JPH0413343A - Study type link capacity control system - Google Patents

Abstract

PURPOSE:To determine the assignment of an optimum link capacity even to a call having any property by using a function study device so as to study the relation among a calling rate, a line speed and a call loss rate based on an observed value from a communication network actually in operation thereby deciding optimum link capacity. CONSTITUTION:An observed data corresponding to a respective link is written in an observed value storage area 11. A function study device 13 selects one observed value from observed values stored periodically in the observed value storage area 11 as a study data and gives a value representing a link load such as an observed calling rate and a link capacity as an input signal and a relevant observed call loss rate as a correct solution output signal to the function study device 13 to execute the study. A link capacity assignment evaluation circuit 12 uses the function study device 13 to evaluate the line capacity assignment. A link capacity optimizing circuit 14 decides an optimum link capacity by using the result of evaluation.

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention relates to a learning type link capacity control method that is incorporated and used inside a switching equipment or a communication network management center, and is particularly applicable to voice communication, image communication, data communication, etc. Communication in which multimedia communication with various speeds is mixed, the characteristics of each communication are unknown, communication in which the characteristics change during communication,
Furthermore, since the exchange uses learning-based call admission decision control, it can flexibly respond to situations where the number of calls that can be multiplexed on the same link changes over time, and can always distribute link capacity optimally. 5 C Conventional Technology 3 Concerning a Learning-type Link Capacity Control System Capable of Dynamic Link Capacity Control A communication network is composed of switching nodes and links connecting the switching nodes. In addition, switching equipment has a large number of links, and these links are realized by dividing physical lines with larger capacity using technologies such as time division multiplexing or label multiplexing (packet format multiplexing method). has been done. Therefore, in principle, the link capacity can be changed within the capacity of the physical line.

Here, we will briefly discuss the relationship between link capacity and communication quality such as call loss rate in a communication network.

A user of a communication network sends a call connection request to a switching node. The originating switching node receives the request, selects a destination switching node that accommodates the destination user, and a link between the originating node and the destination node,
The call is connected only when there is free capacity on that link to meet the request. In this case, the frequency of situations where the link capacity is full and calls cannot be connected is calculated as the call loss rate 1
This value is determined by the ring capacity (■), the call origination rate (calling rate) g that uses the link, the value of parameters representing call characteristics such as call holding time, and the distribution of these values. be done.

In addition, link capacity control refers to adjusting the allocation of link capacity within the physical line capacity according to the load and nature of communication using each link, thereby reducing the average call loss rate and Reduce the call loss rate,
This is a control to achieve a purpose such as improving the usage rate of the mesh body.

Conventionally, the number of calls per unit time for each link, that is, the call rate, the probabilistic distribution of the call rate, and the call holding time and its distribution are predicted or observed from the operating communication network. Based on these values, analysis and simulations are performed to estimate the call loss rate. Then, based on this estimate,
With the physical line capacity as a constraint, the optimal link capacity was searched and determined in order to achieve the above objectives.

On the other hand, with the spread of multimedia communications, communication networks that multiplex various calls onto the same line have been proposed.

For example, in the apparatus described in Japanese Patent Application No. 63-176242 and the drawings, a learning type call acceptance determination control is proposed in which the number of calls that can be multiplexed on a link is determined by learning. When this learning type call admission determination control is adopted, the number of calls that can be multiplexed on a fixed number of lines changes from time to time.

[Problem to be solved by the invention]

In the above-described conventional technology, sufficient consideration has not been given to always optimally distributing link capacity in a communication network that handles multimedia communications.

For example, in multimedia communications, communication speeds and hold times differ for each call, and some calls are being considered in which the communication speed changes during communication. However, it is not easy to observe and understand the details of the communication characteristics of each such call. Furthermore, in such communication networks, the capacity required by one call is not clear, so the maximum number of calls that can be multiplexed by a link with a certain capacity is not clear. Therefore,
With conventional methods based on analysis and simulation, it is extremely difficult to estimate the call loss probability, and it is also not easy to determine the optimal allocation of link capacity. Furthermore, because analysis and simulation require a very large amount of calculation, it is difficult to realize real-time control in situations where a new communication service is added or the nature of a call changes.

