JPS58192161A - Constitution system for knowledge base - Google Patents

Abstract

PURPOSE:To provide capability of improvement and of gradual acquiring of knowledge for a knowledge base, by giving an evaluation representing the propriety of chracteristics and an age started from (0) at its formation and then increasing for a code, and eliminating codes with deteriorated characteristics and aged codes from the knowledge base. CONSTITUTION:The knowledge base KB200 is accumulated codes, and a knowledge base handler KBM reads a code of an address of a base KB represented with the content of an area 220 in a knowledge base buffer 255 with a read/ write function 215, and a code on a base buffer 260 is stored in the address of the base KB indicated with the content of an area 225. Further, a knowledge base manager KBM calls a code in the base KB, executes processing and stores the code to the base KB. Thus, the code is provided with an evaluation representing the propriety of characteristics and an age started from (0) at its production and then increasing, and codes having deteriorated characteristics and aged codes are eliminated from the base KB.

Description

DETAILED DESCRIPTION OF THE INVENTION (1) Technical Field of the Invention The present invention relates to a method of structuring a knowledge pace, and in particular to a nine-dimensional structure that gives the knowledge pace the ability to grow, that is, the ability to acquire new knowledge or improve acquired knowledge. Regarding the method of

Recently, a proposal for knowledge engineering was made by American researchers involved in the development of artificial intelligence research, especially researchers from Stanford, Carnegie, and MTT.

In knowledge engineering, experts' knowledge and know-how are combined with conditions and
Under these circumstances, the procedures to be carried out are revealed in bare form, and this is accumulated to form a knowledge base. As a specific example, the software MMCIN, which helps in diagnosing and treating diseases, is known. Once the knowledge pace is established, 84
It is possible to demonstrate the ability to surpass a doctor in quantity like YCIH, but at present the knowledge pace needs to be created by programming the bearer death of (condition 9 procedure) in the same way as conventional program creation work. It's very inconvenient.

■] Purpose of the Invention The purpose of the present invention is to eliminate all the inconveniences associated with the conventional knowledge base structure of pre-programming the condition 1 procedural bare bones, and to improve the ability to acquire knowledge gradually and to The idea is to provide a method for keeping the pace of knowledge at a pace that allows the ability to grow or adapt, to improve the knowledge acquired.

(3) A pair (hereinafter referred to as a code) of the outline of the invention (condition 9 procedure) has a condition part that matches any condition, and if it is called under a certain condition, it will not receive external input. A function that generates procedures to be performed under those conditions, merges all the conditions and procedures, and stores them in the knowledge pace? Add the mother code that you have.

By doing this, when a new condition occurs that does not match the condition part of the code in the knowledge base, it becomes possible to generate corresponding code on the spot. Next, the code should be able to generate a mutated copy of itself (with a slightly different conditional or procedure part) each time it is used.

Furthermore, when generating or using each code, it is possible to give its characteristic evaluation value and age, so that there is a gap between the codes in the knowledge pace regarding the characteristic evaluation value and age given above. I'll keep it. That is, codes with bad evaluation values are forced out of the knowledge pace by codes with good evaluation values, and older codes are also gradually pushed out.

With the above functions, the code generated by the mother code generates all mutated copies of itself every time it is used, and if the mutated copies have good characteristics, they will remain in the knowledge pace, and if the characteristics are bad, they will remain in the knowledge pace. , banned from the knowledge base. In other words, when viewed as a whole, the knowledge pace has acquired the ability to adapt.

(4) Embodiment Although the present invention can be applied to various objects, an example of application to a solution search system will be described for the sake of clarity.

As is well known, solution search techniques are one of the core systems techniques. Recently, researchers aiming to create automatic simulations by connecting them with simulators have become popular.

Many solutions have been proposed in the past depending on the nature of the problem to be optimized, but from the user's perspective, a considerable amount of usage knowledge is required in selecting a solution method and setting parameters according to the nature of the problem. .

This difficulty could be overcome if there was a universal solution, but it is difficult. For this reason, we have prepared a variety of methods and solutions, and are looking for compatibility between the two. An approach has been proposed to gain knowledge through experience. Inspired by this proposal, a solution search system has been proposed in which a collection of solution methods for each building is collected and an appropriate solution procedure is selected depending on the diagnostic results of the nature of the problem and the state during the solution search. However, the solution procedure and Part 1. ? It was necessary to define in advance how to respond to the scene and situation in which it would be called.

