Patent Application
20240232667
The present invention relates to an information processing device, an information processing method, and an information processing program.
Artificial intelligence for IT operations (AIOps) for achieving automation and efficiency of IT businesses by causing AI to learn various types of data such as big data used in IT business is known. For example, there is an orchestrator that sequentially uses AI that identifies an event, AI that identifies a cause of occurrence of the event, AI that searches for a method of coping with the event, AI that identifies an influence of the event, and AI that determines a method of coping with the event.
In the AIOps, when the AI makes an error in determination, there is concern of service quality deteriorating due to a longer time of failure recovery, and of it being difficult to withstand an actual operation. In order to correct the error in the determination of AI, it is necessary for an advanced operator with an operation skill, a machine learning skill, and mathematical knowledge to take a complicated measure. Therefore, there is a need for a technology for establishing an AI teaching technique to infer validity of output results of various types of AI and implementing safe and secure operation by speeding up failure recovery and allowing operators to be less skilled.
Accordingly, there is a technology in which AI can be safely embedded in an NW operation by verifying the determination of the AI and explaining the reason for the determination by a Bayesian network that simulates the flexible determination of an experienced operator. In the Bayesian network software, design and operation of the Bayesian network can be performed on a GUI basis.
When a Bayesian network is designed, a node indicating a determination element and an edge (arrow line) indicating a causal relation between nodes are set, and a probability of a slave node is set in a conditional probability table (CPT). For example, in the case of a Bayesian network regarding determination of Loses, Sick, Dry, and Loses are set, and Sick and Dry are set as a connection source and Loses are set as a connection destination. An occurrence probability and a non-occurrence probability of Loses are set for a combination of an index value or a probability of Sick and an index value or a probability of Dry (see Non Patent Literature 1).
In the Bayesian network, a determination model in which human experience is reflected can be designed by a graphical model by nodes, edges, and CPT. However, the node and the edge can be intuitively set based on knowledge, but a probability of a slave node set in the CPT is manually calculated based on the statistical information of operational data. Therefore, there is a problem that an amount of work at the time of generation of the CPT is large, and thus a high degree of mathematical knowledge regarding conditional probabilities such as Bayesian inference and mathematical knowledge for shaping data according to a statistical method is required.
The present invention has been made in view of the above circumstances, and an objective of the present invention is to provide a technology capable of designing a Bayesian network without requiring mathematical knowledge.
According to an aspect of the present invention, an information processing device that designs a Bayesian network includes: an input unit configured to input, as elements of a Bayesian network, information regarding a plurality of nodes, an edge indicating a causal relation between a master node and a slave node in the plurality of nodes, and a weight indicating the degree of influence of a definition content of the master node on a probability distribution of the slave node; and a generation unit configured to generate a Bayesian network using the input information.
According to another aspect of the present invention, an information processing method performed by an information processing device that designs a Bayesian network includes: a step of inputting, as elements of a Bayesian network, information regarding a plurality of nodes, an edge indicating a causal relation between a master node and a slave node in the plurality of nodes, and a weight indicating the degree of influence of a definition content of the master node on a probability distribution of the slave node; and a step of generating a Bayesian network using the input information.
According to still another aspect of the present invention, an information processing program is an information processing program that causes a computer to function as the information processing device.
According to the present invention, it is possible to provide a technology capable of designing a Bayesian network without requiring mathematical knowledge.
Hereinafter, an embodiment of the present invention will be described with reference to the drawings. In the drawings, the same portions are denoted by the same reference signs, and description thereof is omitted.
According to the present invention, in addition to a node, an edge, and a CPT, a weight indicating the degree of influence of definition content of a master node on a probability distribution of the slave node is further input as a definition element of the Bayesian network, and the Bayesian network is generated using the input information. That is, since the Bayesian network is generated based on the weight indicating the degree of influence of the definition content of the master node on the probability distribution of the slave node, it is possible to design the Bayesian network in which human experience is reflected without requiring mathematical knowledge.
According to the present invention, a CPT of a slave node is automatically generated based on the weight. Accordingly, it is possible to reflect the business knowledge, the business experience, and the like in the Bayesian network only with a parameter of the weight without directly setting the CPT. As a result, it is possible to design a Bayesian network in which human experience is appropriately reflected without requiring mathematical knowledge.
According to the present invention, the average value of the probability distribution is varied in accordance with the range of probability values which the slave node is able to take in accordance with the weight and the probability of the master node, and a bias of the probability distribution of the slave node is adjusted based on the varied average value. Accordingly, it is possible to reflect the business knowledge, the business experience, and the like in the Bayesian network only with a parameter of the weight without directly setting the CPT. As a result, it is possible to design a Bayesian network in which human experience is more appropriately reflected without requiring mathematical knowledge.
[Configuration of Information Processing Device]
The input unit 11 is a functional unit that inputs, a definition elements of a Bayesian network, information regarding a plurality of nodes indicating determination elements, an edge (arrow line) indicating a causal relation between a master node and a slave node in the plurality of nodes, a CPT for setting a probability (for example, an occurrence probability and a non-occurrence probability) of the slave node, and a weight indicating the degree of influence of definition content of the master node on a probability distribution of the slave node.
In the present embodiment, as a probability of a slave node, an occurrence probability and a non-occurrence probability of the slave node are used as an example. The slave node can take three or more values in addition to such two values. The master node can also take two or more values.
