JP2002007998A - Interactive optimizing device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an interactive optimizing device capable of reducing fatigue of a worker by shortening working time for interactively optimizing a system. SOLUTION: An output characteristic of an object system 1 varies according to a characteristic parameter of n-dimension. An estimating part 2 outputs an estimated value according to the characteristic of the object system 1. The characteristic parameter given to the object system 1 is generated at an optimizing part 3, and the generated characteristic parameter is shown in a two-dimensional space of a second interface 42 through a mapping part 5. A person M can input a point in the two-dimensional space of the second interface 42 and the mapping part 5 generates an inverse map to the n-dimensional space from the point. The inverse map generated by the mapping part 5 is given to the optimizing part 3 as a solution candidate through an optimum solution addition part 6. An optimum solution is selected from the solution candidate generated on the basis of the estimated value at the optimizing part 3 and the solution candidate generated at the optimum solution addition part 6 and is applied to the object system 1 as the characteristic parameter.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to the generation of outputs intended to stimulate human senses, such as sight, hearing, and force, such as computer graphics, live-action video, music, voice, and animation. TECHNICAL FIELD The present invention relates to an interactive optimizing apparatus for interactively optimizing characteristic parameters of a system.

[0002]

2. Description of the Related Art In general, in a simple equipment system, optimization is performed by trial and error based on an evaluation criterion based on an operator's experience in designing and adjusting the system. Even workers vary from day to day, and individual differences also occur between workers. Therefore, obtaining reproducible results requires a lot of experience and a high degree of skill. For example, tuning of musical instruments is often performed only by the experience of an operator and trial and error.

On the other hand, various techniques have been proposed for automatically optimizing the characteristics of an equipment system when designing or adjusting the equipment system. As a technique of this type, a required physical quantity in an equipment system is used as an evaluation criterion, and an error between an output value obtained when the equipment system is operating and a preset target value is obtained, and the error is minimized. The criterion of optimizing the characteristics of the device system (hereinafter referred to as “minimum error criterion”) has been widely adopted. This kind of control is performed in a general servo system.

On the other hand, in designing and adjusting various equipment systems, it is often necessary to use not only physical quantities as evaluation criteria but also human subjectivity (mainly satisfaction) as evaluation criteria. For example, if a hearing aid is taken as an example of a device system, the final evaluation criterion is not only the adjustment of physical quantities such as the frequency range and the degree of amplification, but also whether or not the user can easily hear. In addition to equipment systems, in the field of arts and the like, human sensitivity is the main evaluation criterion. For example, computer graphics (hereafter,
In the design by “CG”, the evaluation criterion is whether or not the design conforms to the concept intended by the designer.

As described above, various systems for generating an output for the purpose of stimulating human sensory organs (hereinafter, the term "system" is used for equipment systems, graphic arts, industrial design, music, etc.). ), The system cannot be optimized without the use of subjective evaluation criteria. Therefore, it is difficult to automatically optimize the system without human intervention like the minimum error criterion. Impossible.

In view of the above, various techniques have been proposed in the past for human intervention when optimizing the system. This type of technology is realized by a device that works interactively using a computer, and this type of device is called an interactive optimization device.

The simplest interactive optimizer is one that optimizes a target system based on an optimization technique. This interactive optimizer basically optimizes the characteristic parameters of the target system (parameters that determine the output characteristics of the target system) by an optimization method using an evaluation function or the like that sequentially evaluates the performance of the target system. It is. This interactive optimization device terminates optimization while monitoring the output of the target system on the screen of the display as an interface, and optimizes the part involving humans by monitoring the screen of the display. Only the point instructing the end of

FIG. 5 shows an example of this type of interactive optimization apparatus. The output characteristic of the target system 1 changes according to n characteristic parameters (that is, n-dimensional characteristic parameters), and the output of the target system 1 is evaluated by the evaluation unit 2. The evaluation unit 2 evaluates the output of the target system 1 using an appropriate evaluation function or the like.
The optimization unit 3 generates an n-dimensional characteristic parameter by applying an optimization method based on the evaluation value which is the output value of the above. By applying the characteristic parameters generated in this way to the target system 1, the target system 1 is optimized. The output of the target system 1 is presented to the person M through the first interface 41. Hereinafter, "interface" means a man-machine interface that includes a display device and an input device.

