Artificial neural networks and their applications to manufacturing scheduling problems

Neural networks and their applications to manufacturing scheduling problems are presented in this dissertation. Most of the scheduling problems are NP-complete problems which have no existing polynomial time complexity solutions. Basically, human decisions still dominate scheduling processes even in an automated environment. Artificial neural networks are proposed to be decision aids to solve scheduling problems. Bidirectional Associative Memory (BAM) is used as a simple paradigm to show how a neural network works and how we many improve a neural network. There are two concepts introduced in this dissertation to improve BAM capacity. One is called multiple training, the other one is called dummy augmentation. Multiple training, which puts different weights for different training pairs, is a very intuitive way of improving a neural network. Linear programming methods can be used to find the weights of the training pairs to achieve the highest capacity under the Hebbian rule. Dummy augmentation method uses different augmented vectors for different pairs to increase the Hamming distance between two pairs (i.e. to make each pair more different from the others), such that all pairs can be recalled. The capacity study in chapter three shows the reason why multiple training can improve the capacity of the BAM. The multiple training concept is extended to back propagation neural networks. By putting different weight for different patterns and applying Basic Differential Multiplier Method (BDMM), simulations show that the new algorithm, Multiple Training Back Propagation Algorithm, can learn patterns faster than the original back propagation algorithm. A neural network approach is introduced to solve a manufacturing scheduling problem. A manufacturing scheduling problem includes routing and sequencing. In this approach, the routing problem is solved by a computer algorithm based upon lightest load assignment strategy. A mapping style neural network, back propagation net for instance, is used to select a nondelay strategy among a set of nondelay sequencing strategies for solving the sequencing scheduling problem. A computer simulation shows that this approach has better performance than that of using a single strategy. Possible future research directions are addressed in the conclusions chapter.