The generation unit 12 is a functional unit that forms a Bayesian network using the input information. The generation unit 12 is a functional unit that generates a CPT of the slave node based on the weight. For example, the generation unit 12 generates the CPT of the slave node by calculating the occurrence probability and the non-occurrence probability of the slave node based on the weight and setting the occurrence probability and the non-occurrence probability in the CPT. For example, the generation unit 12 generates a CPT of a slave node by calculating an average value of a probability distribution in accordance with a range of probability values which the slave node is able to take based on the weight, calculating the occurrence probability and the non-occurrence probability of the slave node based on the probability distribution of the slave node that has the average value, and setting the probability and the non-occurrence probability in the CPT. When the occurrence probability and the non-occurrence probability of the slave node are calculated, the generation unit 12 varies the average value of the probability distribution in accordance with the range of the probability value that the slave node can take in accordance with the weight and the index value or random variables (for example, the occurrence probability and the non-occurrence probability) of the master node, and adjusts a bias of the probability distribution of the slave node based on the varied average value.
The display unit 13 is a functional unit that displays a definition element of the Bayesian network, a design diagram of the Bayesian network, a CPT, and the like on a screen.
[Operation of Information Processing Device]
Step S1;
The display unit 13 first displays a GUI for setting a definition element of the Bayesian network on a screen. Next, the input unit 11 inputs a plurality of nodes indicating a determination element, an edge indicating a causal relation between a master node and a slave node in the plurality of nodes, and a weight indicating the degree of influence of the definition content of the master node on the probability distribution of the slave node, which are input to the GUI by the user.
For example, a node X1 of Pollution, a node X2 Smoker, and a node X3 of Cancer are input. In addition, an edge E1 that has the node X1 as a connection source and has the node X3 as a connection destination and an edge E2 that has the node X2 as a connection source and has the node X3 as a connection destination are input. Furthermore, a weight w1 for the edge E1 and a weight w2 for the edge E2 are input.
The weights w1 and w2 increase weights that are desired to have a greater influence on the slave node X3 between the master nodes X1 and X2. For example, since it is considered that Smoker has a higher probability of causing Cancer than Pollution (Smoker has a stronger influence on Cancer), 0.7 is input as the weight w2 and 0.3 is input as the weight w1. A sum of the weights w: and w2 is 1.
Accordingly, the Bayesian network illustrated in
Step S2;
Next, the generation unit 12 calculates the probability distribution of the slave node X3 using Expressions (1) and (2).
Expression (1) is an expression that defines a probability distribution f(z) of the slave node X3. Expression (2) is an expression that defines an average value μ of the probability distribution f(z) using the weight w input in step S1. z is a random variable of the slave node X3. σ is a variance value of the probability distribution f(z). xj (where j ∈ 1, 2, 3) is an index value or a random variable of the node Xj. M is a total value of cardinality numbers of the master nodes X1 and X2. The probability distribution f(z) of the slave node X3 defined in Expressions (1) and (2) is a normal distribution, as illustrated in
As in Expression (2), the average value μ of the probability distribution f(z) is obtained by a product of the weight w and the index value or the random variable of the node X. Therefore, the probability distribution f(z) of the slave node X3 is biased such that the probability of the slave node is a higher value as the index value or the random variable of the node X is larger. The probability distribution f(z) of the slave node X3 is biased such that the probability of the slave node is a higher value as the master node of the edge has a larger weight w.
As in Expression (2), a sum of a product of the weight w and the index value or the random variable of the node X is divided by M. Therefore, when the cardinality value of the slave node tends to increase as the cardinality value of the master node increases, the occurrence probability of an initial value with high accuracy is obtained.
By using Expressions (1) and (2), the generation unit 12 varies the average value μ of the probability distribution in accordance with the range of the probability value which the slave node X3 is able to take in accordance with the weight w and the index value or the random variable of the master node X, and adjusts a bias f(z) of the probability distribution of the slave node X3 based on the varied average value p.
Next, the generation unit 12 calculates a probability P of the slave node X3 using Expression (3).
For example, when the index value of the slave node X3 is max (X3), that is, 1, the generation unit 12 sets a value integrated in a range from max (X3) to +∞ as the occurrence probability of the slave node X3. The generation unit 12 calculates the occurrence probability and the non-occurrence probability of the slave node X3 for all the combinations of the possible index values which the three nodes X: to X3 are able to take.
Finally, the generation unit 12 sets the occurrence probability and the non-occurrence probability of the slave node X3 of all the combinations in the CPT (see
Expressions (1) to (3) are related to each other and can be said to be an expression group for generating the CPT of the slave node based on the weight and an expression group for calculating the average value of the probability distribution of the slave node based on the weight and generating the CPT of the slave node based on the probability distribution of the slave node that has the average value.
According to the present embodiment, the information processing device 1 includes, the input unit 11 configured to input, as elements of a Bayesian network, information regarding a plurality of nodes, an edge indicating a causal relation between a master node and a slave node in the plurality of nodes, and a weight indicating the degree of influence of a definition content of the master node on a probability distribution of the slave node; and the generation unit 12 configured to generate a Bayesian network using the input information Therefore, it is possible to design the Bayesian network in which human experiences are reflected without requiring mathematical knowledge.
[Others]
The present invention is not limited to the foregoing embodiments. The present invention can be modified in various forms within the gist of the present invention.
For example, as illustrated in
The information processing device 1 may be mounted by one computer. The information processing device 1 may be mounted by a plurality of computers. The information processing device 1 may be a virtual machine that is mounted on a computer. The program for the information processing device 1 can be stored in a computer-readable recording medium such as an HDD, an SSD, a USB memory, a CD, or a DVD. The program for the information processing device 1 can also be distributed via a communication network.
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/JP2021/007449 | 2/26/2021 | WO |
| Number | Date | Country | |
|---|---|---|---|
| 20240135213 A1 | Apr 2024 | US |