In the interactive optimizing apparatus shown in FIG. 5, the optimizing unit 3 first generates n-dimensional characteristic parameters. When information of the target system 1 is not given to the optimization unit 3 when the optimization unit 3 generates the characteristic parameters for the first time, a method of generating the characteristic parameters by random numbers is generally adopted. When the n-dimensional characteristic parameters generated by the optimization unit 3 are applied to the target system 1, the characteristics of the target system 1 are determined, and the target system 1
Output is also determined. The output of the target system 1 is evaluated by the evaluation unit 2, and the evaluation value from the evaluation unit 2 is fed back to the optimization unit 3. At this time, the output of the target system 1 is simultaneously presented to the user through the first interface 41. In the optimization unit 3 to which the evaluation value from the evaluation unit 2 is passed, the optimization of the target system 1 is performed in a direction to improve the evaluation value.
Generates n-dimensional characteristic parameters to the target system 1
Apply to By repeating such processing, the performance of the target system 1 can be improved. Here, the loop between the application of the characteristic parameters to the target system 1, the output of the evaluation values by the evaluation unit 2, and the generation of the characteristic parameters by the optimization unit 3 is performed by the person M via the first interface 41. 1 is monitored and repeated until it is determined that the required performance has been obtained. As is clear from the above description, in the interactive optimization apparatus shown in FIG. 5, the person M is involved only in the determination of the end of the optimization.

[0010] On the other hand, there has been proposed an interactive optimization apparatus in which a person actively participates in the evaluation. As an example of this type of interactive optimization device, there is a technology to which a technology called an interactive evolution calculation method is applied. For example, Hideyuki Takagi, Tatsuo Unemi, Takao Terano: Research Trends of Interactive Evolutionary Computation, Journal of the Japanese Society for Artificial Intelligence, Vol. 13, No. 5, pp. 692-703 (19
98), an application example of interactive evolutionary computation
The outline is roughly divided into the engineering field and the education / game field. Interactive evolutionary computation is a generic term for engineering optimization techniques inspired by biological evolution.
In particular, the genetic algorithm has been published in many manuals and specialized books, and has recently been applied to air conditioner control and the like.

FIG. 6 illustrates an interactive optimization apparatus using this technique. This interactive optimizing apparatus has a configuration in which the evaluation unit 2 is omitted as compared with the interactive optimizing apparatus shown in FIG. That is, the target system 1 is the optimization unit 3
And the output of the target system 1 is configured to be monitored by the person M through the third interface 43. However,
The first interface 41 has only a function of receiving an instruction at the end point of the optimization by the person M other than the function of presenting the output of the target system 1 to the person M,
The third interface 43 has a function of presenting the output of the target system 1 to the user and a function of receiving the evaluation by the person M and returning the evaluation value to the optimization unit 3.

In the interactive optimizing apparatus shown in FIG. 6, first, as in the interactive optimizing apparatus shown in FIG.
Generate dimensional characteristic parameters. When the optimization unit 3 generates characteristic parameters for the first time, a method of generating characteristic parameters by random numbers is generally adopted. When the n-dimensional characteristic parameter generated by the optimization unit 3 is applied to the target system 1, the characteristics of the target system 1 are determined, and the output of the target system 1 is also determined. In the configuration shown in FIG. 6, the output of the target system 1 is presented to the person M via the third interface 43. The person M evaluates the performance of the target system 1 on the basis of the presented output of the target system 1 based on sensitivity and knowledge, and returns an evaluation value to the optimization unit 3 through the third interface 43. That is, the person M subjectively evaluates the output of the target system 1 and evaluates it (hereinafter, referred to as an evaluation value).
"Subjective evaluation value" is input to the third interface 43, and the third interface 43 transfers the subjective evaluation value to the optimizing unit 3. The optimization unit 43 generates an n-dimensional characteristic parameter for the target system 1 based on the subjective evaluation value from the third interface 43 so as to improve the subjective evaluation value.

By repeating such processing, the output of the target system 1 is adjusted so as to improve the subjective evaluation value of the person M. Here, the loop between the application of the characteristic parameters to the target system 1, the input of the subjective evaluation values through the third interface 43, and the generation of the characteristic parameters in the optimization unit 3 is performed by the person M via the third interface 43. The process is repeated until the output of the target system 1 is monitored and it is determined that the required performance has been obtained. As is clear from the above description, since the person M is involved in the evaluation of the output of the target system 1 in the interactive optimization apparatus shown in FIG. 6, the output of the target system 1 satisfies the subjective evaluation criteria of the person M. It becomes possible to adjust.