Furthermore, in a communication network that handles multimedia communications, when switching equipment adopts learning-based call acceptance decision control, link capacity is increased to keep up with the ever-changing number of calls that can be multiplexed on a fixed number of lines. It is difficult to make an optimal allocation of

SUMMARY OF THE INVENTION An object of the present invention is to provide a learning type link capacity control method that can improve the above-mentioned problems and always optimally distribute link capacity in a multimedia communication network.

[Means to solve the problem]

In order to achieve the above object, the learning type link capacity control method of the present invention provides a multi-input/output variable nonlinear functional algorithm and learns actual observation data obtained from the communication network in operation. Find a function for estimating call loss probability from observed values of load and link capacity, use this function to estimate actual call loss probability for link capacity allocation,
It is characterized by searching and determining link capacity allocation that optimizes the communication quality and utilization efficiency of the entire network while evaluating link capacity allocation.

Furthermore, the multi-input/output variable nonlinear function learning device uses a neural network realized by connecting neuronal circuits composed of a weighted sum of multiple input variables and a nonlinear output function in a network to estimate call loss probability. and performing link capacity allocation evaluation.

Note that a multi-input/output variable nonlinear function learner (hereinafter referred to as a function learner) has a large number of input signal lines, a large number of output signal lines, and a correct output signal line, and the relationship between the input signal and the output signal with respect to the input signal is It is possible to freely change the values of many registers inside the function learning device by combining them. Learning means that when you set an input value to the input signal line of this function learning device and at the same time set a correct output value to the correct output signal line, the function learning device internally allows you to derive the correct output value from the input value. There are many parameters to adjust. When this operation is performed on various "pairs of input values and correct output values" (hereinafter referred to as training data), the values of the internal parameters will eventually converge, and any input value will be used as the function learning data. When added to the device, the corresponding correct output value will be automatically output.

[Operations] In the present invention, the following three operations (i) to (ii)
Learning type link capacity control is performed by i).

(1) The link capacity control device periodically monitors the load (calling rate, hold time, etc.) and call loss rate applied to each link from the actually operating communication network, and calculates the link capacity and call loss rate at that time. Save in observation value storage area.

(ii) Use the observed values saved in (i) as learning data,
A function learning device is made to learn a function that estimates the call loss probability from the observed load on the link and the link capacity. Specifically, learning is performed by adding the observed load and link capacity to the function learning device as input signals, and at the same time adding the observed call loss rate corresponding thereto to the function learning device as a correct output signal.

(iii) When a certain capacity is allocated to each link for the observed load, the value of the call loss rate and usage rate of each link is (u)
It can be easily obtained by using a function learning device trained in . Therefore, the maximum call loss rate, link usage rate, etc. are evaluated for various link capacity allocations, and the link capacity allocation most suitable for the purpose is determined. The link capacity allocation determined in this way is notified to the exchange or the device that manages link capacity.

Note that these operations can be executed independently, and are actually executed in parallel.

By using a function learning device in this way, it is possible to constantly learn observed values and estimate the actual call loss rate, so it is possible to estimate the actual call loss rate even if the nature of the call changes or the call acceptance decision control changes during control. Even in such a situation, it is possible to accurately evaluate link capacity allocation. Therefore, by searching for optimal link capacity allocation based on this evaluation, it is possible to realize optimal link capacity control in real time in accordance with the actual network situation.

〔Example〕

An embodiment of the present invention will be described below with reference to the drawings.

(First Example) In this example, a case will be described in which N links are accommodated in M physical lines. Note that, for simplicity, the call rate is used as the value representing the load on the link.

FIG. 2 is a configuration diagram of a communication network in the first embodiment of the present invention.

In FIG. 2, 21 is the physical line, 22 is the physical line 2
1 into multiple links 25,
23 is a switching node; 24 is a network control device that receives and receives observed values for each ring 25 from the switching node 23, determines the optimal link capacity, and returns link capacity allocation to control the cross-connect and the switch; It's a ring.

This embodiment is a case where N=6 and M=3, and the physical line 21 is distributed as a plurality of virtual lines (links 25),
The exchanges are interconnected by their links 25.

The configuration of this link capacity control device 24 is shown in FIG.

