Data clustering, data deduction and data visualization. Using advnaced skills to encode the free format articles to cluster data by using LLM pre-trained models.
Data clustering, data deduction and data visualization. Using advnaced skills to encode the free format articles to cluster data by using LLM pre-trained models.
15. 基因演算法 (Genetic algorithms) 簡介 GA 是一種模擬「物競天擇」、「適者生存」的搜尋法則;每個物種在某個生存環境中彼此互相競爭、淘汰,只有適應性強的物種得以存活及繁衍;並透過複製、交配、突變等演化方式產生下一代的物種,如此反覆進行,最後留下適應性最強的物種。而 GA 的遊戲規則就是「適者生存」。 GA 主要是以操作染色體 (Chromosome) 來進行演化過程,在反覆演化的過程中,可看成是在問題的可行區域中做系統化的多維空間搜尋;而其特點為多點搜尋、只需適應值資訊、轉移規則是隨機性而非決定性的。 另外 GA 的搜尋方式是屬於平行式而非循序式的,這點和 Greedy 、 Tabu 、模擬退火法等的搜尋方式有很大的不同。
26. Connector 為基與 GAs 結合 2 International Journal of Production Research , 42(11), 2243-2261 (2004)
27. Focus How to turn the connector concept into genetic coding? What are the operation mechanisms of a GA? How to apply (1) and (2) in practical cases?
29. Basic idea for connector A connector example : bolt-nut-washer (b) Information on bolt-nut-washer connector
30. Classification of fastener types 3 Races and ball-bearing balls MND Not disassembled 1 snap ring, bearing, spring MD Disassembled Movable fastener 4 pressing fits, riveted joints, welding FND Not disassembled 2 Screw, bolted joint, key, spline, wedge FD Disassembled Fixed fastener level Example Code Type
31. Example: stapler (a) Part drawing (a) Part information Pivot rod 18 Guide rod 9 Rivet4 17 Bottom track 8 Rivet3 16 Staple spring 7 Fastener piece 15 狠狠撸 foot 6 Rivet bottom 14 Pivot spring 5 spring 13 Steel top 4 base 12 Rivet1 3 Impact plate 11 Bracket spring 2 Rivet2 10 Steel cover 1 Part name Part No. Part name Part No.
33. Note: (1) FD: Fixed fastener disassembled (2) FND: Fixed fastener Not disassembled (3) MD:Movable fastener disassembled (4) MND: Movable fastener Not disassembled (5) T 1 : hand (6) T 2 : screwdriver (7) T 3 : a hand vice Connector information of stapler 1,4,8,12,18 T 3 z FND Interference fit C 8 1,2,3 T 3 y FND Interference fit C 7 6,5,4 T 1 -y MD Snap fit C 6 8,9 T 1 x FND Insert C 5 6,7 T 1 -x MD Spring C 4 6,9 T 1 -x FND Insert C 3 7,9 T 1 -x MD Spring C 2 12,15,16,17 T 3 y FND Interference fit C 1 10,11,12,13,14 T 3 -y FND Interference fit C 0 Component owned by connector Tool Direction Combination type Connector name No.
37. Concept for fitness function The lower the frequency of alternation of the precedence sequence for a product, the higher the probability it will fit and survive.
39. Weight design wc + wd + wt = 1. If the result of ranking order is w1 ≥ w2 ≥ w3 ≥…≥ wn ≥ 0 , the summation of the number of weights is referred to as sum-of-ranks. the relative weight of the most important attribute will be n/(sum-of-ranks)
45. Calculate the fitness value Connector information Generate initial populations Reproduction: Roulette Wheel method PMX crossover Insert mutation method Optimal solutions obtained (1) (2) (3) (4) (5) (6) (7) No Satisfy the stopping criteria ? Generate new population number of populations, mutation rate, crossover rate, stopping criteria. (1) Yes The flow chart that combines the connector concept and the GAs
49. Result If the weight value of each engineering feature is assigned to equal (i.e., wc = wd = wt = 1/3). the connector-based assembly sequence of the computer hard disk will be as follows: {12, 7, 0, 10, 8, 5, 6, 2, 4, 1, 3, 22, 18, 17, 11, 14, 9, 15, 16, 13, 21, 19, 20} The average computation time over 10 trials at hard disk was 0.9108 second.