Although the interactive optimization apparatus shown in FIGS. 5 and 6 uses an optimization method, the optimization is performed by trial and error based only on subjective evaluations of humans without using the optimization method. And an interactive optimization apparatus in which a computer supports trial and error during optimization. An example of this type of interactive optimizer is J. Mar
ks, B. Andalman, et al .: Design Galleries: A gener
al approach to setting parameters for computer gra
phics and animation, 24th Int'l Conf. on Computer
Graphics and Interactive Techniques (SIGGRAPH'97),
pp. 389-400 (Aug., 1997)
This document shows an example used for generating graphics. Similarly, A. Koenig, FE Blutner, et.
al .: Anacoustic data base navigator for the intera
ctive analysis of psycho-acoustic sound archive, 5
th Int'l Conf.on Soft Computing and Information / I
ntelligent Systems (IIZUKA '98), Iizuka, Japan: Wo
rld Scientific, Singapore, pp. 60-63 (Oct., 1998)
In this reference, there is shown an example used for designing a violin tone.

FIG. 7 illustrates an interactive optimization apparatus using this technique. Compared with the configuration shown in FIG. 5, the interactive optimization apparatus does not include the evaluation unit 2 and the optimization unit 3 and instead employs a second method for interactively interacting with the person M.
An interface 42 and a mapping unit 5 for visualizing the n-dimensional characteristic parameters by mapping the n-dimensional characteristic parameters into a space of three dimensions or less are added.
The second interface 42 presents the data visualized by mapping the n-dimensional characteristic parameters in the three-dimensional space or less in the mapping unit 5, and outputs the new n-dimensional characteristic parameters specified by the person M to the target system 1. And a function to be provided to the mapping unit 5.

With respect to the interactive optimizer shown in FIG.
The operation will be described assuming that the mapping unit 5 maps (maps) an n-dimensional characteristic parameter into a two-dimensional space. Here, in the mapping unit 5, visualization is possible if n-dimensional characteristic parameters are mapped to a three-dimensional space or less. Here, mapping to a two-dimensional space will be described as an example. In the interactive optimization device shown in FIG. 7, a two-dimensional space corresponding to n-dimensional characteristic parameters is presented on the second interface 42, and a person M first selects a desired point in this two-dimensional space. The second interface 42 inputs the coordinates of the selected point to the mapping unit 5. In the mapping unit 5, the correspondence between the two-dimensional space presented by the second interface 42 and the n-dimensional space is set, and the coordinates of a point in the two-dimensional space selected by the person through the second interface 42 are displayed. When given, it is converted to coordinates in n-dimensional space.
The coordinates in the n-dimensional space obtained by the mapping unit 5 in this manner correspond to n-dimensional characteristic parameters, and the n-dimensional characteristic parameters obtained by the mapping unit 5 are applied to the target system 1. When the n-dimensional characteristic parameters generated in the mapping unit 5 are applied to the target system 1, the characteristics of the target system 1 are determined, and the output of the target system 1 is also determined. The output of the target system 1 is presented to the person M through the first interface 41. When the output corresponding to the characteristic parameter designated by the person M is obtained from the target system 1 through the second interface 42 in this manner, the output is presented by the first interface 41, and thus the target system presented by the first interface 41 It is possible to know the correspondence between the output of No. 1 and the coordinates of the point in the two-dimensional space selected in the second interface 42. As a result, the person M can specify the evaluation of the output of the target system 1 as the coordinates of a point in the two-dimensional space of the second interface 42. Here, the second interface 42 displays all points designated by the person M in the two-dimensional space. Therefore, by repeating the operation of designating a point in the two-dimensional space presented on the second interface 42 and the monitoring of the output of the target system 1 presented on the first interface 41 in response to the operation, the second In the two-dimensional space presented on the interface 42, a distribution of evaluation points is formed. When the first n-dimensional characteristic parameters are given to the target system 1, the person M does not specify a point in the two-dimensional space through the second interface 42, but randomly assigns a plurality of n-dimensional characteristic parameters. After the generation, the person M may start the operation of specifying the characteristic parameters through the second interface 42.