However, a system that aims to make full use of accumulated experience must be gradually perfected through a process of trial and error. In order to make this possible, the entire system must be configured as a thread that can constantly change its structure and form due to the inflow of information from the external environment, that is, as an open system. A solution search system based on this viewpoint will be described below as an embodiment of the invention.

An explanation will be given keeping in mind the conversational non-i planning system (FIG. 1) as a representative example of a solution search system.

When used, first the unknown variable/ is converted to the target/constraint value f(1)
Define a procedure for calculating 10) and program it (20). To search for a solution, analyze the nature of the problem, the progress of the solution search, the status of the solution, etc. (50), select an appropriate solution from the solution method (30), and instruct the supervisor (40) (60).

Scenes and situations that should be noted in the solution search process, procedures and parameter values to be used in those cases are based on the user's experiential knowledge (7
Selected based on 01r. In order to improve the solution search ability, it is necessary to enrich the procedure set (30) by developing a solution method (80), and many proposals have been made in the past. However, it is also important to improve the functions (50) to (70) that have been carried out by the user. This part will be hereinafter tentatively referred to as the knowledge tube ring (KM) (90). The functions of KM90 are
It calls an appropriate procedure based on information from the monitor function, and mathematically it is nothing but the following mapping ψ.

ψ: K(S) → A(P) Here, K(8) is a set of observation information that depends on the scene or situation S, A(P) is a set of the arrangement of elements of (30),
P indicates the dependence of the procedure on variations. Below, we will consider the automatic formation of ψ and its use in order to improve the functionality of the solution search system. Let's consider the issues at this time.

(1) Other than the knowledge that can be registered such as (1) ψ's exchange formation secondary force), we have to add it one by one every time we notice it.

(2) Adaptation to environmental changes, etc. 2 problems (10), monitor function (50), solution procedure set (30), knowledge pace (70)
), but these are gradually changing. ψ needs to be gradually perfected in such a dynamic environment.

At the same time as the ability to specialize, which improves the ability to respond to similar situations.
It is also necessary to have the flexibility to deal with unfamiliar situations. In general, information about environmental forces is often subject to disturbances, and adaptive ability in a statistical sense is required.

(3) Applied behavior: In order to control the pace of knowledge and save time for knowledge acquisition, it is necessary to be able to apply something more similar to the current scene or situation from the accumulated experience.

Next, we will discuss how to solve the problem.

(1) Knowledge representation by code: It is possible to use a density table to represent ψ, but there are problems with the above-mentioned points. Another method that can be considered is the family of learning machines represented by the persebtron. C fan is the element A(P) of A(P)
For, the discriminant function (f(
A(P), K(8), W)IA(P)EA(P), to(
S)EK(S),WER,',m=dim(ni(S))
+1) and select it as the A(P)'ik action corresponding to the discriminant function that takes the maximum value based on the input (S). If A (P) is a small finite set, then the variable W
(2) above can be achieved to some extent by adjusting A.
A problem arises because (P) generally becomes an infinite set due to perturbation of the parameter P.

Therefore, although it is primitive, here we will introduce the parameter vector K(8) that characterizes the scene S and the corresponding solution procedure sequence A.
Use the following code consisting of (P).

(K(S), A(P)) = (Kl (S), Kt(S),...K, (S
), At (Ptl.

At(Pt), ......, ^, (P,)) If you use this code as necessary, the above (1) will be realized. By parameterizing the situation of the scene and creating a key, the set of codes described above can be searched using this key, and the appropriate procedure can be called.

(2) Modularization of solution procedures, etc. and Boolean: the above (
The procedure for monitoring the scene situation K(S) mentioned in 2) and the solution procedure A(P) are collectively pooled.

This will be newly expressed as A(P) below. The elements of A(P) should be modularized as much as possible. In order to avoid complicating information communication, communication between procedures is performed in a manner in which each procedure refers to information entered by each procedure in a common storage area. Changes or additions to procedures are implemented by changing elements of pool A(P).