FIG. 1 is a block diagram of a link capacity control device in a first embodiment of the present invention, FIG. 3 is an explanatory diagram showing a learning method of a function learning device in the first embodiment of the present invention, and FIG. FIG. 3 is an explanatory diagram showing variable conversion during call loss probability learning/estimation in the first embodiment of the present invention.

In FIG. 1, 11 is an observed value storage area for storing observed values obtained from the exchange node 23, and 12 is a function learning device 13.
13 is a function learning device, and 14 is a link capacity optimization circuit.

Next, each operation will be described.

Observed value data corresponding to each link is written in the observed value storage area 11. This observed value data is, for example, a call rate, a call loss rate, and a link capacity, and is periodically collected from the switching node 23. If there is no free space in the storage area 11, you can erase the oldest observed value in the storage area and write over it, or erase the observed value randomly selected from among the observed values in the storage area. A method is used in which the data is written on the data.

Further, the function learning device 13 included in the link capacity allocation evaluation circuit 12 periodically performs learning. As shown in Figure 3, one observation value is selected as learning data from among the observation values stored in the observation value storage area ll, and a value representing the link load and link capacity for equality of observation calls is calculated. The input signal and the corresponding observed call loss rate are added to the function learning device 13 as correct output signals to execute learning. As a method for selecting learning data from the observed value storage area 11, a method of selecting at random or a method of selecting in the order stored in the storage area is used.

In learning the call loss rate estimation function, learning is performed using small values such as the call loss rate and large values such as the number of lines, but if the input value and correct output value to the function learning device 13 is abnormally small. (too large), the time required for learning to converge may become longer, or learning may not converge.

Therefore, a method is used to improve the learning accuracy of the function by multiplying the link capacity or the observed call rate by a constant depending on the characteristics of the function learning device 13. At this time, as shown in FIG. 4(a), if the observed call loss rate is given as the correct output, it is multiplied by a constant, and as shown in FIG. 4(b), the function learning device
When using the output value of 3 as the call loss probability estimate, conversely, by dividing the output value of the function learning device 13 by a constant,
Convert to correct call loss probability estimate.

In addition, in F-online learning, which performs learning in parallel with control execution, observed values are accumulated over time and the accuracy of the call loss probability estimation function is improved through learning, but immediately after the start of control, there are no observed values. , control may become unstable.

Therefore, before using the function learning device I3 for control, a method is used in which the function learning device I3 is trained in advance, or a method in which learning data is written in the observed value storage area in advance.

That is, a rough relationship between the calling rate, link capacity, and call loss rate is determined by simple analysis, and the function learning device 13 is trained before starting actual control, or the approximate value is stored in the observed value storage area 11 in advance. Write it down. This makes it possible to perform stable control immediately after starting control when there are few observed values.
This is gradually corrected by learning during control, and adjusted to be accurate according to the actual call interval distribution, holding time, and call acceptance determination control.

In this case, for example, if the function learning device 13 is implemented using parallel hardware, it becomes possible to calculate the call loss probability estimate and the link capacity allocation evaluation function at high speed.

FIG. 5 is a configuration diagram of a function learning device in the first embodiment of the present invention.

In FIG. 5, (a) shows a neural network forming the function learning device 13, and (b) shows one of the neuron circuits forming the neural network 51. In FIG. Further, 51 is a neural network, 52 is a neural circuit, and 53 is a weight register.

In this embodiment, the function learning device 1 uses the neural network 51.
3. That is, as shown in (a), the neuron circuits 52 are connected in multiple stages to realize a multiple input/output nonlinear function.

Furthermore, as shown in (b), one neuron circuit 53 is composed of a weighted sum of multiple input signals and a nonlinear output function, and the weight for each input signal is stored in the weight register 53. There is.

This output function is a bounded continuous monotonically increasing function.

Therefore, determining the input-output relationship of the entire neural network is
This is a combination of values of the weight register 53. In addition,
As a method for correcting the value of the weight register 53 during learning, for example, the sum of squares (square error) of the difference between the output value of the neural network 51 and the correct output value for the input is expressed as a function using each weight register as a variable. The steepest descent method is known, which calculates the gradient end of this squared error and uses this value to modify the weight.

Next, a method for evaluating link capacity allocation using the function learning device 13 in the link capacity allocation evaluation circuit 12 will be described.