51. Conclusions As a comparison, the traditional heuristic algorithm developed from liaison graph, connector-based approach was another novel research direction for the assembly planning. To make this research efficient and flexible, the following features are introduced: Object-oriented programming and standard template library (STL) syntax was used to combine the connector concept and GAs; Parameter values found by experiment were significant for the quality of GA’s solution.
52. Guided-GAs 引導式基 因演算法 3 International Journal of Production Research, 44(3), 601-625 (2006)
69. 電風扇交配率為 70% 、突變率為 30% 、母體大小為 51 、最大世代數為 1500 代 的測試環境 0 137 17.3333 16.6666 Guided-GAs 6 879 17 6.5999 Traditional-GAs Times of infeasible solution Average generations of convergence Max fitness value Average fitness value Method
71. Part drawing of the laser printer 0 76.6666 75.5333 Guided-GAs 10 0 0 Traditional-GAs Times of infeasible solution Max fitness value Average fitness value Method
90. B Cell and T cell B Cell 是 B 淋巴細胞 (Lymphocyte) ,是由骨髓 (Bone Marrow) 所產生的細胞,主要是受到抗體刺激後產生反應,並分泌抗體跟抗原結合。 T Cell 也就是 T 淋巴細胞,是在胸腺 (Thymus) 中成熟,主要功能辨識外部的侵入者並加以消滅;即 B 細胞要有效產生抗體對抗外來的抗原,必須依靠 T 細胞的幫助, B 細胞及 T 細胞表面會有許多受體 (Receptor) ,當抗原經由免疫階段後會產生不同的決定部位 (Epitope) ,當兩者接觸後會發生生物化學作用,其結合程度稱之為親和力 (Affinity) 。當 Receptor 跟 Epitope 結合越緊密,也就是表示親和度越高;這就是免疫系統中的專一性特點。
91. 親和力成熟 (Affinity Maturation) 在免疫辨識之中可以得知, B 細胞經由跟抗原親和力判定,通過門檻值後就被刺激活化,這樣親和度選擇現象稱為親和力成熟 (Affinity Maturation) ,免疫系統會依此產生專一結合的細胞選殖 (Cloning) 若將抗原可定義為 Connector 為基的組裝規劃的目標式,抗體可定義為對應目標式所產生的答案,親和力是衡量 AIS 抗體間或是抗體跟抗原之間結合程度,而在 ASP 中是維持組裝順序的多樣性。
92. 發展 ASP 專屬的 AIS 演算法的幾項觀念 本研究應用 AIS 的專一性來產生初始解,而其初始解需要符合先行圖所描述的限制條件以產生可行解,再應用株落選擇中選殖 (cloning) 概念,將比較好的抗體選入記憶細胞區中,並分裂成符合題意記憶細胞區中所需個數。 本研究應用 GAs 中的 crossover 及 mutation 產生多個可行解,以達到抗體多樣化的目的。 在記憶性方面,每一次搜尋的可行皆予記錄,因應每次記憶好的抗體再跟抗原快速反應。最後最佳化的抗體再從記憶細胞區中輸出。
93. 分支度 (Outdegree) 及內分支度 (Indegree) Outdegree: 以節點 Ci 為例 , Outdegree 就是接在 Ci 之後的節點數。如果先行順序中 Ci ? Cj , Cj 為 Ci 的 Outdegree ,其 Outdegree 為 1 。 Indegree: 以節點 Ci 為例 , Indegree 就是接在 Ci 之前的節點數。如果先行順序中 Cj ? Ci , Cj 為 Ci 的 Indegree ,其 Indegree 為 1 。
100. 最佳抗體 次佳抗體 1 次佳抗體 2 相同個數 k = 6 66.7 % 相同個數 k = 5 55.6 % 100% 親和力挑選示意圖
101. Comparison between three algorithms for fan. 18.667 18.365 3.045 AIAs 18.333 18.285 3.951 Memetic Algorithms 16.667 16.133 2.808 Guided-GAs Max objective value Average objective value Average time Method
103. Comparison between three algorithms for laser printer 82.33 81.432 19.067 AIAs 80 79.096 25.965 Memetic Algorithms 76.67 75.595 18.543 Guided-GAs Max objective value Average objective value Average time Method
105. Conclusion 在 ASP 問題中,良好的初始解品質的效益將大於傳統的 immune algorithms 的做法;其次,本研究所 proposed 的 AIS 與傳統 GAs 的不同處在於應用人工免疫的多樣性特性,將可免於陷入 local optimal 的困境中。 與過去代表性的範例來作比較,以電風扇為例,發現 AIS 在平均求解時間優於 Guided-GAs 約 20%, 最大目標值優於 Guided-GAs 約 12% ,比起 MAs 改善約 2% ,而以印表機為例,可得知 AIS 在平均求解時間優於 MAs 約 26.6 % 、最大目標值優於 Guided-GAs 約 7.4% 及 MAs 約 2.9% ,進而證明 AIS 可以有效針對限制式過多,易落入局部最佳解及求解時間過長問題做有效的解決。因此,綜觀而言, AIS 可以用來解決 Tseng (2006) 應用 Guided-GAs 求解組裝規畫問題易陷入局部最佳解的問題,而在 MAs 方面 (Tseng et al. ,2007) ,求解品值差異不大,但是在求解時間卻可以有效的縮短
106. 绿色导向产物模组化之研究 Modular design to support green-life cycle engineering 7 Expert Systems with Application, 34, 2524-2537 (2008)
107. Green life-cycle engineering Fierce market competition is shortening the product life cycle. Passive resource recycling approach can no longer cope with the ever-increasing burden current products have on the environment. It is important to maximize the usage percentage of resources and minimize the damage to the environment in the early product design stage. Product life cycle refers to the total amount of time from material, manufacturing, assembly, consumer use, and final disposal or recycle of a product.
