新型之基因演算法與智慧演化於光子學之應用

Abstract

本論文針對頻率解析光閘此超快雷射量測技術所產生的雷射圖樣，及在光纖布來格光柵設計中所需之光學參數，建立了一個基因演算法。在以傳統基因演算法演化一些例子後，我們發現其演化的速度及準確性均不盡理想，因而發展出一個新型之基因演算法的架構及一些新的演化法則，展示在有頻率-時間結構之頻率解析光閘圖形及擁有特定反射圖樣的布來格光柵之設計的反向求解問題中，改善基因演算的速度及準確度。此一新的架構模擬自一些生物體的生命循環，如一種真菌，黑黴菌。它的生命循環同時具有有性生殖及無性生殖，這似乎是它能蓬勃生長且難以消滅的原因。經過適當地安排基因演算法中的族群，以使它分成兩群染色體後，子代的產生經由不同的生殖方式，即有性生殖及無性生殖。而此二生殖方式的分別，在於父母染色體中是否有基因交換。以此觀念而建立之新基因演算法與傳統的的基因演算法需作一比較，經過用同樣的圖樣放於兩者，新型之基因演算法已被證實能較快且較為準確的演化出結果。特別是在為了設計布來格光柵的演化試驗中，相同於之前文獻的一些反射率圖樣，被放入新基因演算法中，其最後之演化結果讓誤差值達到零值，這相較於該文獻中10-5~10-6 的誤差，明顯地改善許多。而若相同的誤差要求被用於新基因演算及文獻中之基因演算法，我們發現新基因演算法會節省約一半的時間。 我們更進一部將機器學習技術及專家系統用於此新基因演算法，在知識庫的幫助下，從演化而得的經驗被儲存起來，並用於對新的演化要求提出建言，特別是會針對所要求演化之超快雷射圖樣或反射率圖樣，建議演化的起始點，因而使新基因演算法的速度及準確獲得進一步的改良。知識庫的建立是透過對雷射圖樣及反射率圖形作特徵萃取，並將這些特徵及相對應的光學參數同時放於知識庫中，在演化開始時才聯於基因演算法。當演化開始時，待演化之圖樣要與知識庫內的所有圖樣做近似度的比較，以決定哪些圖樣適於形成起始族群。這些智慧系統能自我建立以適應問題領域，也能在情況需要時自我擴充。如此，一次又一次的演化，將使得基因演算法變得越來越聰明。 我們也從粒腺體染色體在演化中傳遞並攜帶演化訊息的生物現象中，模擬而發展出一個基因演算法的探勘技術。藉由物件導向技術，訊息串被當作一個物件特性而加於基因演算法的物件中。此訊息串記載在所有世代中的交配及突變過程，它攜帶演化訊息並允許為生存下來的所有後代，推算歷代的祖先。使用此探勘技術，可以觀察到所有演化過程的細節，也能根據其記錄來調整基因演算法至其最佳狀態。

We are able to inverse both the ultrafast laser traces from the measurement technique, the frequency-resolved optical gating (FROG) and the reflectivity shapes in optical fiber Bragg gratings (FBGs) by establishing a genetic algorithms (GA). After evolving the samples by traditional GA, it is found that the convergent speed is slow and the accuracy is also unsatisfactory. To improve the performance of speed and accuracy in the GA evolution, a new structure and some evolution rules have been developed and used to demonstrate the inverse problems in both the FROG traces with frequency-time structure and the reflectivity spectrum shapes in FBG design. This new structure is imitated from the life cycle of some organisms like the mold, one kind of fungi. The coexisting character of the sexual and asexual reproduction seems to make it blossom and hard to beat. After arranging the population in GAs properly to have two groups of chromosomes and to reproduce offspring in different ways, the sexual and asexual are implemented and discerned by with and without gene exchanges. The GA with this structure is observed to evolve faster and more precise as compared to traditional GA in evolving the same examples. In the evolution tests of FBG design, our GA evolves the same examples in the literature to the error of zero that is far better than the error of level 10-5~10-6 recorded. It also saves about half of the consumed time for the traditional GA to converge to the same error level. The machine learning technique and the expert system have been adapted to the GA evolution too. The performance both in speed and accuracy is further improved after a knowledge base (KB) is set to store the experiences from evolution and to provide suggestion for evolution start points. The FROG trace and the FBG reflectivity shapes are retrieved to get their features for the KB. The KB with the information of the features and their corresponding parameters are connected to GA when the evolution starts. A similarity comparison is taken to determine which samples in the KB are the better to form the initial population for GA to evolve. These intelligent systems can build themselves to suit the problem domains and expand themselves when needed. The GA becomes wiser after more and more evolutions are applied. A probing technique is developed imitating another biological phenomenon in mtDNA that transfers with the evolution to carry the evolution information. By the object-oriented technique, the GA object is adapted by the information string that can record all the mutations and crossovers of all generations. This technique records the evolution information and allows tracing all the ancestors for every survived offspring. The user can observe all the detail evolution processes and is able to tune the GA setting according to the records.