1. 程式人生 > >閱讀Book:MultiObjective using Evolutionary Algorithms(9)--Bensons' Method & Value Function Method第三章完

閱讀Book:MultiObjective using Evolutionary Algorithms(9)--Bensons' Method & Value Function Method第三章完

(1)Bensons' Method

This procedure issimilar to the weighted metric approach,except that the reference solution is taken as a feasibel non-Pareto-optimal solution. 

Pareto-optimal region lies at the extreme of a feasible search space.  

                                                              

(2)Value Function Method

 

For the nature of the contours showns in this figure, the solution A, where a contour of the value function is tanfential to the pareto-Optimal front, is the preferred solution.  The metods can find only one solution at a time. By changing the parameters invoved in the value function ,different Pareto-Optimal solution can be find.

 (2 ) Goal Programming Methods

The main idea in goal programming is to find solution attain a predefined target for one or more objective function. If there exists no solution which achieves pre-specified targets in all objective functions(the user is being optimistic),the task is to find solutions which minimize deviations from the targets .

對於目標規劃方法:  主要是為 一個或者是多個目標找預定解。若是 所有的目標都沒有辦法達到預定目標則改為“找與目標偏差最小的解”。

個人覺得有一點: 通過答案做數學題,會規劃這道題的大致的結果t,然後 驗證,若是沒有滿足的 可能會允許一點誤差,就 偏離一點。

(1) single-objective

出於允許獲得的目標與t之間的誤差會有:

 

(3)Weighted Goal Programming

(1) parameter features

less-than-equal-to type goals:  βj的值為0.  對於grater-than-equal-to type goals, αj的值為0。 and for range-type goals, there exists a pair of constraints for each objective function. Uaually,the wighted α and β are fixed by the decision-maker.

 

(2)例子:

題目:  

  決策空間---------------------------------------------------------------------------------------- 目標空間

       

在jFigure35 侷限於(3.21)F1以及F2 的範圍不大於2. Target space 並不能在Objective space 沒有辦法訪問到目標空間。 故而藉助加權的方法,逐步的僅僅目標 空間。 (偏差儘量小)

(3) 綜述各種Mothods

 All of the classical algoritms described here suggest a way to convert a multi-objective opeimization problem into a single-objective problem. The weighted sum approach suggests minmizing a weighted sum of multiple objectives,the ε-constraint methods suggests optimizing one objective funvtion and use all other objectives as constraints, weighted metric methods suggests minimizing an lp metric constructed from all objectives,the value function method suggests maximizing an overall value function(or utility function) relating all objectives ,and goal programming methods suggests minimizing a weighted sum of deviations of objectives from user-specified targets.

(4)   自我總結---第三章完結

這一章節主要說明各種傳統的方法處理多目標問題。 首先是對於凸函式可以很方便的訪問到每一個front 中的點(主要是基於  )獲得切點,求解的是最小化目標值,然後根據凸函式的定義,當相切的時候總會得到最小值的點。Pareto-optimal solution)

但是艱難的就是遇見非凸函式的時候,因為他的特殊性用以前的方法有的 時候根本沒有辦法訪問到那個 點。 然後就 各種想辦法,我覺得最好的不可思議的是ε-constraint methods,將一個space空間,劃分為很多部分,逐步的將非凸轉化為區域性的凸函式,然後得到點。 但是沒有辦法保證你是適當的切割啊。 然後是將目標函式整合不管是加權矩陣,指值函式, 都是在試圖對多個目標函式進行一種操作,然後通過這個操作我們可以得到一個“最終的求解的約束的目標,滿足這個額目標我們就可以獲得pareto-optimal solution”。覺得 大都是一樣的。

然而, 通過這章節就是覺得

(1)  手算的計算量還是有的

(2)  有好多種引數變數, 無可避免的需要user 去考慮,這個要怎麼選擇?

----------------------------------------------Then next chapter -------------Evolutionary algorithms