JP6721555B2 - Software development apparatus, software development method, and program - Google Patents

Description

Embodiments of the present invention relate to a software development device, a software development method, and a program.

Software optimization is generally carried out using multivariate analysis of input data, in which case the coefficient of determination that can explain the model of interest is around 50%. That is, it cannot be estimated once in two times. Also, satisfactory results are not obtained when the T test is used.

JP, 2016-167262, A Japanese Patent Laid-Open No. 01-173134

For example, in optimizing the software of a system mounted on an aircraft or the like, the values of various parameters such as CPU usage rate, memory usage rate, and net load are measured, and if a certain event occurs based on those measurement results. Perform a certain operation (for example, if the net load exceeds a threshold, take measures such as lowering the priority of the task of a certain program).

By the measures as described above, optimization may be temporarily achieved to satisfy the system requirements. However, the make-and-make method returns to its original, unoptimized state over time, making it difficult to meet system requirements.

As long as each measured value of various parameters is regarded as a mere discrete value, such a problem cannot be solved.

The problem to be solved by the invention is to provide a software development device, a software development method, and a program capable of realizing software in which various parameters are optimized.

The software development device of the embodiment is a software development device used for the development of software whose behavior is controlled by a parameter group, and calculates the optimum value of each parameter from the measured value of each parameter obtained when the software is executed. In this case, a neural network is used, and a superfunction is applied to the measured values of each parameter input to the neural network, and the superfunction is also applied to each node of the neural network. And software recreating means for recreating the software by using each parameter calculated by the value calculating means.

The figure which shows an example of a structure of the whole system containing the evolutionary type software development apparatus which concerns on one Embodiment. The figure which shows an example of the model of the target software. The figure which shows an example of an evaluation function. The figure which shows an example of an environment-friendly neural network. The flowchart which shows an example of operation|movement by the evolutionary type software development apparatus.

Embodiments will be described below with reference to the drawings.

The evolutionary software development apparatus of the embodiment described below acquires communication data, system lease information, etc. at the time of execution of target software, and machine learning is performed by a neural network so that the system has a target performance value. To find optimal values of various software control parameters (hereinafter, may be abbreviated as “parameter”) capable of controlling the software, and further reflect them in the software. Will enable the development of software that autonomously makes the best use of system resources and evolves to satisfy system requirements.

Further, the data obtained during execution of highly sophisticated and complicated software is, for example, 14.4 Tbytes when the operation time is 4 hours at 1 Gbyte/sec. , It is also possible to utilize such big data.

Therefore, the evolutionary software development apparatus of this embodiment handles the behavior of software in a quantum theory. That is, each measured value of the software control parameter is not treated as a mere discrete value, but is treated as a representative value of values that have been integrated while being influenced by other elements. Specifically, by extending the input values of multivariate analysis to hyperfunctions, it becomes easier to handle discontinuous data, and the parameters themselves extended to hyperfunctions can be applied to neurons (nodes) of the neural network. Apply. Since the superfunction itself can be differentiated freely, smooth operation on the neural network is guaranteed. Furthermore, it is possible to solve the constraint of the local optimum value by a genetic algorithm.

FIG. 1 is a diagram showing an example of a configuration of an entire system including an evolutionary software development device according to an embodiment.

As shown in FIG. 1, this system includes an evolutionary software development device 1 and recording media 2 and 3. The evolutionary software development device 1 includes a software acquisition data formulation unit 4, an optimum value calculation unit 5, and a parameter determination/software re-creation unit 6 as various functions.

The configuration of the evolutionary software development device 1 is not limited to the example of FIG. For example, the evolutionary software development device 1 may be configured to include the recording media 2 and 3.

The evolutionary software development device 1 is a device used for developing software whose behavior is controlled by various parameters. The various functions provided in the evolutionary software development apparatus 1 are realized in the form of programs executed by a computer, for example.

The recording medium 2 temporarily stores data (software acquisition data) obtained when the target software is executed. This data includes measured values of various parameters observed during execution of the software. The data stored in the recording medium 2 will be used by the evolutionary software development device 1.

The recording medium 3 is for temporarily storing the software that has been re-created after the value of each parameter is determined by the evolutionary software development device 1. The software stored in the recording medium 3 will be executed as needed.

The software acquisition data formulation unit 4 has a function of reading out the software acquisition data from the recording medium 2 and formulating the data so that the data form is suitable for the optimum value calculation unit 5 to process.

The optimum value calculation unit 5 has a function of calculating optimum values of various parameters of the target software using the data formulated by the software acquisition data formulation unit 4. The optimum value calculation unit 5 uses a neural network when calculating the optimum value of each parameter from the measured value of each parameter, and applies a superfunction to the measured value of each parameter input to the neural network, The superfunction is applied to each node of the neural network. The details will be described later.

