JPH01154231A - Optimization system for production system - Google Patents

Abstract

PURPOSE:To minimize the join arithmetic value by optimizing the method (join topology) for sequence and combination of the join arithmetic operations at the condition part of each rule of a production system based on the dynamic characteristics of a program. CONSTITUTION:An initial topology is produced with input of an unoptimized production rule and a set of nodes is obtained. Then two nodes are selected out of said set of nodes for production of a test node which performs the joint arithmetic operation. The cost of the test node is evaluated. Then the test node is eliminated if it is not necessary and then added to the set of nodes if necessary. Then the termination node having the lowest cost is selected out of the set of nodes and an optimized production rule is produced based on the selected node. The cost of the test node is evaluated by applying the equivalent conversion to the sequence and combination of the join arithmetic operations for decision of conditions based on the measurement data on the dynamic characteristics of a production.

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention relates to an optimization method for a production system, and particularly to a method for minimizing the amount of join operations and speeding up the processing of a production system. It is something.

[Conventional technology]

The production system is a highly independent set of rules.
It is a system that expresses knowledge through a collection. The production system is composed of a production memory (hereinafter abbreviated as PM) that stores rules and a working memory (hereinafter abbreviated as WM) that stores data necessary for executing the rules.

Note that the data in the WM is called a working memory element (hereinafter abbreviated as WME). Each rule consists of a concatenation of conditional elements called a left-hand side and a right-hand side.
The execution of the production system is realized by repeating a cycle consisting of the following three phases.

(b) For the Match→All rule, compare the left side with the WM at that point.

(b) Conflict Re5solution→
One of the successfully matched rules is selected according to a predetermined strategy.

(c) Act→Execute the selected rule to add and remove data.

Conventionally, the condition part of a rule is converted into a data flow graph called a Rete network (for a Rete network, for example, [rRete: A high-speed algorithm Jl (F Orgyt C-L v'A
Fast Algorithm for
the ManyPattern/Many 0b
eject Pattern Match Probl
em', Artificial Intelligence
nce, Vol, 19°pp, 17-37 (1982)
reference).

When data is added to the WM in the Act phase, the data is poured into the Rete network, and the network is updated in accordance with the change. This update (Matc
(corresponding to the h phase) is proceeded in the following steps. First, a test (1-1nput-test) that is completed within one condition is performed for each condition on the left side, and data that passes the test is stored in the network. Thereafter, a test h (2-1nput-test) is performed to check whether the values of pattern variables are consistent between one condition or not, and data that passes the test is stored in the network. t of the data that satisfies all the conditions on the left side
fl (instantiation) reaches the terminal of the Rete network and makes the corresponding rule fireable. Once the Rete network update is complete, processing will proceed to Conflict Reso
Move to the solution phase.

[Problem that the invention seeks to solve]

In conventional production systems, the order and combination method (j.

in topology).

However, in production systems, performance deteriorates rapidly as the amount of data in working memory increases. This is a series of JOIs included in the condition determination of the rule.
In a normal processing system, the time required for n operations is 2 times the amount of data.
This is thought to be due to the following: (1) Inappropriate description of conditions generates a large number of intermediate calculation results, further increasing the amount of processing for join calculations. To solve this problem, for example, r Art Programming Tutorial Jl (Clayton, B, D
, , ”ARTP programming
trial”I nference Corp.

(1987), etc., the user can specify the order of join operations and the combination method (j) in the condition part of the rule.

in topology) can be specified freely.

However, in the end, production systems require automatic optimization functions for the following reasons.

(a) It is not easy to program an efficient join topology, and it is too burdensome for users to perform optimization.

(b) Since the optimal join topology depends on the operating characteristics of the production system program, it is necessary to measure statistical data during execution and make a selection based on it. This is true even if the rules are the same, but the target data is different (for example.

In the fault diagnosis system for switching equipment, if the target model is different, the optimal join topology will also be different. In other words, optimization is not only a problem on the rule development side, but also needs to be performed on the expert system operation side.