108. Motivation(Taking green life cycle into consideration) The author attempted to apply the green modular concept to product design. Advantages for this study: Reexamination of product functions ensures that the goal of environmental protection can be achieved. Products or product components can be recycled, reused and disposed of more easily. The life-cycle cost estimation enables designers to bring product cost into control.
109. Module definition Modules can be defined as the integral physical structures corresponding to specific product functions. (Otto and Woods, 2001) A proper modular design is able to cope the green objective and assemble components effectively into new products.
110. This study comprises three parts: Liaison graph is used to describe the product models, and liaison intensity of components is introduce. To assign the components whose liaison intensities are stronger in the same module, a clustering method is needed. The grouping genetic algorithm (GGA) is employed to solve the clustering problem. The evaluation of the clustering result: green pollution analysis and cost viewpoint are proposed in this study.
112. Liaison graph of a pen (a) (b) A higher LI indicates a more difficult type of combination and a smaller LI means a simpler type of combination. Liaison intensity(LI)
113. Estimate of liaison intensity among components A higher LI indicates a more difficult type of combination and a smaller LI means a simpler type of combination. In this study, a number from 0 to 100 is employed to describe the liaison intensity between components. The liaison intensity is decided by four engineering attributes of components such as the contact type, combination type, tool type, and accessed direction.
114. Table 1 Intensity of contact type Strong combination high Score Many faces will be contacted. 30 Multi-face contact Many points will be contacted. 24 Multi-point contact The contact part is a face. 18 Single face contact The contact part is a line . 12 Line contact The contact part is a point. 6 Point contact Description Liaison intensity Attribute Contact type
115. Computing intensity LIi = CTi + CBi + TLi + ADi (1) where LIi represents the total liaison intensity of the ith component; CTi represents the contact type intensity of the ith component; CBi represents the combination type intensity of the ith component; TLi represents the tool type intensity of the ith component; ADi represents the accessed direction intensity of the ith component.
116. Encoding for grouping genetic algorithms (GGA). Each gene stands for a module. For a chromosome composed of five modules “ABCDE”, the number of modules can be expressed as A={1}, B={3, 6}, C={4}, D={2}, E={5}.
117. Fitness Design A stronger Li intra indicates that it is easy to assemble components in a module
120. Green analysis Poll = Weight × Indicator Where Poll indicates the pollution value of a component; Weight denotes the weight of the component (Kg); and Indicator represents the unit pollution index of a component. The pollution value offered by Eco-indicator99 ( http: // www.pre.nl / ).
121. Cost analysis Total Cost = material cost + process cost +assembly cost Material cost = unit cost of material (NT$/Kg) × material weight (Kg) Process cost = unit cost of process (NT$/sec) × time for process (sec) Assembly cost = the unit assembly cost (NT$/sec) × assembly time (sec).
