2001 Volume 13 Issue 5 Pages 496-505
Recently, a lot of researches for knowledge acquisition using Genetic Algorithm(GA)have been reported. Two approaches, Michigan approach and Pittsburgh approach, have been mainly used for these researches. Michigan approach, in which one rule corresponds to one chromosome, has the advantage to acquire effective each rule. However, the evaluation function for each movement of the robot has to be established and it is difficult to evaluate each rule. Pittsburgh approach evaluates a set of rules according to the movement of the robot. It has the advantage to establish the evaluation function easily. However, some of the rules can be meaningless, ineffective, or anomalous one another while the robot acquired the objective movements. The acquired rules can have low stability for change of environment. On the other hand, the mechanisms of the biological developmental process from only one egg cell have been investigated in biology. The existence of some of the controlling genes called "Homeobox Genes" in DNA has been discovered in many animals including human beings. These genes are controlling activation/inactivation of other structural genes as transcription factor in the developmental process. The biological DNA acquires the indirect design system by these homeobox genes. DNA coding method, which has been proposed by authors, is suitable for knowledge representation. This method can be thought that it is able to realize the biological structure like homeobox genes and the mechanism of the developmental process effectively. This paper introduces a part of the mechanism of the homeobox genes, inactivation of other structural genes, into the DNA coding method. In this method, each rule can inactivate the other rules according to the situation. It is expected that this method makes effective, understandable for human, and high stability rules. It is also expected that this method can make the set of rules separated into some effective roles. This paper applies this method to the acquisition of control rules for a mobile robot, and studies the effectiveness of this method showing the acquired rules by the simulation. The robot moves to goal avoiding the obstacles using inputs from the sensors. The result shows that only a few but effective and meaningful rules for the movement of the robot can be acquired by this method. The acquired rules also show they are separated into some important roles automatically. This paper also compares this method with the normal GA using a penalty into the evaluation function to reduce the number of rules. It shows the effectiveness of this method.