FIG. 6 is an explanatory diagram of ring capacity allocation evaluation in the first embodiment of the present invention, and FIG. 7 is an explanatory diagram of constraint condition check in the first embodiment of the present invention.

The evaluation of link capacity allocation varies depending on the purpose of control. For example, the objective is to minimize the average call loss rate.
Possible methods include maximizing the link usage rate and minimizing the worst-case call loss rate.

In this embodiment, a case will be considered in which the objective of minimizing the worst-case call loss rate of links within the network is selected and the link capacity is evaluated in accordance with this objective.

The evaluation open number E for link capacity allocation is E (v)=max (1,,1,,-,IN)
It is expressed as (1). When selecting the objective of minimizing the average loss rate, the evaluation function E is expressed as E(v)=Σ1□/N (1)'. Here, ■ is a vector (Vl,..., VN) whose elements are capacity allocations for each of N links, and ■ represents the call loss probability for link 1 (i = 1..., N). There is. Note that the call loss rate is a value determined from the link capacity and the load applied to the link.

In the second section, as an example, only the call rate is considered as a parameter representing the load applied to the link, and the call loss rate 11 is expressed by the following nonlinear function estimated from the link capacity Vl and the call rate gi.

1t=F(vi+gt) (2) This function F is learned by the function learning device 13 described above. Therefore, if the call loss rate for each link is estimated using a large number of function learning devices 13 and the maximum value is calculated,
The value of evaluation function E is determined.

For example, the configuration of a circuit that evaluates the maximum call loss probability from the link capacity allocation Vi and the observed call rate g is shown in FIG. In addition, in FIG. 6, one function learning device 1 is used for each link.
3 is used, however, the number of actually required function learning devices 13 can be reduced by using a method of sharing the function learning devices 13 among several links and performing time-division integration.

Furthermore, the link capacity allocation evaluation circuit 12 also includes a function that determines whether the link capacity does not exceed the physical line capacity.

The condition that restricts the maximum value of the sum of ring capacities based on physical line capacities is expressed as ΣIV, -cm≦O(3). Here, C (i = 1, ・
M) is the physical line capacity, and Σ' is C in the link capacity VJ.
This is an operator that calculates the sum only for links passing through 4.

Here, we define this condition as Heaviside.
) Consider an evaluation function EC of the following formula using function H.

Note that the Heaviside function is defined by the following formula.

The link capacity allocation in which the value of the evaluation function EC is O satisfies the constraint regarding the physical line capacity.

A circuit representing this evaluation function is shown in FIG.

In FIG. 7, (a) shows the configuration of a constraint condition check circuit for preventing link capacity from exceeding physical line capacity. 5(b) is the Heaviside function H of FIG. 5(a), and is an implementation example in which the Heaviside function is used as the output function in the neuron circuit 52 of FIG. 5(b).

Note that 71 is a weight register.

Finally, the method for determining the optimal link capacity using the link capacity optimization circuit [4] will be described.

FIG. 8 is a flowchart explaining the optimal link capacity allocation method in the first embodiment of the present invention, FIG. 9 is an explanatory diagram of link capacity evaluation in the first embodiment of the present invention, and FIG.
FIG. 0 is a configuration diagram of a link capacity optimization circuit in a first embodiment of the present invention.

Link capacity allocation control is to find the link capacity allocation that minimizes the evaluation function (1) among the link capacity allocations that satisfy the above constraint condition (3).

In this embodiment, as a method for realizing the search, a method will be described in which a large number of test link capacities are evaluated using the above-mentioned function, and the link capacity allocation with the best evaluation among them is determined. As this method, two methods (a) and (b) as shown in FIG. 8 can be considered.

In method (a), the link capacity generator generates a test link capacity allocation (8001), and the constraint condition is checked using the constraint evaluation function EC (8002).
. As a result, if it is determined that the constraint condition is not satisfied, another test link capacity is newly generated. Furthermore, if a link capacity that satisfies the constraint condition is found, then an evaluation E of the link capacity according to the purpose is calculated (8003). If this evaluation result is better than the already discovered link capacity evaluation, this link capacity allocation and its evaluation value are saved (8004).

