CN111753892A - An Interpretation Method for Global View Network System Based on Deep Learning - Google Patents

Abstract

The invention relates to the field of Internet information technology, in particular to an interpretation method of a global view network system based on deep learning. The method of the invention carries out causal interpretation and transformation for the decision of the computer network system based on deep learning in the global perspective. First, the original network system is trained by the method of deep reinforcement learning. After completing the training of the original deep learning-based system, the generated global configuration results are modeled by means of a hypergraph, and the key points in the hypergraph are analyzed- Hyperedge connections, which score the impact of each point-hyperedge connection on the final global configuration result, enabling network administrators to understand key components of decision-making. This method greatly reduces the difficulty of understanding the original global vision network system based on deep learning, and facilitates network administrators to understand the decision-making process. When the explanation method is deployed on the actual system, it is helpful for network administrators to understand and correct the decision-making process of the original global view network system.

Description

An Interpretation Method for Global View Network System Based on Deep Learning

technical field

The invention relates to the field of Internet information technology, in particular to an interpretation method of a global view network system based on deep learning.

Background technique

Computer network systems can generally be divided into systems with local vision and systems with global vision. Local vision refers to the deployment of network systems on servers, clients, middleware, and switches (such as congestion control systems). These systems with partial vision can only observe and make decisions by observing the information of a point in the system. decision making. Systems with a global view are such as network management controllers, traffic engineering scheduling systems, software-defined network controllers, etc. These systems with a global view can observe and make decisions on multiple devices in the network. In terms of decision-making logic, some existing global-view network systems use deep learning technology as their decision-making algorithms, such as software-defined network routing optimization algorithms based on deep neural networks.

In the decision-making strategy of the traditional network system, decisions are often made through simple strategies such as "increase, multiply and decrease". These existing deep learning-based systems make decisions in a way that network administrators cannot understand: neural networks often contain thousands of neurons, and come to a final conclusion after a series of non-linear calculations. Therefore, despite good performance in training, network administrators are not able to understand the logic of their decisions and thus often have difficulty gaining trust.

SUMMARY OF THE INVENTION

The purpose of the present invention is to propose an interpretation method for a global vision network system based on deep learning, which can causally interpret and transform the decision of a computer network system based on deep learning in the global vision situation. First, the deep reinforcement learning method is used to train the original network system. After completing the training of the original deep learning-based system, the global configuration results generated by it are modeled by hypergraph, and the hypergraph is analyzed. Key point-hyperedge connections, scoring the influence of each point-hyperedge connection on the final global configuration result, to allow network administrators to understand the critical components of the decision.

The interpretation method of the deep learning-based global view network system proposed by the present invention includes the following steps:

(1) Input resources and requests into the global vision network system S to be explained, and output a global configuration result set, which is denoted as O;

(2) Construct a global view computer network system hypergraph H including resources and requests, and the global view computer network system hypergraph H has the following four forms:

a. When the global view network system S is a software-defined network routing optimization system, the physical links in the hypergraph H are constructed as points, and the routing paths are constructed as hyperedges;

b. When the global view network system S is an optimization system for virtual network function placement, the physical server in the hypergraph H is constructed as a point, and the virtual network function is a hyperedge;

c. When the global view network system S is an ultra-dense cellular network optimization system, the mobile users in the hypergraph H are constructed as points, and the base stations are constructed as hyperedges;

d. When the global vision network system S is a cluster task scheduling optimization system, the requested tasks in the hypergraph H are constructed as points, and the dependencies between tasks are constructed as hyperedges;

An association matrix I is used to represent the hypergraph H of the global view computer network system. The association matrix I is a V×E dimension 0-1 matrix. When the position (v, e) in the association matrix I is 1, it indicates that the point v and the hypergraph are The edge e has a connection relationship, and when the position (v, e) is 0, it means that the point v has no connection relationship with the hyperedge e;

(3) carry out feature calculation to the global view computer network system hypergraph H of step (2), including the following steps:

(3-1) Construct an evaluation matrix W, which is used to characterize the importance of each non-zero element in the correlation matrix I;

(3-2) Calculate the performance loss of the above evaluation matrix W:

Judging the global configuration result set O in step (1), if the output result of the global configuration result in step (1) is a discrete variable, then use the following formula to calculate the KL divergence D(W, I); If the global configuration result output result in step (1) is a continuous variable, then utilize the following formula to calculate the mean square error D(W,I) of the global configuration result correlation matrix I and the evaluation matrix W:

In the above formula, f(W) is the evaluation result of the evaluation function of the global view network system S to be explained on the evaluation matrix W, and f(I) is the evaluation result of the correlation matrix I of the global view network system S to be explained;

(3-3) Use the following formula to calculate the simplicity of the evaluation matrix W in step (3-1), and define the simplicity of the evaluation matrix W as:

In the above formula, W _ev is the element value of the evaluation matrix W at the (e, v) position of the correlation matrix I;