In the interactive optimizing apparatus shown in FIG. 7, the distribution of the evaluation points visualized in the two-dimensional space on the second interface 42 is presented to the person M, so that it is relatively difficult for a computer, but for the person M. Uses the relatively easy global grasping ability, and makes the person M guess and select a position in the two-dimensional space where an optimal solution is likely to exist. Therefore, the optimization is performed by trial and error. Since the state of the trial is visually grasped on the second interface 42 and the result of the trial can be grasped on the first interface 41, the target system 1 is optimized. It can be said that the number of trial and error required for this is relatively small.

[0018]

The above-mentioned configuration examples of the interactive optimization apparatus have the following problems.

That is, since the interactive optimization apparatus shown in FIG. 5 employs an optimization method using an evaluation function or the like in the evaluation unit 2, it is necessary to set evaluation criteria in a computer. In the target system 1 in which the subjectivity of the person M, such as appearance, how to hear, and taste, must be used as an evaluation criterion, there is a problem that it is difficult to use it for optimizing characteristic parameters.

The interactive optimization device shown in FIG.
Since the interactive evolutionary computation method is adopted, as pointed out in the above-mentioned document (research trend of the interactive evolutionary computation method, the journal of the Japanese Society for Artificial Intelligence), the worker M
The problem of fatigue is inevitable. In other words, while the interactive evolutionary computation method is an algorithm that sequentially searches based on a large number of search points, the number of search points that can be evaluated by a person is limited by the storage capacity of the person, and a sufficient number of search points is required. Cannot be handled.

More specifically, when the target system is a hearing aid or a CG, considering the required number of characteristic parameters to be optimized, at least about 100 search points (individuals) are secured per generation and at least I want to search several tens of generations, but the number of search points that can be evaluated by humans is
The number of repetitions is at most about 20 per generation, and the number of repetitions is generally about several generations.
The 20th generation is the limit from the degree of fatigue and practicality. Also,
In the interactive optimization apparatus shown in FIG. 6, the person M only gives an evaluation of the result of the target system 1, and the characteristic parameters given to the target system 1 are automatically generated by the computer. Also, it may take time before the output of the target system 1 is optimized.

Since the interactive optimizing apparatus shown in FIG. 7 does not use the optimizing unit 3, both the evaluation of the output of the target system 1 and the search for the characteristic parameter for optimizing the output of the target system 1 are all performed. Person M will do it. In this configuration, a characteristic parameter that optimizes the output of the target system 1 is determined by the person M globally judging from the distribution of points in the visualized low-dimensional space presented on the second interface 42. Since it is generated, it is possible to make a global judgment that is good for the person M and optimize the characteristic parameters in a relatively short time. However, since the mapping is performed from the high-dimensional space (the space of the characteristic parameter) to the low-dimensional space (the space presented to the second interface 42), the information of the positional relationship in the high-dimensional space is completely converted in the low-dimensional space. It is impossible in principle to keep
The characteristic parameters are searched for by trial and error, relying on the distribution information of points on the distorted space. as a result,
Even the configuration shown in FIG. 7 requires skill in grasping the distribution information of the points presented on the second interface 42, and as a result, the degree of fatigue of the worker M increases, It takes a lot of time to search for a characteristic parameter that optimizes the output.

As described above, in any of the configurations of the interactive optimizing apparatus shown in FIGS. 5 to 7, it takes a lot of time to optimize the output of the target system.
As a result, the degree of fatigue of the worker is increased, and these problems are factors that hinder practical use.

The present invention has been made in view of the above circumstances, and an object of the present invention is to make it possible to optimize a system interactively in a relatively short period of time, and consequently, to optimize the system. An object of the present invention is to provide an interactive optimization device capable of reducing the degree of fatigue.