(3) Code proliferation and creation: A mother code program that can be called with any key and, when called, generates a code with an arbitrary sequence of procedures based on user input. Additionally, the code has the ability to create a copy of itself each time it is used and to change parameters on the copy by user input. Next, we will make it possible for users to give characteristic ratings to the codes they use. moreover,
Each code is given an age that is 0 years old at the time of generation and is increased by one year each time the code group is accessed thereafter.

A limit is set on the number of codes that can be stored, and after the total number of codes reaches the limit, competitive selection is performed between the codes based on the above-mentioned score and age.

The mother code enables the registration of actions for unexperienced scenes,
Self-propagation and selection based on ratings enable specialization, that is, the improvement of ψ's characteristics for similar situations. Next, age-based selection eliminates codes that are used less frequently and therefore have fewer opportunities to proliferate, thereby ensuring the plasticity of ψ, which increases the number of codes that are effective in the current environment. When calling a code, a method is used in which a number of similar codes with good characteristic scores are retained and one is randomly selected from among them. Additionally, even if there are disturbances in the code characteristics and their ratings, codes with a high probability of exhibiting good characteristics can gradually occupy a dominant position. As described above, the above-mentioned problem (2) can be addressed by proliferation, mutation formation, and pottery.

(4) Division of the set K(S): In order to deal with the problem in (3) above, the domain K(8)'e of the scene situation vector is divided into several categories, and those in the same category are regarded as similar codes. .

However, the division of K(S) itself should be formed through experience, and it is problematic to define the division in advance. Therefore, in the following, a method is used in which provisional divisions are given in advance and each category of the provisional division is gradually filled with codes with high adaptability using the above-mentioned adaptive function. Other methods will be described later.

show. In the same figure, KB is knowledge pace, KBM is knowledge pace manager, KBH is knowledge pace manager, KPM
indicates knowledge pace supervisor and PSF indicates problem solving function. Each part is explained below.

(1) KB (FIG. 2(A) 200): A collection of codes. Figure 2(B) shows forgery of the code.

Here, the code's characteristic rating and age were also added as some factors. The part consisting of the code number and the status parameter will be expressed as (S) below.

(4KBM (205 in FIG. 2(A)): Calls and executes the code in the KB, and stores the code in the KB.

(3) KBH (210): Read and write function (R&W) (2
15), the K indicated by the contents of the BAD area (220)
All codes for address B K B B U F 1
'1 (255) is read and the SAD area (225
) to the KB address indicated by the contents of KBBUF2 (260)
Store the above code. WKEY (230) is Wari,
5KEY (235) is a key storage area for searching the storage location of the code in KB, and RKEY (240) is a key storage area for searching the address in KB of the code called to KBB[JFl. The format is the same as the part A in FIG. 2(B). 8
1 (AD or S) using the keys on KF, YfRKEY.
The task of finding the address in KB to write in AD is KB B
tJ F l Fill in the code above ni 2nd (A)
Pronation is carried out by the 5OLV (275) element.

(4) KSM (245): KBBIJFl is the storage area for the called code, and KBHIJF2 is the storage area for the code after execution. The format of each is the same as the code. KBCT
TR, (265) is an execution procedure storage area whose format is the same as that of the solution procedure section in FIG. 2(B). EXEC(25
0) interprets the code on KBBUFI and executes a predetermined procedure in 5OLV.

Details will be described later.

(5) PSF (270): 80LV (details will be described later), which is a set of programs for individual elements of set A (P), and SO
The elements of L2 consist of array 8A (280) used for solution search and storage area C (2851 units) of constant C.

(6) SA, C: Format Kintei 2 (C) shown in Figure.

If it is a procedure called by KSM (245) or a solution search method, Nana et al. generate a vector term in their own way, drive the simulator, calculate f(s, C1, We will continue to improve.cjL and others will store the memory in the me and f parts of the diagram.In addition to this, there are also procedures in 80LV that change the contents of individual elements in the diagram, such as the solution search definition line. C is a storage area for constants given to the simulator. KBM operation Figure 3 shows K B M (7) Operation sound BO
The cycle of calling code from , executing it, and storing the code in KB after execution is repeated.

(301 in the figure) Preprocessing: Import the contents of the KB saved on the external storage device, and set default values for variable names, function names, etc. on the SA.

(302) Initial setting of SAD, f'tAD 2) Address in KB of the code to be executed first 1) In LAD, add the garbage collect code on KB (the last code in KB prepared for the purpose of removing unnecessary chord tones). code) address to SAD.

