Orhan Dengiz

A tabu search algorithm for the training of neural networks

The most widely used training algorithm of neural networks (NNs) is back propagation (BP), a gradient-based technique that requires significant computational effort. Metaheuristic search techniques such as genetic algorithms, tabu search (TS) and simulated annealing have been recently used to cope with major shortcomings of BP such as the tendency to converge to a local optimal and a slow convergence rate. In this paper, an efficient TS algorithm employing different strategies to provide a balance between intensification and diversification is proposed for the training of NNs. The proposed algorithm is compared with other metaheuristic techniques found in literature using published test problems, and found to outperform them in the majority of the test cases.

Grain boundary detection in microstructure images using computational intelligence

Two computational intelligence approaches, a fuzzy logic algorithm and a neural network (NN) algorithm, for grain boundary detection in images of superalloy steel microstructure during sintering are presented in this paper. The images are obtained from an optical microscope and are quite noisy, which adversely affects the performance of common image processing tools. The only known way to accurately determine the grain boundaries is digitizing by hand. This is a very time-consuming process, causes operator fatigue, and it is prone to human errors and inconsistency. An automated system is therefore needed to complete as much work as possible and we consider a fuzzy approach and a neural approach. Both methods performed better than the widely available standard image processing tools with the neural approach superior on images similar to those trained while the fuzzy approach showed more tolerance of disparate images.

Optimization of manufacturing systems using a neural network metamodel with a new training approach

In this study, two manufacturing systems, a kanban-controlled system and a multi-stage, multi-server production line in a diamond tool production system, are optimized utilizing neural network metamodels (tst_NNM) trained via tabu search (TS) which was developed previously by the authors. The most widely used training algorithm for neural networks has been back propagation which is based on a gradient technique that requires significant computational effort. To deal with the major shortcomings of back propagation (BP) such as the tendency to converge to a local optimal and a slow convergence rate, the TS metaheuristic method is used for the training of artificial neural networks to improve the performance of the metamodelling approach. The metamodels are analysed based on their ability to predict simulation results versus traditional neural network metamodels that have been trained by BP algorithm (bp_NNM). Computational results show that tst_NNM is superior to bp_NNM for both of the manufacturing systems.

Two-stage data mining for flaw identification in ceramics manufacture

Advanced ceramics are commonly manufactured by sintering high-purity powders. The design of ceramic elements is governed by its fracture strength, which is greatly influenced by microstructural flaws. Three ceramic powder preparation methods for ceramics manufacturing are considered in this paper—uniaxial pressing followed by isostatic pressing, flocculated slip casting, and dispersed slip casting. Their effects on the growth and characteristics of microstructure flaws and damage on the ceramic surface are investigated using a two-stage data-mining approach. In the first stage, digital microstructural images are mined to characterize the flaws and surface damage. In the second stage, an extreme value probability distribution is fitted using the information from stage 1. The extreme value distribution estimates large flaws which are highly correlated with subsequent fractures. Results of the two-stage data mining show that ceramic production method significantly affects flaw characteristics that, in turn, determine the ceramics’ fracture strength.

Improving Network Connectivity in Ad Hoc Networks Using Particle Swarm Optimization and Agents

In a mobile ad hoc network (MANET) the nodes can serve as both routers and hosts and can forward packets on behalf of other nodes in the network. This functionality allows the MANET to form an instant, autonomous telecommunication network without an existing infrastructure or a central network control. This chapter introduces a dynamic MANET management system to improve network connectivity by using controlled network nodes called agents. Agents have predefined wireless communication capabilities similar to the other nodes in the MANET. However, the agents’ movements, and thus their locations, are dynamically determined to optimize network connectivity. A particle swarm optimization (PSO) algorithm is used to choose optimal locations of the agents during each time step of network operation.

Topsis based taguchi method for multi-response simulation optimization of flexible manufacturing system

This study presents a simulation design and analysis case study of a flexible manufacturing system (FMS) considering a multi-response simulation optimization using TOPSIS (Technique for Order Preference by Similarity to Ideal Solution) based Taguchi approach. While in order to reduce expensive simulation experiments with the Taguchi design, the TOPSIS procedure is used to combine the multiple FMS responses (performance measures) into a single response in the optimization processes. Thus, TOPSIS carries out an important role to build a surrogate objective function that represents multiple responses of the system. The integrated approach finds a new design considering discrete factors (physical and operational parameters) which affect the performance measures of FMS. Optimal design configuration is obtained for the considered system with improved performance.

Non-deterministic decoding with memory to enhance precision in binary-coded genetic algorithms

A non-deterministic decoding algorithm for binary coded genetic algorithms is presented. The proposed algorithm enhances the precision of the GA solutions by introducing a Gaussian perturbation to the decoding function. This non-deterministic decoding enables individuals to represent any point in the continuum instead of finite discrete points. As the generations evolve, information gathered from the most fit members is continuously used to rearrange the binary representation grid on the search space, thus establishing a search memory such that the best known individual is always positioned at the center of the Gaussian offset.

Neural Computing Approach to Shape Change Estimation in Hot Isostatic Pressing

A neural network approach is presented for the estimation of shape change during a Hot Isostatic Pressing (HIP) process of nickel-based superalloys for near net-shape manufacture. For the HIP process, shrinkage must be estimated accurately; otherwise, the finished piece will need excessive machining and expensive nickelbased alloy powder will be wasted (if overestimated) or the part will be scrapped (if underestimated). Estimating shape change has been a very difficult task in the powder metallurgy industry and approaches range from rules of thumb to sophisticated finite element models. However, the industry still lacks a reliable and general way to accurately estimate final shape. This paper demonstrates that a neural network approach is promising to estimate post-HIP dimensions from a combination of pre-HIP dimensions, powder characteristics and processing information.