The parameter determination/software re-creation unit 6 determines the values of various parameters of the software to be re-created based on the result calculated by the optimum value calculation unit 5, and reflects the determined various parameter values in the software. This is a function to recreate software.

<About parametric software model>
FIG. 2 is a diagram illustrating an example of a target software model.

Here, an example of a parametric software model that enables software control by parameters is shown.

As shown in FIG. 2, the target software includes mission software group (business logic) 11, integrated information service 12, and data centric framework 13.

The mission software group 11 includes various applications (applications corresponding to planning processing, integrated sensor control, sensor information correlation/class identification, target information integration, mission simulation support, human interface, etc.).

The integrated information service 12 includes a common service function (each function such as ASSW service, ASW service, communication service, CDS service) 21 having a resource management/secure management function, and an integrated information service interface 22.

The data-centric framework 13 includes an IO service function 31 that belongs to an operating system function, a platform-specific service function (functions such as device driver, GUI library, and inspection/monitoring firmware) 32, and a transmission service function (transmission gateway, DDS, CORBA). , WEB services, etc.) 33.

The individual applications that make up the mission software group 11 shown in the example of this model control the behavior of the software in order to control the behavior of the software, such as QOS (precision, degree of class identification, screen resolution), processing execution time, It has a structure in which CPU usage rate, memory usage, communication band, etc. are built in as parameters, and can be dynamically controlled to optimally utilize system resources and obtain maximum results.

Further, in the example of this model, the request from each application is interpreted by the integrated information service interface 22, and the operating system function having the transmission service function 33, the platform-specific service function 32, and the IO service function 31 is stored as data. The data-centric framework 13 is based on the Centric architecture.

In the range A in FIG. 2, each application realizes a use case service through the integrated information service interface 22. The values of the parameters that control the objects of these applications are determined by the evolutionary software development apparatus 1 performing machine learning using a neural network.

In the range B in FIG. 2, the request from each application interpreted by the integrated information service interface 22 is realized by processing such as arrangement/setting/management of the common service function 21. The resource management function provided in the common service function 21 is based on the optimization parameters determined by machine learning by the evolutionary software development apparatus 1 based on network control, CPU, memory usage allocation, priority management of threads, etc. It is monitored and reflected in the arrangement/setting of the common service function 21. The transmission gateway of the transmission service function 33 can connect a system having a legacy interface in the same manner as DDS.

In a range C in FIG. 2, the IO service function 31 and the platform-specific service function 32 provide services specific to the platform.

<About the quantum treatment of software>
Next, the quantum treatment of software will be described.

In the present embodiment, when the optimum value calculation unit 5 obtains the optimum value of each parameter, the measured value of each parameter is quantum Treat it properly.

Quantum treatment of software means that each parameter measured at time t is regarded as a representative value of the behavior from the time the software is started to the time t. Therefore, the input value is not a mere real number, but is, for example, a superfunction obtained by multiplying the measured value by a sudden decrease function and integrating the time.

Further, in the present embodiment, the optimum value calculation unit 5 uses the evaluation function in the optimum value calculation of each parameter. The evaluation function used here is also a super function, and the target space of the input value data is expanded.

An example of the evaluation function is shown in equation (1) of FIG.

The optimum value calculation unit 5 obtains an SP that satisfies the condition (that is, the condition that the differential value of the evaluation function is 0 and the value of the evaluation function is the minimum) shown in the formula (2) for the expression (1).

However, in the case of the expression (2′) in the expression (1), there is a possibility that the software design is broken, and therefore information indicating the necessity of redesign is output. Examples of the cause of software design failure include design mistakes, software bugs, and unauthorized access to the server.

LP represents the Lp norm. Here, it is assumed that calculation is performed using the L2 norm (p=2) or the like according to the sensitivity of the data. When the Fourier-transformed data is targeted, selection such as p=2 is performed.

f() represents the Schwarz superfunction. This superfunction can take as an argument the discrete value data observed by software (such as a multiple of Diracδ).

g() represents a neural network having an(t,gop) as a neuron (node). This will be described in detail later.

Round delta represents the derivative of the superfunction.

SO represents the system target value. Specifically, it is represented by SO(v1,v2,v3,...,vn).

SD represents the software acquisition value (parameter measurement value). Specifically, it is represented by SD(d1,d2,d3,...,dn).