(3-4) The entropy H(W) of the evaluation matrix W is used to characterize the certainty of the evaluation matrix W:

(3-5) Establish an optimization model for solving the evaluation matrix W, and the objective function of this optimization model is:

minD(W,I)+λ ₁ ||W||+λ ₂ H(W)

The constraints of the optimization model are:

Among them, D(W,I) is the mean square error in step (3-2), ||W|| is the simplicity in step (3-3), and H(W) is in step (3-4) The entropy of the evaluation matrix W, λ ₁ and λ ₂ are the calculation parameters of simplicity and entropy, respectively;

(4) Using the gradient descent method to solve the optimization model of step (3-5), the evaluation matrix W is obtained, and according to the evaluation matrix W, the representation of the importance of each connection relationship in the association matrix I is obtained, and the global vision network based on deep learning is realized. System explanation.

The features and advantages of the deep learning-based global view network system interpretation method proposed by the present invention are:

The interpretation method of the global vision network system based on deep learning of the present invention converts the output of the global vision network system based on deep learning into an equivalent hypergraph, and quantitatively displays the importance of each connection relationship in the hypergraph , which greatly reduces the difficulty of understanding the original global vision network system based on deep learning, and facilitates network administrators to understand the decision-making process of the results. When the explanation method is deployed on the actual system, it is helpful for network administrators to understand and correct the decision-making process of the original global view network system.

Description of drawings

FIG. 1 is a global routing result O of a network routing system S constructed in an embodiment of the method of the present invention.

FIG. 2 is a hypergraph H converted from a global routing result O in an embodiment of the present invention.

FIG. 3 is an important connection relationship result obtained in the embodiment of the present invention.

Detailed ways

The interpretation method of the deep learning-based global view network system proposed by the present invention includes the following steps:

(1) Input resources and requests into the global vision network system S to be explained, and output a global configuration result set, which is denoted as O; for a deep learning-based software-defined network routing optimization system, The set O of the configuration results is the path that the traffic at any two points in the network should be forwarded along;

(2) Constructing a global view computer network system hypergraph H including resources and requests. A network system with a global view is the allocation of resources and requests, such as link resources are allocated to traffic requests, physical machine resources are allocated to virtual machine services request and so on. Therefore, resources and requests can be represented as points and hyperedges in a hypergraph, respectively. The global view computer network system hypergraph H has the following four forms:

a. When the global view network system S is a software-defined network routing optimization system, the physical links in the hypergraph H are constructed as points, and the routing paths are constructed as hyperedges;

b. When the global view network system S is an optimization system for virtual network function placement, the physical server in the hypergraph H is constructed as a point, and the virtual network function is a hyperedge;

c. When the global view network system S is an ultra-dense cellular network optimization system, the mobile users in the hypergraph H are constructed as points, and the base stations are constructed as hyperedges;

d. When the global vision network system S is a cluster task scheduling optimization system, the requested tasks in the hypergraph H are constructed as points, and the dependencies between tasks are constructed as hyperedges;

An association matrix I is used to represent the hypergraph H of the global view computer network system. The association matrix I is a V×E dimension 0-1 matrix. When the position (v, e) in the association matrix I is 1, it indicates that the point v and the hypergraph are The edge e has a connection relationship, and when the position (v, e) is 0, it means that the point v and the hyperedge e have no connection relationship, as shown in Figure 1;

(3) carry out feature calculation to the global view computer network system hypergraph H of step (2), including the following steps:

(3-1) Construct an evaluation matrix W, which is used to characterize the importance of each non-zero element in the correlation matrix I;

(3-2) Calculate the performance loss of the above evaluation matrix W:

Judging the global configuration result set O in step (1), if the output result of the global configuration result in step (1) is a discrete variable, then use the following formula to calculate the KL divergence D(W, I); If the global configuration result output result in step (1) is a continuous variable, then utilize the following formula to calculate the mean square error D(W,I) of the global configuration result correlation matrix I and the evaluation matrix W:

In the above formula, f(W) is the evaluation result of the evaluation function of the global view network system S to be explained on the evaluation matrix W, and f(I) is the evaluation result of the correlation matrix I of the global view network system S to be explained;

(3-3) Use the following formula to calculate the simplicity of the evaluation matrix W in step (3-1), and define the simplicity of the evaluation matrix W as:

In the above formula, W _ev is the element value of the evaluation matrix W at the (e, v) position of the correlation matrix I;

(3-4) The entropy H(W) of the evaluation matrix W is used to characterize the certainty of the evaluation matrix W: at the same time, it is also hoped that the importance should be aggregated to 0 or 1 to indicate the dichotomous nature of the importance: either important or unimportant .

