بهينه‌سازي همزمان شبكه‌ي عصبي بازگشتي براي بهبود شناسايي و مدل‌سازي ديناميك غيرخطي هواپيما

Simultaneous Optimization Of Recurrent Neural Network To Improve Identification And Modeling Of Aircraft Nonlinear Dynamics

در اين نوشتار چگونگي استفاده از الگوريتم ژنتيك در آموزش شبكه‌هاي عصبي، و به‌طور همزمان بهينه‌سازي ساختاري آن‌ها به‌منظور مدل‌سازي ديناميك غيرخطي هواپيماهايي با قابليت مانور بالا، بررسي مي‌شود. ارتباط‌هاي وزني، معماري شبكه و قوانين يادگيري از مشخصاتي هستند كه نقش بسيار مهمي در كيفيت آموزش و تعميم‌ شبكه‌هاي عصبي براي مدل‌سازي سيستم‌هاي غيرخطي ايفا مي‌كنند. لذا تنظيم درست اين پارامترها كمك شاياني به بهبود قابليت تعميم‌دهي شبكه‌ي عصبي آموزش ديده مي‌‌كند. در اين كار از روش الگوريتم ژنتيك عادي و اصلاح شده در كنار ساختارهايي متفاوت از شبكه‌هاي عصبي براي اين منظور، استفاده شده است. اعتباربخشي روش از طريق مقايسه و ارزيابي نتايج تحليلي با داده‌هاي تجربي حاصل از آزمايش پرواز يك هواپيماي جنگنده‌ي نسل چهارم صورت پذيرفته است. بررسي‌ها نشان‌گر دقت بالاي روش در مدل‌سازي ديناميك هواپيماست.

In this paper, using the modified genetic algorithm (MGA) as an optimization method and combining it with neural networks, the nonlinear dynamics of a highly maneuverable aircraft has been modeled. Generalization has long been considered a dilemma in dynamic system identification, especially for dynamic systems with various possible inputs, like aerospace vehicles. Therefore, the focus of this paper is to obtain methods for improving generalization of neural network based aircraft models that are going to be used in aircraft simulators. Weighted connections, network architecture, and learning rules are some features that play key roles in the quality of neural network training and generalizability, in order to model nonlinear systems. Furthermore, this paper seeks to particularly focus on applying evolutionary methods to optimize the parameters of recurrent neural networks, in order to improve the identification and modeling of aircraft nonlinear dynamics. The proposed method in this study is to apply MGA. In original genetic algorithms, genetic operators are regular seeded selection, elitism, random selection, crossover, and mutation, and the appropriate fitness function is the inversed mean squared error between the network output and target. In MGA, a new operator called mutation 2 is used. This operator randomly nullifies some weights and rules them out with a very small probability. Moreover, a penalty on non-zero weights (C) must be included in the fitness function to encourage the algorithm to reach a structure with a minimum number of connections. MGA improves generalization through zeroing unnecessary weights (or connections). MGA can additionally be used to simultaneously train and optimize three different types of recurrent neural network. To further validate this study, the reported results were compared with the recorded experimentally obtained data from a fourth-generation fighter aircraft. In conclusion, the results of training with the two methods applied in this study (modified and original genetic algorithms) clearly show that the simultaneous optimization of network architecture improves neural network generalization. This, of course, imposes a cost of longer computation time and an increased number of required generations to reach the desired mean squared error for in-sample data. Thus, the optimal network has better performance in the identification and modeling of aircraft nonlinear dynamics.