(303) Calling and storing the code: Call the code group KBBUF1 at the address indicated by R and AD, and store the contents of KBBUF2 at the address indicated by SAD.

(304) Judgment: If execution of the code on KBBUFl is completed, go to (308); if not completed, go to (305).

(305) AI (Pt) Copy: Copy procedure AI (PI) on KBBUFl by EXEC to K RCU
Copy to R.

(306) Execution: F', Eli to XEC, K B CI
The information on TR is interpreted and the corresponding procedure in SQL~' is called and executed. The called procedure is KBCUR
The above parameter P1 is displayed on the terminal, changed to P+tPt' on KBCUR by user input, execution is started using P1', and the execution progress and results are displayed on the terminal. On the code, create the continuation code search key tRKEY,
Use this to search KBt and enter the result in 1RAD, evaluate the characteristics of the code after execution, 8KEY
It also includes searching KB' using the contents of , and entering the result in t=sAD.

(307) Copying KBCUR to KBBUF2: EX
Implemented by EC.

(308) Key merge: 5KEY contents to KBBUF
Merge into part of the code on Z.

As will be explained later, the contents of RKEY used to call the code to KBBtJFl before image processing are transferred to WKEY, and after entering the rating etc., it is transferred to 5KEY. Therefore, (306) to (308), KBBUF
A code is generated on Z that has its own calling condition parameters and score in its key part and its own mutated copy A(P') in its procedure division.

(309) End operation When the system ends, the contents of the KB are saved in the external storage device.

Set A(P) or 5OL in the system of solution procedure set embodiment
The essential SV of v (Figure 2 (A) 275) is shown in Figure 4. In the figure, [410] indicates the storage of 5AJPC in an external storage device, etc. 2
is to change the contents of SA and C, 3 is to calculate ffΔf, 4 is to correct the limit and f tolerance, etc., 5 is to register or delete / to SA,
6 and 7 are outputting solutions and analyzing their properties, diagnosing the properties of BFi problems, searching for solutions in 11 to 13, 21 is an editor for code in KB, 22 is calling code from KB, storage control, 23 Vi, code 90 performs a termination process. Some of them will be explained in detail below.

(5, 3): Delete the vector / in SA.

(7, 3): Parameters Pl, P! , Ps t- has.

The degree of increase in the extended objective function value on the P2 division point from the P1-th point to the other points in SA is checked, and cluster numbers P,'t- are given to points belonging to the same cluster as the former. If there is a point in another cluster, its number is displayed, and if there is no point, 0 is output.

(7,4): When P, 〆0, from each cluster P, set (
X, f) k Output together with the cluster number. No output when P1=00.

(8, 1): Create a continuation code search key using the continuation code number and the nature diagnosis result of the problem and enter it in RKEY. As the continuation code number, the default one for the three parameters PlsP, , P is used. It can also be changed by the user.

When P,=00, user input is used. Nature of the problem 1d/
The number of dimensions of the extended objective function defined by the following formula M,
It is expressed as the time required to weigh f. These parameters are obtained by driving the simulator T-drive.

-(L(μ,f,/i)+L(μ,f,/j))/
2')) Here, E is the expected value, U(t)=1 (t>
0), =0(1≦03.L(μ, f, /) is the extended objective function value of f at point /.

(12, 3): Neelder-Mead sympresis method (14). Pl is initial simplex size determination, Pl is convergence judgment, Pl to P are Kasuri image, degeneracy.

The expansion projection coefficient, PI, is an output control parameter.

Minimize the extended objective function.

(12+5): W, L, Pr1ce projection method. We create p, Minimize the extended objective function to 1. The purpose is to search for an optimal solution for a multimodal function, but the required dt
There are many calculations. It has parameters from Pl to pH, but their explanation will be omitted.

(21, 3): Incorporate the solution procedure and its parameters desired by the user. Since the contents of KHCUR are transferred to K l'3 B U F 2, KBB
The desired code is generated on UF2.

It has no parameters.

(22,1): Using the contents of WKEY, B B
Find the storage location of the UF 21 code in the KB and enter it in the SAD. At P, 20, the unused first address in KB,
If there is no unused portion, write the address of the worst code. The worst code is determined as follows. First, the number N' of codes in the KB to be matched (note later) with the contents of WKEY is determined.