(c) Optimized programs have poor readability. In other words, in order to obtain short-term performance improvements, source programs are changed at the expense of long-term maintainability, but this has many problems.

As described above, in the conventional method, the join topology was specified manually, which caused many problems.

The purpose of the present invention is to solve these conventional problems and to
An object of the present invention is to provide a production system optimization method that can execute a production system program at high speed by automatically optimizing the in topology.

[Means for solving problems]

In order to achieve the above object, the production system optimization method according to the present invention inputs pre-optimization production rules in a production system that expresses knowledge by a set of rules consisting of a concatenation of one conditional element and a set of actions. a first process of creating an initial topology and generating a set of nodes; a second process of selecting two nodes from the set of nodes and generating a test node to perform a join operation; Test node cost 1-
A third process is to evaluate the node, discard it if unnecessary, and add it to the set of nodes if necessary, and select the terminal node with the lowest cost from the set of nodes. and a fourth process of generating post-processing production rules, and in the third process, the order and combination method of join operations in condition determination are equivalently converted based on the measured data of the operational characteristics of the production system. The feature is that the cost of the death 1-node is evaluated.

[For production]

In the present invention, the order and combination method of join operations in the condition part of each rule of the production system (these are referred to as joinh topology) is equivalently converted based on measurement data of the operational characteristics of the production system. By doing so, the amount of join calculations is minimized. Also, multiple rules are optimized at once, and join operations are shared as much as possible across the rules. To this end, we use an inference engine, a process that measures the operating characteristics of the production system.

Initial topology creation process, 2-1nput-tes
A t-node generation process, a cost evaluation process, and an optimal production rule generation process are provided. This avoids creating topologies that are more costly than join topologies that have already been created, and eliminates expensive topologies that already exist. Then, nodes with common join variables are preferentially combined, and 1-1n
token and ll1eIo in put-test.

Join operations are executed in order from the smallest ry. As a result, a post-optimization production rule is generated, thereby making it possible to minimize the amount of join calculations.

〔Example〕

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

Figures 2 (1) to (4) show the j used in the present invention.
It is an explanatory diagram showing an example of oin topology.

In production systems, ConflictRe
Due to solution issues, the position of conditional element (a) shown in Figures (1) to (4) cannot be changed.6 In what order should the other conditional elements (b), (c), and (d) be placed? For example, Figures (1) to (4)
The following combinations are possible. In Figure (1), the whole is satisfied by satisfying conditional elements (a) and (b), satisfying both the result and (c), and satisfying both the result and (d). In addition, in Figure (2), the result of satisfying conditional elements (a) and (b) and one conditional element (C)
and (d) are satisfied. Also,
In Figure (3), the result of satisfying conditional elements (b) and (c), the satisfaction of both (a), and the result of satisfying both (d). Furthermore, in Figure (4), the conditional element (b
) and (C), satisfy both (d), and satisfy both that result and (a). In addition, (a
) to (d) are 1-1nput -test nodes, and the node that performs their join operation is 2-1nput
- test node.

Considering the combinations in this way, since there are six combinations for each of (1) to (4), there are 24 join topologies (generally, the number of combinations is on an exponential order).

In the present invention, the one with the minimum cost is selected from these join l-topologies.

FIG. 3 is an explanatory diagram of the cost calculation method used in the present invention.

FIG. 3 shows a cost model for join operations. The cost model consists of the following parameters.

■ token→Total number of data that has reached each node.

∎memory→This is the average value of data held in each node.

J) j o i n → 2 executed on each node
-1nput is the number of tests.

(f C09j→2 executed up to each node
-input - This is the total number of times of testing.

■ ratio → 2-1nput -te at each node
st is the success rate.

FIG. 3 shows calculation models for each case.