[0025]

According to the first aspect of the present invention, there is provided a first interface for presenting an output of a target system in which an output characteristic changes according to an n-dimensional characteristic parameter; An evaluation unit that outputs an evaluation value according to the characteristic; an optimization unit that generates a characteristic parameter to be given to the target system; and an n-dimensional characteristic parameter output from the optimization unit, which maps to a low-dimensional space that can be visualized. A mapping unit having a function and a function of inversely mapping an n-dimensional characteristic parameter from a point in the low-dimensional space; a function of presenting a mapping of the characteristic parameter to the low-dimensional space by the mapping unit; A second interface that receives an input of points to the mapping unit and receives the input of points, and provides the optimization unit with the characteristic parameters generated by the mapping unit based on the input from the second interface. An optimization solution adding unit that optimizes the solution based on the characteristic parameter solution candidate generated based on the evaluation value given from the evaluation unit and the characteristic parameter solution candidate given from the optimal solution adding unit. It is characterized in that it is selected and applied as a characteristic parameter to the target system.

According to a second aspect of the present invention, in order to achieve the above object, a function of presenting an output of a target system whose output characteristic changes according to an n-dimensional characteristic parameter and an input of an evaluation value based on human judgment are provided. A third interface to receive;
An optimization unit for generating a characteristic parameter to be given to the target system; a function of mapping the n-dimensional characteristic parameter output from the optimization unit to a low-dimensional space that can be visualized; and an n-dimensional characteristic from the point of the low-dimensional space A mapping unit having a function of inversely mapping parameters, and a function of presenting a mapping of the characteristic parameter to the low-dimensional space by the mapping unit and receiving a point input into the low-dimensional space based on a judgment of a person. A second interface provided to the mapping unit; and an optimum solution adding unit providing the characteristic parameter generated by the mapping unit based on the input from the second interface to the optimizing unit. Characteristic parameters by selecting the optimal solution from the characteristic parameter solution candidates generated based on the evaluated values and the characteristic parameter solution candidates given by the optimal solution adding unit. It has the feature to be applied to the target system by.

By employing the above configuration, the optimizing unit uses the optimizing method without trial and error.
In addition to searching for the optimal solution in the three-dimensional space, visualizing the n-dimensional characteristic parameters given to the target system and showing it to the operator enables searching for a solution candidate based on the global judgment that the person is good at, By adding the solution candidate selected by the person to the solution candidate generated by the optimizing unit by the adding unit, both solution candidates can be integrated. As a result, it is possible to converge to the optimal solution at a relatively early stage, and it is possible to reduce the fatigue of the operator in the interactive optimization and to realize a practical level of optimization.

[0028]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS (First Embodiment) In this embodiment, as shown in FIG. 1, a mapping unit 5, an optimum solution adding unit 6, and a second unit are added to the conventional configuration shown in FIG. Interface 4
2 is added. The second interface 42 is configured to present a space of three dimensions or less as an evaluation space and to specify a characteristic parameter given to the target system 1 in the evaluation space. Further, the mapping unit 5 has a function of performing mutual conversion between the n-dimensional space and the above-described evaluation space, and the n-dimensional characteristic parameter generated in the optimization unit 3 is displayed on the second interface 42 in the evaluation space. And the evaluation of the person M specified in the evaluation space displayed on the second interface 42 is represented by n
Inverse maps to dimensional characteristic parameters. The optimum solution adding unit 6 has a function of giving the n-dimensional characteristic parameter inversely mapped by the mapping unit 5 to the optimizing unit 3 in accordance with the evaluation of the person M input from the second interface unit 42.

The interactive optimizing apparatus according to the present embodiment shown in FIG. 1 first has an optimizing unit 3 similar to the conventional configuration shown in FIG.
Generates an n-dimensional characteristic parameter. When information of the target system 1 is not given to the optimization unit 3 when the optimization unit 3 generates the characteristic parameters for the first time, the characteristic parameters are generated by random numbers for the first time. When the n-dimensional characteristic parameter generated by the optimization unit 3 is applied to the target system 1, the characteristics of the target system 1 are determined, and the output of the target system 1 is also determined. The output of the target system 1 is evaluated by the evaluation unit 2, and the evaluation value from the evaluation unit 2 is fed back to the optimization unit 3. At this time, the output of the target system 1 is
1 to the user. The optimization unit 3 to which the evaluation value from the evaluation unit 2 has been passed generates an n-dimensional characteristic parameter to the target system 1 in a direction to improve the evaluation value, and applies it to the target system 1.