If N<P, use the address of the code with the second score>P4 among the elderly codes. If there is no such code, use the address of the Gabetsuji Ken collect code. When N≧pH, the address of the N code with the highest internal temperature 2 score is used. When P,=1, the direct address of P is used, and when P,=2, the address of the gap correct code i, when p,=3, the user input address is used.

When P, = 1, clear KF (CUR, but do not clear when PR〆l. After the above processing, the age entry field on WKEY is set to O, and the storage code is set to 0 years old.Finally. ,W
Move the contents of KEY to 'ksKEY and age all codes in KB by 1 year.

(Note) The two keys are determined to match only when the code number and status parameter part (8) each element matches. Each element matches only when at least one of the elements is a "side dish" and when it is included in a predetermined category. As an example of a category, for example, if the dimension of 〆 is 1 to 5, the category name is 1.6 to 10 is 6.11 to 20 is 11.21 or more', and the above is 21. stipulate. The multimodal degree is set to 0 if M≦0 and 1 if M>0.

(22, 2): Obtain the KB address of the code to be called in KBBUFl and write it in R and AD. Parameter P, ~P4
have. The roles of P and -P are similar to (22,1). First, the contents of RKEY' (after moving to WKEY,
If P, = O, among the codes in the KB that match WKEY, leave at most P4 from the small power of the first score, and select one code at random from among them, 'k R,
Fill in A.D. For PI) VO, (22,1)
See section.

(22, 3) Nio (1) Create a continuation code search key in RKEY based on user input, etc. Parameters P1 to P,
have money After filling in RKEYt-', (811
) Follow the procedure explained in the first part.

(23, 1): Does not have any parameters. After the code is executed, the user is informed of the number of times the simulator has been driven and the amount of time the computer has been used, and the user is asked to input two scores with a domain of 0 to 100, and the WKEYGIF is imported.

The first +FF score is a criterion for calling a code, and the second score is a criterion for discarding a code. The lower the score, the easier it is to be called out and the less likely it is to be expelled.

Here is a code example using the above procedure with various characteristics? Shows that you can define waiting code. For convenience, the codes (560) to (590) related to the later 69 light will also be explained.

(510) Mother code: Fig. 5 (510) C', - C matches all keys, but the most
(1st rating 100), least likely to be expelled (2nd rating 0)
). In case it is called with a certain key K (S), the user's desired code 1K is returned by the procedure (21, 3) in the code.
Create in BBUF2.

After that, RKE requires the user to enter the continuation code number.
Dig into Y (22,3), K on WKEY (S)
Give the score of the generated code in the score column of (23, 1),
After merging this with the code on KBBIJF2, KB
Calculate the destination to store the code (22.1) and the destination to call the continuation code (22.2). As you can see from the above, this code has the called condition as a key part, and generates code that has the entire action part of the solution procedure under that condition.

(520) Non-proliferative code: Since the code at (520) in the figure is (22, 1) P = 2, it discards its own copy on KBBUFZ to the gearbage self-collect code in KB. In other words, it does not have the ability to proliferate. Use when you want to catalog and save all specific codes.

(530) Proliferative coat (5-30) Vi, ('
22, "1) to) P, = 0 'T:To(
7) Te, the copy of itself on KBBUFZ weeds out the recessive code in the KB and enters the KB. Therefore, it is used when you want to improve the characteristics of the code by gradually improving the parameters on the code.

(540) Processing start code: (5
When the 4H code is called, the procedure list (Fig. 4) is output (1, 3), and then the user is prompted to enter the continuation code number.

(550) Termination code: The code (550) in the figure executes the system termination operation (90, 1)' (f-).

(560) Problem diagnosis code: For the code (560) in the figure, after diagnosing the problem, a key consisting of the code number (1000゜2, 饅) and the parameters of the problem diagnosis result is created and displayed to the user. After making the necessary corrections, find the KB address of the continuation code. It is non-proliferative.

(570) Simbrex solution search: The code (570) in the figure performs a solution search using the simplex method (12, 3). It is a proliferative code.

(580) Solution search using projection method #: The code (580) in the figure performs a solution search using the projection method by W, L, and Pr1ce. It is a proliferative code.

