شناسايي تركيب غيرمسلط عوامل كنترلي در مسيله‌ي چندپاسخه با استفاده از شبكه‌ي عصبي مصنوعي و الگوريتم ژنتيك

Determination of Non-Dominated Solutions in Multiple Response Problems using Artificial Neural Network and Genetic Algorithm

يافتن بهترين تركيب عوامل كنترلي براي بهينه‌سازي توامان چندين متغير پاسخ كه اكثراً با يكديگر در تضادند، يكي از مهم‌ترين نيازهاي مسايل صنعتي است. روش معمول براي حل اين‌گونه مسايل استفاده از رگرسيون چندجمله‌يي براي شناسايي روابط بين عوامل كنترلي و متغيرهاي پاسخ است، در حالي كه شبكه‌ي عصبي مصنوعي در حالاتي كه اين روابط پيچيده باشد قابليت مناسب‌تري از خود نشان مي‌دهند. در اين نوشتار، برخلاف حالت‌هاي به كار برده شده، متغيرهاي پاسخ به‌عنوان ورودي و عوامل كنترلي به‌عنوان خروجي شبكه‌ي عصبي مصنوعي در نظر گرفته شده‌اند تا با تركيب شبكه‌ي عصبي مصنوعي، تكنيك محدوديت جزيي و الگوريتم ژنتيك بتوان تركيبات غيرمسلط كارايي در مسيله‌ي چندپاسخه ارايه داد. قابليت روش ارايه‌شده در قالب مثال عددي بيان شده است كه نشان‌دهنده‌ي كارايي روش پيشنهادي نسبت به ساير رويكردهاي موجود است.

Simultaneous optimization of multiple response problems is an important problem in manufacturing cases. Polynomial regression is a common method for finding the relationship between controllable factors and responses. Some researchers have showed that artificial neural networks have better performance when the relationships are far too complex. In the multiple response problems, determination of non-dominated solutions is more valuable than finding only one solution as an optimum treatment, while this solution is one of the obtained non-dominated solutions. Unlike other existing research into using neural networks for multiple response problems, in the proposed method, responses are assumed as inputs, and controllable factors are assumed as targets of the neural network. This kind of input and target definition for neural networks helps us to determine non-dominated solutions by employing a neural network, an epsilon constraint technique and a genetic algorithm. The proposed method includes three major steps: 1) modeling the relation between responses and controllable factors by employing a neural network, 2) finding non-dominated solutions using an epsilon constraint and a genetic algorithm, 3) sieving solutions obtained from the last step and determining strong non-dominated solutions. For showing the efficiency of the proposed method, non-dominated solutions for a numerical example from the literature are determined by using the proposed approach. Comparing the results shows that obtained non-dominated solutions obtained by the proposed method for the example, are often better than other research results for the same example.