The operation up to this point is the same as that of the conventional configuration shown in FIG. 5, but in this embodiment, the following operation is also performed in addition to the above operation. That is, the n-dimensional characteristic parameter generated by the optimization unit 3 and the evaluation value output from the evaluation unit 2 are passed to the mapping unit 5. Now, assuming that the evaluation space in the second interface 42 is a two-dimensional space, the mapping unit 5 maps the n-dimensional characteristic parameters to the two-dimensional space, and passes the n-dimensional characteristic parameters to the second interface 42 together with the evaluation values. Therefore, since the position of the point corresponding to the characteristic parameter is displayed in the evaluation space which is a two-dimensional space on the second interface 42, the person M observes these points and finds that an optimal solution exists from a global viewpoint. The likely position is specified as a point in the evaluation space. Here, the evaluation value is indicated by shading in the evaluation space.

The coordinates of points input by the person M in the evaluation space through the second interface 42 are inversely mapped in the mapping unit 5 to the n-dimensional space, so that the mapping unit 5 generates n-dimensional characteristic parameters. Then, the characteristic parameters are passed to the optimum solution adding unit 6. In the present embodiment, the optimization unit 3 employs a genetic algorithm to perform an individual-based multipoint search, and generates a plurality of n-dimensional characteristic parameters inside the optimization unit 3. Here, the solution (that is, a point in the evaluation space) selected by the person M as a candidate for the optimum solution using the global judgment ability is highly likely to be effective in searching for a solution. n
This solution inversely mapped to the dimensional characteristic parameter is newly added to a plurality of candidates of the optimization unit 3 or replaced with an n-dimensional characteristic parameter with poor evaluation.

As described above, the optimization unit 3 performs the search for the solution in the n-dimensional characteristic parameter space to be originally searched based on the algorithm, and the solution candidate which the person M visually supposes as the optimum solution is determined. By adding the optimal solution adding unit 6 to the solution candidates of the optimizing unit 3, a solution search combining the advantages of both can be realized. In other words, the characteristics generated by the mapping unit 5 based on the subjectivity of the person M while reducing the burden on the person M by the optimization unit 3 generating a large number of solution candidates based on the optimization method. By adding the parameter as a solution candidate, it becomes possible to provide the target system 1 with a characteristic parameter reflecting the subjectivity of the person M.

The mapping in the mapping unit 5 includes a linear mapping such as principal component analysis, and a non-linear mapping (NLM) (J.
W. Sammon: A Nonlinear Mapping for Data Structure
Analysis, IEEE Trans. On Computers C-18, No.5, pp.
401-409 (1969)), Visor (A. Koenig, O. Bulmahn,
M. Glesner: Systematic Methods for Multivariate Da
ta Visualization and Numerical Assessment of Class
Separability and Overlap in Automated Visual Indu
strial Quality Control, 5th British MachineVision
Conference, pp.195-204 (1994)), TOPAS (A. Koenig:
A Survey of Multivariate Data Projection, Visuali
zation and Interactive Analysis, 5thInt'l Conf.on
Soft Computing and Information / Intelligent System
s, pp.55-59 (1998)), a method using Genetic Programming (G. Venturini, M. Slimane, F. Morin, JP As
selin de Beauville: On Using Interactive Genetic A
lgorithms for Knowledge Discovery in Databases, 7t
h Int'l Conf.on GeneticAlgorithms, pp.696-703 (19
97)) and self-organizing maps (T. Kohonen
Satoru Kishida / Kikuro Fujimura: Self-Organizing Maps, Springer-Veerark Tokyo, 1996) and other mapping technologies can be used.

When the self-organizing map is used, the correspondence between the thousands and tens of thousands of n-dimensional coordinates and the two-dimensional coordinates is stored in the inverse mapping by the mapping unit 5 when the n-dimensional mapping is performed in advance. Therefore, reverse mapping can be easily realized. With other mapping methods, inverse mapping is possible by combining it with search for inverse mapping using a genetic algorithm, and this technique is described in H. Takagi, T. Noda, and S.-B. Cho, Psycholog
ical Space to Hold Impression Among Media in Commo
n for Media Database Retrieval System, IEEE Int'l
Conf.on System, Man, and Cybernetics (SMC'99), To
kyo, Japan, Vol.VI, pp. 263-268 (Oct. 12--15, 199
It is disclosed in 9).