(590) Huuristic global search: (590) H18A (7) in the figure, 1. fmtri IJ7 (5,
3) After L, perform a complete tentative search using the projection method, switch to the sympresis method midway through to find local optimal solutions, and perform cluster analysis centering on the solution points. After repeating the latter two twice, output the results. It is proliferative.

Experimental Results We will conduct an experiment to determine whether it is possible to form an incremental version of the above-mentioned mapping ψ. As shown in FIG. 6, the system is presented with different problems one after another to search for solutions, and the operation of the system is observed.

(610) Initial code setting: Fig. 5 (510) (54
0) (550) (i'): 7-do*KBvc was registered and (540) t- was called at the same time as startup.

(620) Registration of problem diagnosis code: Code (540) is executed at the same time as startup, and after displaying the list in Figure 4,
I asked for the continuation code number. So I entered (1000, 1). Since the mother code has been called, Figure 5 (56
0) code has been registered. I was asked for a continuation code, so I entered (1000°2.1).

(630) Registration of simplex solution search code: The mother code is called and compared. The code shown in Figure 5 (570) was registered.

(4) Problem setting: In the simulator shown in Figure 2 (A),
The following objective function was set. The problem is to minimize this.

fo (, voice≦□nIA+ 11%-7ku(11
1Bs +C+ lHere, (A, Bt, Ct
l jEI, II 1=2) is a constant. fo
(8) is a synopsis l) Set broken point %t (iε■)
It is a function of the distance between Jl and the end, and has a local minimum value Cat at the end. As the first problem, we solve the unimodal
lII=-+) in four dimensions, me = (0, 5, 0, 5, 0
,5,0,:a) The solution was set.

(650) Calling and executing the problem diagnosis code: Enter the code number (1000, 1) and call the code in FIG. 5 (560). As a result, a single peak was detected and the following key was generated. (1000゜2, fight-, 4,0,
0,316E-2,11n.

(660) Continuation code call and execution: -v in the above key
Of Figure 5 (510) and (570), (51
0) was called, so be sure to register the same code as explained in (630) above.

(670) Repeated use of registered code: Figure 5 (56
0) All done again friend. As a result, FIG. 5 (570) was called. The above steps (650) and (66o) were repeated while changing the parameters of the sympresis method. FIG. 7 shows the screen display during the second repetition. A summary of the progress is shown in Figure 8.

During this period, the empty area of KB disappeared. From this figure, it can be seen that the number of function calculations required to obtain a solution gradually improves. This is due to the recessive code porcelain.

(680) Problem change = 8-dimensional unimodal, /, = (0,5
, 0, 5, 0, 5, 0, 5, 0, 5, 0, 5, 0, 5° 0.5). As a result of running the problem diagnosis code, the mother code was called because it was an unexperienced event. Therefore, I registered Figure 5 (570) and Mr. Sue's code. Thereafter, the same steps as above (670) were repeated. The solution search efficiency gradually improved.

(681) Change the problem: 〆, = (0,5,0,5゜0.5,0.5) and 7. = (−0,5,−0,5
, −0,5°−0,5) and a four-dimensional bimodal function that takes the value 1.0 was set. As a result of problem diagnosis, non-convexity was detected and the following key was generated. (1000,2, tatami, 4,0.3
°0.273E-2,-). Since the mother code was called δ, the code shown in FIG. 5 (580) was registered. As a result of running the problem diagnosis code, the result in Figure 5 (5) that matches this key is
10) (570) of (570) (580) was called. Since this code is for a unimodal function, I gave it a bad score. As a result of repeating the process of giving a good score when (580) was called, (580) was reliably called.

(682) Change of question: 8 similar to (680') above
A friend who sets up dimensional problems. As a result of problem diagnosis, the mother code was called. Therefore, I registered the code shown in Figure 5 (590). The characteristics of the code improved through repeated use.

(683) Presentation of similar problem = 66-dimensional unimodal problem was set. As a result of problem diagnosis, the code obtained in the above (680) was called. As a result of running the code without changing the parameters, a good solution was obtained in one run.

(684) Presenting a similar problem: A three-dimensional bimodal problem was set. The code obtained in step (681) above was called and two solutions were found without changing the n1 parameter.