In the method (b), after generating a link capacity allocation without considering constraints (8011), a new evaluation function is generated by a weighted sum of the evaluation function of the constraints and the evaluation function for the purpose as shown in FIG. E′ is used to search for the optimal ring capacity. In addition, in FIG. 9, 91 is a function learning device configured as a call loss probability estimation function.

Since this evaluation function E is defined as follows, it is possible to evaluate the constraint conditions and the purpose at once.

E' (v)=E(v) 10aEc (6) Here α
is a sufficiently large positive constant, and in order to minimize E', the second term EC on the right side must be O. That is, minimizing E' (Vi) is equivalent to minimizing the evaluation (1) for the objective within the range that satisfies the constraint (3).

Using such an evaluation function E', the test link capacity is evaluated in the same manner as method (a) in FIG. 8 (8012
), and stores the link capacity allocation with the best evaluation among them and its evaluation value (8013).

Link capacity optimization circuit 14 that implements the method (b)
The configuration is shown in FIG.

In FIG. 10, (a) is the link capacity optimization circuit 14.
(b) shows the configuration of a link l optimization circuit constituting the link capacity optimization circuit 14. Further, 101 is a link lR1 optimization circuit, 102 is a random number generation circuit, 103 is an optimum link capacity storage register, 104 is an optimum evaluation storage register, 105 is a switch, 106 is a comparator, and 107 is a test link capacity storage register. Further, vII vII...+''N is the test link capacity.

The link capacity optimization circuit 14 of this embodiment has both the above-mentioned link capacity allocation generation function and optimal link capacity storage function.

As shown in (a), the test link capacity (v+r ■l
+...l vN) is input to the link capacity allocation evaluation circuit 12, and the evaluation value is returned. This evaluation value is compared with the optimal evaluation value already found, and if the evaluation value is better, this test link capacity is saved. Repeat this operation and
Output the final saved optimal link capacity allocation.

Further, the functions shown in (a) are divided and arranged as a link 1 optimization circuit 101 for each link.

That is, as shown in (b), the optimal ring capacity storage register 103 and the optimal evaluation storage register 104 respectively store the optimal link capacity discovered at that time and the evaluation value for the link capacity. Note that an appropriate initial value is set.

In addition, as a method of generating the test link capacity, the content of the optimum link capacity storage register 103 is
The method of adding the random numbers generated in step 2 is used. As the random number distribution, −-like distribution, normal distribution, etc. can be used.

The link capacity thus generated is output to the outside via the test link capacity storage register 107.

Next, the external link capacity allocation evaluation circuit 12 returns an evaluation of the output test link capacity allocation. This value is compared with the contents of the optimal evaluation storage register 104,
If a better evaluation than the value in the optimum evaluation storage register 104 is obtained, the switch 105 is closed and the value of the test link capacity storage register 107 is transferred to the optimum link capacity storage register 103, and the value obtained from the outside is transferred to the optimum link capacity storage register 103. The evaluation values are each stored in the optimal evaluation storage register 104.

Note that the optimum evaluation storage register 104 and the comparator 106 may be provided in common, instead of for each link.

Through such operations, the optimum link capacity allocation is finally determined by repeating the evaluation of a large number of test link capacities and saving the link capacity that always obtained the optimum evaluation.

(Second Embodiment) In the first embodiment, a case was described in which all calls were handled without classifying them according to their characteristics, but in this embodiment, calls that are multiplexed on one link are We will discuss the case where more accurate control is achieved by classifying and handling based on communication speed and communication service used. Note that the classification of calls with similar characteristics is called a call type.

FIG. 11 is an explanatory diagram showing how to use the function learning device when handling calls classified into call types in the second embodiment of the present invention.

In FIG. 11, (a) shows a method for estimating the average call loss rate for all call types, and (b) shows a method for estimating the call loss rate for each call type.

The configuration of the communication network of this embodiment is the same as that of the first embodiment, and the link capacity control device is composed of an observation value storage area, a link capacity allocation evaluation circuit including a function learning device, and a link capacity optimization circuit. .

In addition, the link capacity control device observes the calling rate and call loss rate for each call type, stores it in a pattern table (not shown), and uses a function learner to calculate the average calling rate for all call types and the call loss rate for each call type. learn a function that estimates the call loss probability.

