Data processing: artificial intelligence – Fuzzy logic hardware – Fuzzy neural network
Reexamination Certificate
1998-04-09
2001-12-11
Davis, George B. (Department: 2122)
Data processing: artificial intelligence
Fuzzy logic hardware
Fuzzy neural network
C706S006000, C706S009000, C706S900000, C706S903000
Reexamination Certificate
active
06330553
ABSTRACT:
BACKGROUND OF THE INVENTION
This invention relates to a system for updating a fuzzy neural network, and particularly to that for obtaining fuzzy rules provided in the neural network in an autonomic manner. The invention further relates to a control system using the fuzzy neural network which is highly responsive to various state changes.
Heretofore, a fuzzy neural network, formed by combining a fuzzy inference system and a neural network, has been known to possess the advantages of both a fuzzy inference system and a neural network. In the above, the fuzzy inference system allows linguistically descriptive algorithms including obscurity, such as decision by humans, using if-then type fuzzy rules. The neural network allows regulating any input-output relationship by updating coupling coefficients using a learning function.
The aforesaid fuzzy neural network allows, for example, modifying the shapes of a membership function by using a learning method such as a back provocation method, wherein a membership function in the first-half portion of a fuzzy inference system is constructed using, for example, a sigmoid function; and the central value and the inclination of the membership function of the first-half portion, as well as output of fuzzy rules, are made to correspond to weighing values for coupling in a neural network.
In the above fuzzy neural network, when the number of fuzzy rules is too small, errors in output become large. On the other hand, when the number of fuzzy rules is too high, the probability of outputting appropriate values for input other than teacher data becomes low, i.e., decreasing adaptability. Thus, it is difficult to obtain appropriate numbers of fuzzy rules which balance the adaptability and the occurrence of errors. Conventionally, the number of fuzzy rules is determined per control object through trial and error, and thus, it takes an extremely long time to obtain the appropriate number of fuzzy rules.
Further, due to changes in the object with time or changes in the surrounding conditions, the appropriate number of fuzzy rules changes accordingly. Thus, when the appropriate number of fuzzy rules is determined through trial and error, if using circumstances change frequently or changes in an object constantly occur with time, such as in the case of controlling an engine, control cannot keep up with the changes in a timely manner, i.e., timely and satisfactory control cannot be achieved.
SUMMARY OF THE INVENTION
An objective of the present invention is to solve the above problems in fuzzy neural networks, and to provide an autonomic system for constructing a fuzzy neural network which itself obtains appropriate numbers of fuzzy rules in an autonomic manner, wherein the adaptability and the occurrence of errors are balanced. Another objective of the present invention is to provide a control system using the above fuzzy neural network.
An important aspect of the present invention is to provide an autonomic system for updating a fuzzy neural network which has layers and receives as input at least two variables each having membership functions located in the layers, and which outputs a parameter to be identified or adjusted, wherein the input-output relationship is constituted using fuzzy rules formed by a combination of the membership functions, comprising the steps of: causing the fuzzy neural network to learn a relationship between input and output of the fuzzy neural network based on an error in output, by changing coupling coefficients between adjacent layers, wherein the membership functions and the fuzzy rules are modified; judging whether a change in an error in output or in coupling coefficients is within a predetermined range; if the change is not within the predetermined range, adding to the fuzzy neural network a membership function related to at least one of the at least two variables, thereby adding fuzzy rules to the fuzzy neural network; judging whether any fuzzy rules are interpolated between the other fuzzy rules or extrapolated from the other fuzzy rules; and if interpolation or extrapolation is found in a fuzzy rule, deleting the fuzzy rule. According to the above system, even if the object has strong non-linearity, fuzzy rules can be formatted in an autonomic manner.
In the autonomic system, preferably, during operation of the object, fuzzy rule(s) is/are added when an error in output cannot be controlled within a permissible range simply by updating the coupling loads (coupling coefficients) through learning. The above system allows minimizing temporal deterioration of controllability occurring due to the addition of fuzzy rules during operation of the object.
In the autonomic system, the fuzzy neural network having plural layers is normally composed of two portions: the first-half portion and the second-half portion. Preferably, linearity of coupling loads at the second-half portion of at least three membership functions in one input direction is determined, and if linearity of coupling coefficients is established, a fuzzy rule corresponding to at least one coupling load is deleted. This system allows preventing the structure from becoming unnecessarily complex by effectively deleting unnecessary fuzzy rules in an autonomic manner.
In the above autonomic system, the system can be designed to have only a deletion function, or to have only an addition function, and can be used independently, depending on the intended use of the system. The present invention includes the above aspects.
Another important aspect of the present invention is to provide a control system for controlling an object using at least one model which is obtained by modeling at least one part of the internal structures of the object, wherein the modeling is conducted using at least a fuzzy neural network which obtains appropriate numbers of fuzzy rules in an autonomic manner by adding and deleting fuzzy rules. This system allows easily and effectively controlling an object having high non-linearity since fuzzy rules can be obtained simply and efficiently, thereby efficiently modeling the object. Thus, the present invention can be applied to various types of objects. In the control system, preferably, the fuzzy neural network(s) is/are the one defined in the aforesaid autonomic system.
In the above, when the present invention is applied to engine control as described above, the fuzzy neural network constituting the engine forward model undergoes the addition and deletion of fuzzy rules in an autonomic manner, thereby obtaining an optimal number of fuzzy rules. Thus, although the engine is an object having high non-linearity, fuzzy rules can be created automatically, and the engine forward model can be constructed easily. Further, even when errors between the forward models and the actual engine caused by the environmental changes and changes in the engine over time cannot be satisfactorily fixed solely by the learning function, the forward model can be adjusted by changing fuzzy rules in the fuzzy neural network of the model, by themselves in an autonomic way, to match the models with the actual engine. Further, in the above, in the autonomic system, which comprises at least one fuzzy neural network for controlling which actually controls the object to be controlled, the system may further comprise at least one fuzzy neural network for learning which has the same structure and coupling loads as does the at least one fuzzy neural network for controlling, and the at least one fuzzy neural network for learning undergoes learning including the addition and deletion of at least one fuzzy rule, and after completion of learning, the at least one fuzzy neural network for learning is made to function as a fuzzy neural network for controlling. The above system allows minimizing temporal deterioration of controllability occurring due to the addition of fuzzy rules during operation of the object. Further, by modeling the behaviors of air and fuel using the fuzzy neural networks, the system overcomes difficulties in modeling the behaviors of air and fuel due to factors which are
Fujime Yoko
Furuhashi Takeshi
Uchikawa Yoshiki
Yamaguchi Masashi
Davis George B.
Knobbe Martens Olson & Bear LLP
Yamaha Hatsudoki Kabushiki Kaisha
LandOfFree
Autonomic system for updating fuzzy neural network and... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Autonomic system for updating fuzzy neural network and..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Autonomic system for updating fuzzy neural network and... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2592990