(685) Presentation of similar problems = All 10-dimensional two-modal problems were set. The code obtained in (682) above is called, and two solutions are found in one execution without modifying the parameters.
The total number of calculations was 1431. For comparison, the fifth
As a result of solving the problem using the code shown in Figure (580), approximately 6,000 attempts were required to obtain two solutions. As a result of repeatedly executing Fig. 5 (570) by roughly changing the initial point, 2
It took about 4,000 calculations to obtain the solution.

(690) Termination operation: The experiment was terminated by calling the code in FIG. 5 (550).

As a result of the above, the above-mentioned problems (1) (2) (3) t-
Furthermore, the process of 681 means that KB, which had only experienced a situation where a code number is present in the code call parameter, is now in a new situation where it can also observe the characteristic parameters of the problem. 17 shows that they have adapted to this situation. In other words, it shows that it has adapted to changes in the set of monitor functions K(S).

Although one embodiment of the present invention has been described below, the following modifications may also be considered.

(1) Empirical partitioning method of set K(8): K(
From the user's point of view, it is convenient to divide σ without dividing S) and to divide σ through experience. As one method, it is conceivable to introduce the territory of each code into K(8).

As a concrete proposal, a vector consisting of a part of the code number and the situation of each code (see Figure 2 (B)),
K, abbreviated as (S)), call score JI, expulsion score J1, age t1 parameter '1 + 'l,
Each code is accompanied by the following function part.

(Ps (V, ni-(8), Js m rt.

Pz(V, ni-(S)-Jt,'-m' 1yeK
, (8)).

Here, P is the only maximum lJ\value J in y=1 (S)
1, P, becomes y=, the only maximum f-direction Jt in (S)
k is a potential function on (S), '1
* 't is a parameter that defines the gap and range of p, , P (see Figure 9). When calling a code with the key y, select randomly from among the codes with the small value p, (y, Nimong (8), J,, rs), and when selecting the worst code to be expelled from the KB, pt(y ,Ni,(S),J
s − ", ) value is larger than the second score of the code having the key y"'. Under this scheme, for two codes with -s (81, K-t (8)), the 13-branch set on (8) is
(ylpt(y, ni-1(8), Js*rt)
=P+ (L K-1(S), J,', r,'
.

It is given by yεK(8)). The above branch set is formed a posteriori through experience 1, depending on how each code is scored.

TI (910), T2 (920) on the y-axis in Figure 9
are respectively code 1. This is code 2 territory.

(2) Automatic setting of scores: In the above embodiment, the actual measured values of the characteristics of the code were presented to the user, and the user manually converted the values into two scores.

This operation can be performed using, for example, two functions Jl''hl(α, .

α1. ...6 mouths, Js = ht (α1
, α3.・・・・・・,α,) Make it into a meeting program,
Put it in the procedure (23゜1) and use this as J.
, J. It can be automated by making gold needles. Here, α, ~αJ are the actual characteristics of the code (l
Hail i.

There are many other procedures that present a series of parameters to the user and then allow the user to make necessary corrections. child! Regarding Shira, if possible, for example, by using a method of automatically correcting parameter values using random number tones, it is possible to reduce opportunities for user intervention in the system and increase the degree of automation.

(5) Effects of the Invention According to the present invention, instead of programming the problem-solving procedure in advance, it becomes possible to program it gradually after the fact. This is gradually known #! This means that it is possible to acquire all of them.

Furthermore, the acquired knowledge can be gradually improved to improve its characteristics.

This is effective in creating systems that gradually learn how to understand and operate objects, such as robots capable of diagnosing system abnormalities, controlling systems, and learning. It is.

[Brief explanation of drawings]

Claims (1)

[Claims] 1. In the Knowledge Pace system, which stores and utilizes codes consisting of conditions and procedures to be performed under those conditions, when a condition that Knowledge Pace has no experience with arises, it is possible to perform the procedure under those conditions. By using a mother code that has the function of generating a new procedure, the code has the function of generating a mutated copy of itself, and the above code is given a score indicating whether its characteristics are good or bad, and the generation time is set to 0. By giving the knowledge base a function to weed out codes with poor characteristics and old codes from the knowledge pace using the above rating of the above code and the above age, the knowledge base is given an age that increases from then on. A knowledge base construction method that waits for gradual acquisition ability and knowledge improvement ability.