As a result, it is possible to easily create a link capacity allocation evaluation circuit that is conscious of call types and minimizes the worst-case call loss probability for all call types and all links. Further, by using such a link capacity allocation evaluation circuit in combination with the above-mentioned link capacity optimization circuit, it is possible to realize control that minimizes the worst call loss rate for all call types and all links.

[Effects of the Invention] According to the present embodiment, a function learning device is used to learn the relationship between the calling rate and line speed based on observed values from an actually operating communication network, and the call loss rate. Since this is used to determine the optimal link capacity allocation, it is possible to implement link capacity control that determines the optimal link capacity allocation for calls of any nature.

Furthermore, even if the number of calls that can be multiplexed for the same line capacity changes by changing the call admission determination control method, it is possible to similarly achieve optimal link capacity allocation.

Furthermore, even when calls are classified into multiple call types, highly accurate link capacity control can be easily achieved.

[Brief explanation of drawings]

FIG. 1 is a block diagram of a link capacity control device in a first embodiment of the present invention, FIG. 2 is a block diagram of a communication network in the first embodiment of the present invention, and FIG. 3 is a block diagram of a communication network in a first embodiment of the present invention. FIG. 4 is an explanatory diagram showing the learning method of the function learning device in the embodiment. FIG. 4 is an explanatory diagram showing variable conversion during call loss probability learning/estimation in the first embodiment of the present invention. FIG. A configuration diagram of the function learning device in the embodiment, FIG. 6 is an explanatory diagram of link capacity allocation evaluation in the first embodiment of the present invention, and FIG. 7 is an explanatory diagram of constraint condition check in the first embodiment of the present invention. , 8th
The figure is a flowchart explaining the optimal link capacity allocation method in the first embodiment of the present invention, and FIG.
FIG. 10 is a configuration diagram of a link capacity optimization circuit in the first embodiment of the present invention, and FIG. 11 is a classification diagram of call types in the second embodiment of the present invention. FIG. 2 is an explanatory diagram showing how to use the function learning device when handling calls made by the user. 11: Observation value storage area, 12: Link capacity allocation evaluation circuit, 13: Function learning device, 14: Link capacity optimization circuit, 2
1: Physical line, 22: Cross connect, 23: Switching node, 24: Link control device. 25: Link, 51: Neural network, 52: Neuron circuit, 53: Weight register, 71: Weight register, 101:
Link l optimization circuit, 102: Random number generation circuit, 103:
Optimal link capacity storage register. 1o4: Optimal evaluation storage register, 105: Switch, 1
06: Comparator, 107: Test link capacity storage register +
C1' Physical line capacity, E, E'': Evaluation against link capacity allocation, Ec: Evaluation against constraint conditions, H: Heaviside function 2g□: Observed call rate. 1□: Estimated call loss rate 1M: Maximum value function +"i 'Link capacity allocation, v,, v□,...+VN' Test link capacity. α: constant, +: linear sum. Figure: Evaluation diagram for link capacity allocation (Part 2) Figure (Part 1): Physical line capacity: Link capacity allocation: Evaluation diagram for constraint conditions (Part 1) Optimal link allocation diagram (Part 2) Optimal link Allocation diagram: Physical line capacity: Link capacity allocation: Observed call rate: Estimated call loss rate: Evaluation of link capacity allocation S - Haruko Nawate

Claims (2)

[Claims] (1) Observe the traffic volume for each link connecting the exchanges, and based on the observed volume, adjust and determine the allocation ratio of each link capacity within a range that does not exceed the physical line capacity accommodating the link. In the link capacity control method, which has a large number of input signal lines, output signal lines, and correct output signal lines,
A function learning device is provided that can change the relationship between an input signal and an output signal with respect to the input signal by a combination of internal register values, and the function learning device is used to learn observed values obtained from a communication network in operation. By doing so, a function is generated to estimate the call loss probability from the observed value of the load applied to the link and the link capacity, and the optimal link capacity allocation is searched for while evaluating the link capacity allocation based on the output value of the function,
A learning-type link capacity control method characterized by determining. (2) The above-mentioned function learning device uses a neural network realized by connecting neuronal circuits composed of a weighted sum of multiple input variables and a nonlinear output function in a network to estimate call loss probability and allocate link capacity. The learning link capacity control method according to claim 1, characterized in that an evaluation is performed.