(i) 2-1 to at nput-test node
When ken, join→toker1 is input from the left, the number of tokens (Ta) multiplied by the number of data in the right memory (Mb) is 2-2 of the total number (TaMb).
1nput-test is performed. Also, token
When is the right input, the opposite is true (T b M a ). The sum of both (T a M b + T b
M a ) is executed on this node 2 −1np
It is the number of times (Jc) of ut-test. In the case of a positive condition element of the right input, the number of tokens (Tc) generated as a result of the test is a value (JcRc) obtained by multiplying the number of 2-1nput-testM by the test success probability (Rc). On the other hand, to sing a negative conditional element, the left input token
(T a ) becomes the filtered value (TaRc).

(ji) 2-1 M at nput-test node
memory→The average number of data (Me) is 2-1npu when the right input is a positive conditional element, the average number of data in the memory of the left and right nodes is multiplied by the combined value (MaMb).
t-test success probability (Rc) multiplied by (Ma
MbRc). . On the other hand, when the negative conditional element is the right input, the value (MaRc) is the result of filtering the left input memory (M a ).

(iii) 2-1nput -Cost at the test node -> The cost (Ca, Cb) of the left and right nodes plus the calculated cost 1- of the own node. 2-1np
The computational cost of ut -test is generally A's J
It can be expressed as c + B * T c. Here, A and B are appropriate constants. In this embodiment, the goal is to minimize the total number of 2-1 nput-tests, and Cc=Ca+Cb+Jc (ie, A=1.B=0). However, if the data is hashed, then 2-1nput -te
The problem is the number of tokens generated by st. In such a case, for example, Cc = Ca + Cb + T c
It is necessary to appropriately set the cost calculation formula according to the processing system.

(iv) Ratio at 2-1nput-test node→2-1nput-test success probability (
Re) has an approximate value of 1-1nput of the right input.
- Rati of the test node.

(Rb) is used.

In the example shown in FIG. 3, the goal is to reduce the number of times 2-1 nput-test is performed, so cost is simply set to the sum of the times. That is, the conditional element (
When satisfying a) and (b), each parameter of result (c) is tokenTc= (TaMb+TbMa
)Re, that is, each token has the other party's Memory
The sum of the products multiplied by (c) is multiplied by Ratio. Also, Memory is M c = M a M b
Rc, that is, each memory and Rati of (c)
The value is multiplied by o. Also, join is Jc=TaM
b + TbMa, that is, each token has the other party's Mem
It is multiplied by ory. Also, Cost is Cc=
Ca+Cb+Jc, that is, the sum of each cost plus the join of (C). Also, Ra
tio is equal to Rc=Rb, that is, Ratio in (b).

The index expressing performance is not necessarily 2-1nput -tes
The sum of t times is not necessarily good. For example, successful 2
If only -1nput -test determines the performance, it is necessary to set it appropriately according to the processing system, such as setting Cc=Ca+Cb+Tc as described above. Furthermore, in order to set initial values for Ta, Tb, etc. in the cost model, the rules created by the user are executed once to collect statistical data.

In this way, the optimization algorithm in this example is
Basically, possible jojn topologies are generated one after another and the one with the minimum cost is selected using a cost model. The reduction strategy for join topologies that exist in exponential order is explained with reference to FIG.

FIG. 1 is an optimization function block diagram showing an embodiment of the present invention, and FIG. 4 is a block configuration diagram of a computer system for executing the functions shown in FIG.

Inference engine 2 in Figure 1. Operation characteristic measurement processing 3,
The initial topology creation process 4, the input test node generation process 8, the cost evaluation reduction process 10, and the optimal production rule generation process 11 are all executed in the processor (C, PU) 21 of the computer system shown in FIG. Data input for these processes 1, 7.
9 etc. are all read out from the main storage device 22. Also, before being loaded into the main memory 21, or before being loaded into the main memory 21,
1 to the outside, the file control device 24
The data is stored in the file device 25 via.

Input/output processing is performed by the channel device 23, and input/output operations are performed by the input/output device 27 via the input/output control device 26. Note that 28 is a common bus to which each device is connected.

Inference engine 2 uses production rule 1 before optimization.
As input, inference result 5 is output.

