Introduction
Genetic algorithms have been widely used in vary of projects. There are thousand kinds of Genetic Algorithm in the world. However some kinds of Genetic Algorithm often solves some very narrow range of projects. And if we take this algorithm into another project, the performance will be very bad. This is because the code that GAs used and the genetic operation often have strong effect on the result. To give an example to illustrates that the codes have strong effect on performance, Grey code is a good example. My work focuses on how to design an adaptive GAs based on finding the most suitable codebook for algorithm. The most benefits from this algorithm is it doesn't have connects with fitness function directly. So it doesn't constrained by the target function. Therefore, it would have more potential to use in a very wide range.
Main Objection
1.To illustrate the interconnect between code and GAs.
2.To design a robust & efficacy algorithm to adaptive find the good code.
OverView
1. To illustrate the interconnect between code and GAs.
Code is one of most important factors, which can affect the performance of GAs. The interconnect between performance, and code has been described in some papers. If a set of code strategy is given, how to measure the performance? How an efficient strategy can be selected? It could improve the performance of GAs.
2. To design a robust & efficacy algorithm to adaptive find the good code.
In real-world projects, the most suitable code may be changed by the time. So if an adaptive change strategy can be found. It can increase the robust & efficacy of GAs.
Related Work
There are many kinds of adaptive genetic algorithms. And all of them have been widely used in many kinds of projects like JSP problems [1], Data mining of DNA [2], Forecasting [3]. In general, there are 5 kinds of adaptive Genetic Algorithm. They are as follow:
1. Fitness function based algorithm.[4][5]
2. Genetic operation based algorithm [6]
3. Knowledge based algorithm [7]
4. Hybrid based algorithm [8]
5. Code based algorithm [9]
In follow algorithms, seldom researcher focus on code based algorithm because the search space is too large, and the theory in code based algorithm is unclear. In recent years, [10] put forwards that using Tree model to change the code based algorithm.
Novel Points
In this project, we mainly focus on using linear transmission or rotational transmission to discover the hidden regular things in the code based algorithm. The novel ideas are as follow:
1. The method is designed according to Space Theory, which is a novel kind of method.
2. The method is a kind of method to adaptive control the schema, which is also a novel direction for EA&GA.
Result:
1. Overview Formal of Algorithm:
The algorithm I focus on has two kinds of style - static and dynamic. The static describe is like follow:
Note target GA: the GA which the algorithm need to adjust.
Loops untill convergence:
1. Computing fitness function store in vectors f
2. Using transform matrix to change the code of target GA
3. Applying target GA's Selection, Crossover, Mutation Operations with f .
4. inverse transform into the original code.
|
The dynamic describe is:
Note target GA: the GA which the algorithm need to adjust.
Loops untill convergence:
1. Computing fitness function store in vectors f
2. Using transform matrix to change the code of target GA
3. Applying target GA's Selection, Crossover, Mutation Operations with f .
4. Inverse transform into original code.
5. Try to find a way to select another transform matrix which is used to change the code in next term.
|
2. Vaild the Ability of Rotation:
To illustrate that linear transform can change the evolutionary strategy. I choose a special subset of linear transform - rotation transform, as example. The picture above is 30 times results of min value gotten by Different Evolutionary Algorithm with different type of 2-D Real Code. The figure shows that code actually have strong effect on the performance of GA type algorithm.
 
But it brings many extra questions.
1. the computing process is too slow.
2. the relation between anger and fitness function is uncontinous and very complex. Sometimes even become a chaos system.
3. what factors causes this phonema.
The dynamic style of this problem can solve the problem 1 but the policy used to update the anger still unknown. The problem 2 means I can't get help from some traditional method even I deduced the mathematical connection between every angers. The problem 3 required the analysis of this strange random process constructed by anger, genetic operation and fitness function, even in some cases, the both the static and dynamic algorithm become useless. So it is very difficulit in theory.
So I conclued these 3 challenges aspect which I currently should deal with.
The Challenge
Insensitive Example :
There are some type of algorithm with properly that can't change the performance. What properly these algorithms contained? Why they insensitive about the change of code?
The Strange Figures
Sometimes, the algorithm may give us a very regular figure (Like the first figure). Is it only a coincidence example? How to discover this "hidden order"?
How to find the good code?
To try every ways seems cost to much resources. To stress the ability of this algorithm, actually, we don't know the structure of this "strange space". Can we find a simple way to update the transfer matrix?
Download the Reference
Return to Research Page
|