An experiment comparing the present embodiment with the conventional configuration shown in FIG. 5 yielded the following results. Number 1
Is a modified version of Schaffer's second function, which is widely used in the performance evaluation of genetic algorithms. Here, m is a numerical value for changing the complexity of the function, and m = 3 and m = 5 were adopted in the evaluation experiment. In the evaluation experiment, the value of Equation 1 is a characteristic of the target system 1, and the optimizing unit 3 sets the optimal value, that is, xi such that the value of Equation 1 is maximum.
Finding (i = 1, 2,..., M) is the optimization of the target system 1.

[0036]

(Equation 1)

In this embodiment, six subjects searched for the optimal solution. The optimizing unit 3 uses a genetic algorithm (hereinafter, referred to as “GA”) as an optimizing method. Each parameter of the genetic algorithm reduces the number of individuals to 2
The experiment was performed up to the tenth generation with 0, the crossover rate being 0.9, the mutation rate being 1/80, and the number of bits being 16.

FIG. 2A shows m = 3, and FIG. 2B shows m = 5.
In this case, the convergence result is compared with the conventional configuration shown in FIG. The solid line A in FIG. 2 is the average value of the six persons in the present embodiment, and the other lines are the average values of the six persons in the conventional example shown in FIG. In the interactive optimizer, G
Considering the problem of user fatigue, the number of individuals of A must be at most about 20. Searching for the complex function shown in Equation 1 with such a small number of individuals is impossible with a conventional interactive optimizer, and does not converge at all for 40 generations. When the number of individuals is increased to 100 or 1000, convergence starts, but with such a large number of individuals, it is not practical as an interactive optimization device from the viewpoint of human fatigue. On the other hand, in the interactive optimizing apparatus according to the present embodiment, even if the number of individuals is only 20, the number of individuals is 100 or 10 in the conventional interactive optimizing apparatus.
It was found that the same convergence performance as in the case of 00 was obtained. Therefore, it can be said that the problem of the interactive optimizing apparatus, in which the fatigue problem tends to be a hindrance factor for practical use, is greatly improved.

In the evaluation experiment of this embodiment, an example in which the optimization unit 3 uses a genetic algorithm has been described. However, the present invention is not limited to the optimization based on the genetic algorithm, and enables an interactive design. Any optimization technique can be used.

(Second Embodiment) This embodiment is similar to FIG.
As shown in FIG. 6, a second interface 42, a mapping unit 5, and an optimum solution adding unit 6 are added to the conventional configuration shown in FIG. These configurations have the same functions as the configurations of the same reference numerals described in the first embodiment.

In the interactive optimizing device shown in FIG. 3, the optimizing unit 3 first generates n-dimensional characteristic parameters, as in the conventional configuration shown in FIG. When the optimization unit 3 generates the characteristic parameters for the first time, the characteristic parameters are generated by random numbers. When the n-dimensional characteristic parameter generated by the optimization unit 3 is applied to the target system 1, the characteristics of the target system 1 are determined, and the output of the target system 1 is also determined. The output of the target system 1 is the third interface 4
3 to the person M. The person M evaluates the performance of the target system 1 on the basis of the presented output of the target system 1 based on sensitivity and knowledge, and returns an evaluation value to the optimization unit 3 through the third interface 43. That is, the person M subjectively evaluates the output of the target system 1 and inputs the subjective evaluation value to the third interface 43, and the third interface 43 transfers the subjective evaluation value to the optimizing unit 3. In the optimization unit 43, based on the subjective evaluation value from the third interface 43,
An n-dimensional characteristic parameter to the target system 1 is generated so as to improve the subjective evaluation value.

On the other hand, the optimizing unit 3 uses the n-dimensional characteristic parameters inversely mapped in the n-dimensional space by the mapping unit 5 from points in the evaluation space inputted by the person M through the second interface 42 as solution candidates, and In step 3, the solution candidate generated by the mapping unit 5 is added to the solution candidate generated by the optimization method. As described above, the optimization unit 3 performs the solution search in the n-dimensional characteristic parameter space to be originally searched based on the algorithm for realizing the optimization method, and the solution candidate which the person M visually infers as the optimum solution. Is added by the optimal solution adding unit 6, a solution search combining the advantages of both is made possible as in the first embodiment. Other configurations and operations are the same as those of the first embodiment.