During this time, necessary data is collected by operating characteristic measurement processing 3 and measurement data 6 is output. This measurement data is sent from the main storage device 22 to the file 25 and stored therein.

The optimization function is then activated.

First, in the initial topology creation process 4, a set of nodes 7 is generated using the pre-optimization production rule 1 as input. 2-1 nput-test node generation process 8
Select two nodes from the set of nodes 7 by
A 2-1nput-test node 9 that performs a join operation is generated. The cost evaluation/reduction process 10 evaluates the cost of the generated 2-1nput-test node 9, discards it if unnecessary, and adds it to the node set 7 if necessary. Processes 7 to 10 are repeatedly executed until all combinations of nodes in the node set 7 are exhausted (returns to 7 through the thick line for feedback). After that, optimal production rule generation process 1
1 selects the lowest cost termination (
(terminal) note and generates the post-optimization production rule 12 based on it. The post-optimization production rule 12 is stored in the file device 25 from the main storage device 22 via the file control device 24.

The following is used as a reduction strategy for joinhbology that exists in exponential order.

(b) Do not generate a topology that costs more than the join topology that has already been generated. If it has already been created, remove it. To effectively utilize this strategy, larger, lower-cost join topologies can be generated quickly.

(b) Perform the join operation with as many constraints as possible. That is, nodes having a common join variable are preferentially combined.

(c) Among 1-1nput-test, toke
J oin operations are performed in the order of n and memory.

In the experiment, we prototyped and evaluated an optimization function that takes as input the program and operating characteristics written by the user and outputs an optimized program. As a result of optimizing the existing expert system, 2-1nput -test
The total number of CPUs has been reduced to 173, and the CPU performance has been approximately doubled. What is particularly noteworthy is that the results outperformed past manual performance improvements.

Generally, as the number of rules and the number of data increases, the optimization effect further increases. Therefore, it is believed that the present invention will function more effectively as expert systems become more practical.

〔Effect of the invention〕

As described above, according to the present invention, it is possible to automatically optimize the join topology based on the operating characteristics of the program, instead of manually specifying the join topology, and the join operation Since the amount is minimized, production system programs can be executed quickly and at low cost.

[Brief explanation of the drawing]

Fig. 1 is a functional block diagram of a production system showing an example of the present invention, and Fig. 2 is a functional block diagram of a production system used in the present invention.
Figure 3 shows an example of the oin topology, Figure 3 is an explanatory diagram showing the cost calculation method used in the present invention, Figure 4 shows the example of the
1 is a block diagram of a computer system that performs the functions shown in the figure. FIG. 21: CPU, 22: Main storage device, 23: Channel device, 24: File control device, 25: File device, 26
Two-person output control device, 27: Input/output device, 28: Common bus, 2: Inference engine, 3: Operating characteristic measurement process, 4: Initial topology creation process, 8: 2-1nput-tes
t-node generation processing, 10: Cost evaluation/reduction processing, 11
: Optimal production rule generation process. Patent applicant Nippon Telegraph and Telephone Corporation Figure 2 (d)) Figure 3 Token: 'l'c AIemory: bic Join: Jc Cost: CQ 1(atio: Rc Tc=(TaMb+TbMa)Re Me=MaMbRC JC=Ta %jb+'l'blL(a CC=Ca+Cb×JC Rib=Rb Fig.

Claims (1)

[Claims] (1) In a production system that expresses knowledge by a set of rules consisting of a connection of conditional elements and a set of actions, the first step is to input pre-optimization production rules, create an initial topology, and generate a set of nodes. process, select two nodes from the set of nodes, and select jo
a second process of generating a test node that performs an in operation;
A third process of evaluating the cost of the generated test node, discarding it if unnecessary, and adding it to the set of nodes if necessary, and selecting the terminal node with the lowest cost from the set of nodes. , and a fourth process of generating post-optimized production rules based on this. In the third process, the order of join operations in condition determination and the A production system optimization method characterized by equivalently converting combination methods to evaluate test node costs.