SP represents a software control parameter. Specifically, it is represented by SP(a1(t,gop), a2(t,gop), a3(t,gop),..., An(t,gop)).

an(t,gop) represents the n-th neuron (node) of the hierarchy t.

t represents the measurement time.

gop represents a genetic operator (selection, crossover, mutation, apoptosis).

<About environment-friendly neural network>
In the present embodiment, the optimum value calculation unit 5 uses an environment-responsive neural network for the optimum value calculation of each parameter. The neuron group that calculates the input value changes according to the temporal change of the input value. For these neuron groups, for example, genetic operations (selection, crossover, mutation, apoptosis) may be applied according to the environment in which the system is placed to grow the neuron group for calculation. ..

An operation corresponding to apoptosis, which is not found in conventional genetic algorithms, can be realized by giving a numerical penalty to the output value of a node (neuron) according to the penalty method. For example, when the output value is represented by a fraction (the numerator is a fixed value and the denominator is a variable), it can be realized by an operation such as setting the value of the denominator to 100 million.

FIG. 4 shows an example of the environment-responsive neural network.

Here, each measured value measured by the system at time t is regarded as a representative value in the vicinity of time t, and the measured value is handled using a Schwarz's super function. That is, the state of the system is treated quantumally.

The anm(t) of each node (neuron) corresponds to the above-mentioned an(t). anm(t) represents a neuron (node) at the nth layer and the layer depth m. This anm(t) is described by, for example, the product of controllable software parameters and a superfunction.

The following are examples of each parameter.

・Number of threads ・Thread priority ・Thread heap amount ・Thread cycle ・Screen resolution QOS
・Data communication QOS
・Event interrupt QOS
・DatumQOS
The following are examples of candidates for the hyperfunction.

・Δ(t)
・Heaviside function ・Sigmoid function (eg "1/(1+exp(-xgn(an)))")
Note that gn corresponds to a genetic algorithm, and the fitness of an individual to be calculated is determined according to the state of an.

Examples of input values input by the neural network include measured values of various parameters as shown below. However, the respective measured values in this case are input in the form of a superfunction that is time-integrated using, for example, a sudden decrease function or a delta function.

D1: Planning measurement data (effect x11, processing time x12, CPU usage rate x13, memory usage rate x14)
D2: Integrated sensor control measurement data (accuracy x21, processing time x22, CPU usage rate x23, memory usage rate x24)
D3: Detection/tracking processing measurement data (accuracy x31, processing time x32, CPU usage rate x33, memory usage rate x34)
D4: Sensor data processing measurement data (identification accuracy x41, processing time x42, CPU usage rate x43, memory usage rate x44)
D5: Data communication processing measurement data (processing time x51, CPU usage rate x52, memory usage rate x53, communication bandwidth x54)
D6: Display processing measurement data (screen resolution x61, processing time x62, CPU usage rate x63, memory usage rate x64)
D7: Operation processing data (response time x71, processing time x72, CPU usage rate x73, memory usage rate x74)
In this case, the output values E1 to E7 output by the neural network correspond to the target values E1optm to E7optm of various parameters as shown below, for example.

E1optm: Design target value of planning process (effect, processing time, CPU usage rate, memory usage rate)
E2optm: Design target value for integrated sensor control processing (accuracy, processing time, CPU usage rate, memory usage rate)
E3optm: Design target value for detection/tracking processing (accuracy, processing time, CPU usage rate, memory usage rate)
E4optm: Design target value for sensor data processing (identification accuracy, processing time, CPU usage rate, memory usage rate)
E5optm: Design target value for data communication processing (identification accuracy, processing time, CPU usage rate, memory usage rate, communication bandwidth)
E6optm display processing design target values (screen resolution, processing time, CPU usage rate, memory usage rate, communication bandwidth)
E7optm operation processing design target values (response time, CPU usage rate, memory usage rate)
The optimum value calculation unit 5 calculates the Lp norm of the output values E1 to E7 and the respective target values E1optm to E7optm and feeds it back to the neural network.

In this way, by causing the neural network to perform machine learning so that the system has a target performance value, it is possible to obtain optimum values of various parameters capable of controlling software.

<Optimization of non-stationary events>
It is also possible to record the growth process in the above-mentioned neural network and construct a hierarchical neural network for reflecting it in the next step so as to be able to cope with time changes.

<Operation example>
Next, an example of the operation of the evolutionary software development device 1 will be described with reference to FIG.

First, the evolutionary software development apparatus 1 reads the software acquisition data from the recording medium 2 by the software acquisition data formulation unit 4 and converts the data into a data form suitable for the optimum value calculation unit 5 to process. It is formulated (step S1).