(3-5) Establish an optimization model for solving the evaluation matrix W, and the objective function of this optimization model is:

minD(W,I)+λ ₁ ||W||+λ ₂ H(W)

The constraints of the optimization model are:

Among them, D(W,I) is the mean square error in step (3-2), ||W|| is the simplicity in step (3-3), and H(W) is in step (3-4) The entropy of the evaluation matrix W, λ ₁ and λ ₂ are the calculation parameters of simplicity and entropy, respectively; λ ₁ and λ ₂ can be set according to the network administrator's preference for the corresponding properties. In an embodiment of the present invention, λ ₁ The values of and λ ₂ are 0.5 and 0.5;

(4) Using the gradient descent method to solve the optimization model of step (3-5), the evaluation matrix W is obtained, and according to the evaluation matrix W, the representation of the importance of each connection relationship in the association matrix I is obtained, and the global vision network based on deep learning is realized. System explanation.

Below in conjunction with the accompanying drawings, the content of the method of the present invention is described in detail:

Set a software-defined network routing system S based on deep learning, and the resulting routing decision results are shown in Figure 1. In Figure 1, a to g are routers, 1 to 8 are physical links, and blue and red are a to The routing paths from e and a to g are calculated by the network routing system S. Construct a hypergraph H of a network routing system S, in which physical links are constructed as points and routing paths are constructed as hyperedges; as shown in Figure 2. By calculating the evaluation matrix W, the key influence on the final optimization objective in the point-hyperedge connection in the above hypergraph can be obtained. For example, the algorithm finds that the path selection link 6 has a greater impact on the global optimization results, and can prompt the network administrator to pay attention to link 6 as a performance bottleneck. Similar operations can be done for larger-scale topologies, and the final possible visualization results are shown in Figure 3.

Claims (1)

1. A deep learning-based interpretation method of a global visual field network system is characterized by comprising the following steps:

(1) inputting resources and requests into a global view network system S to be interpreted, outputting to obtain a global configuration result set, and recording the global configuration result set as 0;

(2) constructing a global-view computer network system hypergraph H comprising resources and requests, the global-view computer network system hypergraph H having the following four forms:

a. when the global visual field network system S is a software defined network routing optimization system, physical links in the hypergraph H are constructed into points, and routing paths are constructed into hyperedges;

b. when the global visual field network system S is used for placing an optimization system for a virtual network function, a physical server in the hypergraph H is constructed into a point, and the virtual network function is a hyperedge;

c. when the global visual field network system S is a super-dense cellular network optimization system, mobile users in the hypergraph H are constructed into points, and a base station is constructed into a super edge;

d. when the global visual field network system S is a cluster task scheduling optimization system, the request tasks in the hypergraph H are constructed into points, and the dependency relationship among the tasks is constructed into a hypergraph;

representing a hypergraph H of the global visual field computer network system by using an incidence matrix I, wherein the incidence matrix I is a V multiplied by E dimensional 0-1 matrix, when the position (V, E) in the incidence matrix I is 1, the incidence matrix I represents that a point V and a hyperedge E have a connection relation, and when the position (V, E) is 0, the incidence matrix I represents that the point V and the hyperedge E have no connection relation;

(3) and (3) performing feature calculation on the hypergraph H of the global-view computer network system in the step (2), wherein the feature calculation comprises the following steps:

(3-1) constructing an evaluation matrix W for representing the importance of each non-zero element in the incidence matrix I;

(3-2) calculating the performance loss of the evaluation matrix W:

judging the global configuration result set 0 in the step (1), and if the output result of the global configuration result in the step (1) is a discrete variable, calculating KL divergence D (W, I) of the incidence matrix I and the evaluation matrix W by using the following formula; if the output result of the global configuration result in the step (1) is a continuous variable, calculating the mean square error D (W, I) of the global configuration result correlation matrix I and the evaluation matrix W by using the following formula:

in the above formula, f (W) is the evaluation result of the evaluation matrix W by the evaluation function of the global view network system S to be interpreted, and f (I) is the evaluation result of the correlation matrix I of the global view network system S to be interpreted;

(3-3) calculating the conciseness of the evaluation matrix W in the step (3-1) by using the following formula, and defining the conciseness of the evaluation matrix W as:

in the above formula, W_evFor evaluating the elements of the matrix W at the (e, v) positions of the correlation matrix IThe prime value;

(3-4) representing the certainty of the evaluation matrix W by adopting the entropy H (W) of the evaluation matrix W:

(3-5) establishing an optimization model for solving the evaluation matrix W, wherein an objective function of the optimization model is as follows:

minD(W，I)+λ₁||W||+λ₂H(W)

the constraint conditions of the optimization model are as follows:

wherein D (W, I) is the mean square error in step (3-2), | | W | | is the simplicity in step (3-3), H (W) is the entropy of the evaluation matrix W in step (3-4), and λ₁And λ₂Respectively calculating parameters of simplicity and entropy;

(4) and (3) solving the optimization model in the step (3-5) by adopting a gradient descent method to obtain an evaluation matrix W, and obtaining the representation of the importance of each connection relation in the incidence matrix I according to the evaluation matrix W to realize the explanation of the global visual field network system based on deep learning.

Images