computation. Experimental comparison study of the method we propose and based on the Hopfield model with approximation methods based on local search is reported in this paper. The method based on the Hopfield model presents better results than the other methods in the case of over-constrained problems.