Hardware implementation of hybrid intelligent systems based on FPGA

مقدم أطروحة جامعية

Hasan, Sundus Dahham

مشرف أطروحة جامعية

Akkar, Hanan Abd al-Rida

الجامعة

الجامعة التكنولوجية

الكلية

-

القسم الأكاديمي

قسم الهندسة الكهربائية

دولة الجامعة

العراق

الدرجة العلمية

دكتوراه

تاريخ الدرجة العلمية

2010

الملخص الإنجليزي

The most familiar technique in Artificial Neural Network (ANN) learning is called back propagation (Bp) algorithm.

Bp is widely used to solve many real world problems using the concept of Multi-Layer Perceptron (MLP) training and testing.

However the major disadvantages of Bp are its relatively slow convergence rate and being trapped at local minima.

One of the more intriguing possibilities is that of combining a neural network with other Artificial Intelligent (AI) systems, like Genetic Algorithm (GA) and Particle Swarm Optimization (PSO), to enhance the learning process in terms of convergence rate and classification accuracy.

In this work two programs called Particle Swarm Optimization-feed forward Neural Network (PSONN) and Genetic Algorithm-Neural Network (GANN) are proposed.

The two programs have been tested by three examples for pattern classification and their results are compared with each other and with Bp algorithm results for each example.

The simulation of GANN systems and PSONN algorithms are implemented using MATLAB (V-6.5) program.

Tests show that GANN is faster than PSONN but it has less accuracy.

Both algorithms have equal effectiveness but superior efficiency for PSONN over GANN.

For the correct classification percentage, they show that PSONN result is better than GANN with (93.2 %, 96.24 % and 92.83 %) compared to (99.98 %, 99.99 % and 98.42 %) for each example respectively.

GANN significantly reduces the error at small number of iteration (5, 8 and 8) compared to PSONN (23, 58 and 116) for each example respectively.

For overall performance, the experiments show that both algorithms produce feasible results in terms of convergence time and classification percentage compared with Bp algorithm.

The tests show that an ANN designed using GA has better generalization ability and smaller number of iteration for training than one trained by Bp using a human designed architecture.

Field Programmable Gate Arrays (FPGAs) have been used to implement ANN trained by GA and PSO, because of their speed benefits, as well as the re-programmability of the FPGAs which can support the reconfiguration necessary to program a neural network.

Hardware Design of ANN platform (HDANN) is proposed to evolve the architecture ANN circuits using FPGA-spartan3 board (XSA-3S1000 Board).

The HDANN design platform creates ANN design files using WebPACKTM ISE 9.2i program, which are converted into device-dependent programming files for eventual downloading into an FPGA device by using GXSLOAD program from the XSTOOLS programs.

التخصصات الرئيسية

تكنولوجيا المعلومات وعلم الحاسوب

الموضوعات

• الذكاء الاصطناعي
• أساليب المحاكاة
• معالجة البيانات
• الآلات المنطقية
• الشبكات العصبية(الحاسبات الإلكترونية)
• هندسة الإلكترونيات الحيوية

• APA
• MLA
• AMA

لغة النص

الإنجليزية

نوع البيانات

رسائل جامعية

رقم السجل

BIM-305034