123. Table 5 Estimate liaison intensity for table lamp Create the Liaison Intensity for every components. 40 1 Angle Hand Turn on PC 13-14 40 1 Angle Hand Turn on PC 12-14 60 1 Angle Small tool type Put on PC 11-14 36 1 Angle Hand Insert PC 10-14 45 4 Angles Hand Insert MPC 9-10 30 5 Angles Hand Insert LC 8-10 26 5 Angles Hand Put on LC 7-10 54 1 Angle Small tool type Turn on PC 6-20 54 1 Angle Small tool type Turn on PC 6-19 54 1 Angle Small tool type Turn on PC 5-18 54 1 Angle Small tool type Turn on PC 5-17 54 1 Angle Small tool type Turn on PC 5-16 54 1 Angle Small tool type Turn on PC 5-15 32 5 Angles Hand Put on SFC 5-10 60 1 Angle Hand Insert MFC 3-6 54 1 Angle Small tool type Turn on PC 2-22 54 1 Angle Small tool type Turn on PC 2-21 30 5 Angles Hand Insert LC 2-8 26 5 Angles Hand Put on LC 2-7 32 5 Angles Hand Put on SFC 1-6 48 1 Angle Hand Insert SFC 1-4 32 5 Angles Hand Put on SFC 1-2 Liaison intensity Accessed direction Tool type Combination type Contact type Liaison
127. Design modification and the modular component analysis Situation 1: When the bearable green polluted value is not accepted by the designer. Designers choose new materials or designs of lower pollution values. Component 8’s polluted value is the highest (Table 6). Obviously, Component 8 (Soft_pipe) should be improved. Situation 2: With the result of clustering, designer can replace the whole module according to the bearable green polluted value. Major components for the light bulb module are Component 3 and Component 6.
128. Table 7 Material cost and process cost change of Component 8. Change the material to reduce the green polluted value. Choose the alternative material whose pollution and cost are lower. The material and process costs will be changed when the material is changed. 5.1 1.0 5.1 4.76 1.0 4.76 C pc T pcu C pcu C pc T pcu C pcu Cost Weight Unit Cost Weight Unit Process cost 0.6 0.2 3.0 0.56 0.2 2.8 C m W m C mu C m W m C mu Cost Weight Unit Cost Weight Unit Material cost After modification Before modification
129. Fig. 7. (a) Illustration for light bulb module, (b) a revised modular graph for the table lamp. (a) (b) Situation 2 A new design replace the original component 3, 6 ,19 and 20.
130. Conclusion Result a score system using the liaison graph was adopted to evaluate the liaison intensity between components. a GGA was adopted for the clustering of modules. The crossover methodology proposed by Falkenauer was adjusted for the constraints in modular classification problem. Green pollution and cost analysis was conducted to evaluate the clustering result.
131. Future work The green polluted standard should be established as references for designers. Proper decision-making strategies can be added to help evaluate the liaison intensity between components, making the algorithm more objective and flexible.
132. 应用多目标混合基因演算法 整合組裝規劃與線平衡之研究 8 International Journal of Production Research, 21(1), 5951-5977 (2008)
176. 基因局部搜尋演算法 (2/2) MOGLS 流程 步驟 1 :隨機產生新的權重組合 步驟 2 :使用競爭式法則挑選染色體 步驟 3 :進行局部搜尋 步驟 3-1 :利用局部搜尋機率 P LS 決定是否執行以下步驟 步驟 3-2 :利用鄰近解搜尋,產生相對於現在染色體解 x 的新解 y 。 步驟 3-3 :計算 x 與 y 的適應性函數,如果 y 優於 x ,則 y 取代 x 步驟 3-4 :重複步驟 3-2 與步驟 3-3 ,如果已經進行 k 次鄰近搜尋,則結束局部搜尋的步驟。 Back
177. Mass customization 9 Expert Systems with Applications 29, 913-925 (2005)
199. 總結 要將一件事作得稍為像樣,一定要忍受一些孤獨、忍受一些冷漠 、甚致一些嘲諷 。 實力是建立在一步一腳印努力之上,心性的陶養同等重要,必須建立對於作研究的興趣。 研究工作就是冒險之旅,很難預期最後的結果是什麼,要經得起失敗,不要太計較得失。郭台銘名言: 成功是最差的老師, 它只會帶給你膽怯和懦弱。 Do not work hard, work smart. 對於研究的堅持與訓練會讓學生帶入工作職場,通常研究作得好的人工作表現也會理想。 思考清楚兩年研究生涯的定位。