Next, the evolutionary software development device 1 calculates the optimum values of various parameters of the target software by using the data formulated by the software acquisition data formulation unit 4 by the optimal value calculation unit 5 (step S2). Here, in calculating the optimum value of each parameter from the measured value of each parameter, a neural network is used, and a superfunction is applied to the measured value of each parameter input to the neural network, and each node of the neural network is used. Then, after applying the super function, the optimum value of each parameter is calculated.

Finally, the evolutionary software development apparatus 1 determines the values of various parameters of the software to be re-created by the parameter determination/software re-creation unit 6 based on the result calculated by the optimum value calculation unit 5 The software is recreated by reflecting the parameter value in the software (step S3).

As described above, according to the evolutionary-type software development apparatus of the present embodiment, the behavior of software is treated in a quantum theory, and each measured value of the software control parameter is not treated as a mere discrete value. Treated as a representative value of the values that have been integrated while being affected, by extending the input value of multivariate analysis to a hyperfunction, making discontinuous data easier to handle, and extending to a hyperfunction. By applying itself to the neurons (nodes) of the neural network, smooth operation on the neural network is guaranteed by the superfunction that can be differentiated freely. Furthermore, the constraint on the local optimum value can be solved by a genetic algorithm.

In addition, by letting the neural network perform machine learning so that the system achieves the target performance value, the optimum values of various parameters can be found heuristically, and by reflecting them in software, autonomous It is possible to develop software that evolves to satisfy system requirements while making maximum use of system resources.

Further, although the amount of data obtained during execution of highly sophisticated and complicated software is enormous, the evolutionary software development apparatus of the present embodiment can also utilize such big data.

As described in detail above, according to the above embodiment, it is possible to realize software in which various parameters are optimized.

Although some embodiments of the present invention have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These novel embodiments can be implemented in various other forms, and various omissions, replacements, and changes can be made without departing from the spirit of the invention. These embodiments and their modifications are included in the scope and gist of the invention, and are also included in the invention described in the claims and the scope of equivalents thereof.

1... Evolutionary software development device, 2, 3... Recording medium, 4... Software acquisition data formulation unit, 5... Optimal value calculation unit, 6... Parameter determination/software re-creation unit, 11... Mission software group, 12... Integration Information service, 13... Data centric framework, 21... Common service function, 22... Integrated information service interface, 31,... IO service function, 32... Platform-specific service function, 33... Transmission service function.

Claims (10)

A software development device used for developing software whose behavior is controlled by a group of parameters,
When calculating the optimum value of each parameter from the measured value of each parameter obtained during execution of the software, a neural network is used, and a superfunction is applied to the measured value of each parameter input to the neural network, and the neural network Optimal value calculation means that applies a superfunction to each node of the network,
A software re-creating means for re-creating the software using each parameter calculated by the optimum value calculating means. The software development device according to claim 1, wherein the optimum value calculation means applies, as an input value of the neural network, a superfunction obtained by multiplying a measured value of each parameter by a sudden decrease function and time integration. 3. The software development device according to claim 1, wherein the optimum value calculation means applies a function represented by a product of each parameter and a superfunction to each node of the neural network. 4. The optimum value calculation means performs an operation corresponding to at least one of selection, crossover, mutation, and apoptosis, which are genetic operators, on any of the nodes. Software development equipment. The software development device according to claim 4, wherein the optimum value calculation means gives a numerical penalty to an output value of the node according to a penalty method when performing an operation corresponding to the apoptosis on any node. The software development device according to claim 1, wherein the optimum value calculation means calculates the optimum value of each parameter using an evaluation function including a super function. The software development device according to claim 6, wherein the evaluation function is represented by an Lp norm of a difference between an output value of the neural network and a target value. The software development device according to claim 6 or 7, wherein the optimum value calculation means obtains the value of each parameter when the differential value of the evaluation function is 0 and the value of the evaluation function is minimum. A software development method used to develop software whose behavior is controlled by a group of parameters,
When calculating the optimum value of each parameter from the measured value of each parameter obtained at the time of execution of the software by the optimum value calculation means, a neural network is used, and a superfunction is added to the measured value of each parameter input to the neural network. Software development including applying the superfunction to each node of the neural network and recreating the software by the software recreating means using each parameter calculated by the optimum value calculating means Method. A program used to develop software whose behavior is controlled by parameters,
When calculating the optimum value of each parameter from the measured value of each parameter obtained during execution of the software, a neural network is used, and a superfunction is applied to the measured value of each parameter input to the neural network, and the neural network also apply the generalized function to each node of the network, and the optimum value calculation function,
A software recreating function for recreating the software using each parameter calculated by the optimum value calculating function,
A program that makes a computer realize.