FIG. 4 shows an example in which the present embodiment is applied to the target system 1 as a synthesizer. The person M evaluates the sound transmitted from the synthesizer and returns an evaluation value to the optimization unit 3 through the third interface 43. Further, a point X at which an optimal solution is considered to exist in the evaluation space is input through the second interface. In the second interface 43, not only points corresponding to the characteristic parameters given to the target system 1 are displayed, but also the evaluation values input from the third interface 43 according to the density of each point. In the illustrated example, an example is shown in which the evaluation value decreases in the order of black painting, oblique lines, dots, and white outlines. On the screen of the second interface 43, the evaluation value is actually represented by lightness or the evaluation value is represented by color. If the person M newly inputs a point X at a position where it is presumed that an optimal solution exists by visually observing the distribution of such evaluation values, the characteristic parameter corresponding to this point X becomes the optimal solution candidate. This is input to the conversion unit 3. In the illustrated example, the sound of the synthesizer is optimized as the target system 1, but the configuration of the present embodiment is not limited to the case where the synthesizer is used as the target system 1.

[0044]

According to the present invention, as described above, the optimization unit uses the optimization method instead of trial and error to search for the optimal solution in the n-dimensional space and to provide the n to the target system.
By visualizing the dimensional characteristic parameters and showing them to the operator, it is also possible to search for solution candidates based on the global judgment that the person is good at, and the optimal solution adding unit selects the solution candidate generated by the optimization unit By adding the calculated solution candidates, both solution candidates can be integrated. As a result, it is possible to converge to an optimal solution in a relatively early stage (that is, in a relatively short time), and to reduce a worker's fatigue in interactive optimization, thereby achieving a practical level of optimization. There is an advantage that can be.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a first embodiment of the present invention.

FIG. 2 is a graph showing experimental results of performance evaluation of the above.

FIG. 3 is a block diagram showing a second embodiment of the present invention.

FIG. 4 is a block diagram showing an application example of the above.

FIG. 5 is a block diagram showing a conventional configuration.

FIG. 6 is a block diagram showing another conventional configuration.

FIG. 7 is a block diagram showing another conventional configuration.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Target system 2 Evaluation part 3 Optimization part 5 Mapping part 6 Optimal solution addition part 41 1st interface 42 2nd interface 43 3rd interface

Claims (2)

[Claims] 1. A first interface for presenting an output of a target system whose output characteristic changes according to an n-dimensional characteristic parameter, an evaluation unit for outputting an evaluation value corresponding to the characteristic of the target system, and a characteristic parameter given to the target system , A function of mapping the n-dimensional characteristic parameters output from the optimizer into a low-dimensional space that can be visualized, and a function of inversely mapping the n-dimensional characteristic parameters from the point of the low-dimensional space. A mapping unit having a function of presenting a mapping of characteristic parameters to the low-dimensional space by the mapping unit, a second interface for receiving an input of a point in the low-dimensional space, and providing the input to the mapping unit; An optimal solution adding unit that gives the characteristic parameters generated by the mapping unit based on the input from the interface to the optimizing unit, and the optimizing unit evaluates the characteristic given by the evaluating unit. Interactive optimization characterized by selecting the optimal solution from the solution candidate of the characteristic parameter generated based on the value and the solution candidate of the characteristic parameter given from the optimal solution adding unit, and applying it to the target system as the characteristic parameter. apparatus. 2. A function for presenting an output of a target system whose output characteristic changes according to an n-dimensional characteristic parameter, a third interface receiving an evaluation value based on human judgment, and a characteristic parameter to be given to the target system. An optimizing unit; a function of mapping an n-dimensional characteristic parameter output from the optimizing unit to a low-dimensional space that can be visualized; and a function of inversely mapping the n-dimensional characteristic parameter from a point in the low-dimensional space. A mapping unit, a function of presenting the mapping of the characteristic parameters to the low-dimensional space by the mapping unit, and a second interface to the mapping unit upon receiving a point input into the low-dimensional space based on the judgment of a person, An optimum solution adding unit for giving the characteristic parameter generated by the mapping unit based on an input from the second interface to the optimizing unit, wherein the optimizing unit includes a third interface The characteristic feature is that the optimal solution is selected from the solution candidate of the characteristic parameter generated based on the evaluation value given through the process and the solution candidate of the characteristic parameter given from the optimal solution adding unit, and is applied to the target system as the characteristic parameter. Interactive optimizer.