Training device, control method, and program

Abstract

The training device (2000) acquires a first transmission symbol sequence and a second transmission symbol sequence. The first transmission symbol sequence is input to an optical transmission unit (112) and converted into an optical transmission signal. The second transmission symbol sequence is obtained by demodulating the optical transmission signal. The training device (2000) trains a set of a generator (200) and a discriminator (300) using a training data set (10) generated based on the first transmission symbol sequence and the second transmission symbol sequence. The generator (200) is trained to generate data that is determined to be ground truth data by the discriminator (300). The discriminator (300) is trained to distinguish between the ground truth data and data generated by the generator (200). The training device (2000) outputs parameter information (20) including trainable parameters of the generator (200).

Description

[Technical field]

[0001] The present disclosure relates generally to signal processing in optical communication systems. [Background technology]

[0002] Optical communication systems transmit messages from a transmitter to a receiver over an optical fiber medium. At the transmitter, a digital signal is converted into an optical signal. Because this process is not ideal, unintended distortions are introduced into the optical signal.

[0003] Distortions in optical signals can cause erroneous conversion of messages at the receiver. Therefore, a method to mitigate this distortion is necessary. For example, Non-Patent Document 1 discloses a technique for performing calibration of a pre-compensator using a neural network. [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] G. Paryanti, H. Faig, L. Rokach and D. Sadot, “A Direct Learning Approach for Neural Network Based Pre-Distortion for Coherent Nonlinear Optical Transmitter,” IEEE, Journal of Lightwave Technology, vol. 38, no. 15, pp. 3883-3896, August 1, 2020. Summary of the Invention [Problem to be solved by the invention]

[0005] The inventors of the present disclosure have found that there is room for improving the accuracy of compensation for distortion caused by a transmitter on an optical signal. An object of the present disclosure is to provide a technique for improving the accuracy of compensation for distortion caused by a transmitter on an optical signal. [Means for solving the problem]

[0006] The training device of the present disclosure includes at least one processor and a storage element storing instructions, and by executing the instructions, the at least one processor acquires a first transmission symbol sequence, which is a sequence of symbols to be input to an optical transmission unit included in a transmitter, the optical transmission unit acquires the first transmission symbol sequence, converts the first transmission symbol sequence into an optical transmission signal, transmits the optical transmission signal to a receiver, acquires a second transmission symbol sequence, which is a sequence of symbols obtained by demodulating the optical transmission signal, generates a training data set including input data and ground truth data based on the first transmission symbol sequence and the second transmission symbol sequence, and performs training of a set of a generator and a classifier using the training data set, the generator is trained to generate data to be determined as the ground truth data by the classifier, the classifier is trained to distinguish between the ground truth data and the data generated by the generator, and outputs parameter information including trainable parameters of the generator.

[0007] A control method according to the present disclosure is executed by a computer and includes the steps of: acquiring a first transmission symbol sequence, which is a sequence of symbols to be input to an optical transmission unit included in a transmitter; the optical transmission unit acquiring the first transmission symbol sequence, converting the first transmission symbol sequence into an optical transmission signal, transmitting the optical transmission signal to a receiver, acquiring a second transmission symbol sequence, which is a sequence of symbols obtained by demodulating the optical transmission signal; generating a training data set including input data and ground truth data based on the first transmission symbol sequence and the second transmission symbol sequence; training a set of a generator and a discriminator using the training data set; the generator being trained to generate data to be determined as the ground truth data by the discriminator; the discriminator being trained to distinguish between the ground truth data and the data generated by the generator; and outputting parameter information including trainable parameters of the generator.

[0008] A computer-readable medium of the present disclosure stores a program that causes a computer to execute the following steps: acquire a first transmission symbol sequence, which is a sequence of symbols to be input to an optical transmitting unit included in a transmitter; the optical transmitting unit acquires the first transmission symbol sequence, converts the first transmission symbol sequence into an optical transmission signal, transmits the optical transmission signal to a receiver, and acquires a second transmission symbol sequence, which is a sequence of symbols obtained by demodulating the optical transmission signal; generates a training dataset including input data and ground truth data based on the first transmission symbol sequence and the second transmission symbol sequence; train a set of a generator and a discriminator using the training dataset; the generator is trained to generate data to be determined as the ground truth data by the discriminator; the discriminator is trained to distinguish between the ground truth data and the data generated by the generator; and output parameter information including trainable parameters of the generator. Effect of the Invention

[0009] The present disclosure provides techniques for improving the accuracy of compensation for distortions introduced by a transmitter into an optical signal. [Brief description of the drawings]

[0010] [Figure 1] FIG. 1 is a diagram showing an optical communication system for which parameter information is generated by the training device of the first embodiment. [Diagram 2] FIG. 2 is a diagram showing an overview of the training device of the first embodiment. [Diagram 3] FIG. 3 is a block diagram illustrating an example of a functional configuration of the training device. [Figure 4] FIG. 4 is a block diagram showing an example of a hardware configuration of a computer that realizes the training device. [Diagram 5] FIG. 5 is a flowchart illustrating an example of a process executed by the training device 2000. [Figure 6] FIG. 6 shows an example configuration of a generator that includes a set of layers with linear activations and a set of layers with non-linear activations. [Figure 7] Figure 7 shows an example of the training flow for a set of models. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0011] Hereinafter, an example of an embodiment according to the present disclosure will be described with reference to the drawings. Note that the same elements in the drawings are given the same reference numerals, and redundant description will be omitted as necessary.

[0012] EMBODIMENT 1 <Summary> 1 is a diagram showing an optical communication system for which parameter information is generated by the training device of embodiment 1. The optical communication system 100 includes a transmitter 110, a receiver 120, and an optical fiber 130. The transmitter 110 includes an optical transmitting unit 112 that acquires a transmission symbol sequence, converts the transmission symbol sequence into an optical transmission signal, and transmits the optical transmission signal to the receiver 120 via the optical fiber 130. The optical transmitting unit 112 can constitute a Mach-Zehnder modulator, a driving amplifier, or other signal processing and conversion devices.

[0013] The transmitted symbol sequence is a digital signal representing a time sequence of symbols. The transmitted symbol sequence is the result of encoding data to be transmitted to the receiver 120. If the optical communication system 100 is a single-polarized, single-channel optical communication system, each symbol includes an in-phase component and a quadrature-phase component. Thus, in this case, the transmitted symbol sequence includes a set of in-phase and quadrature-phase components at each time point.

[0014] The receiver 120 includes an optical receiver 122 that receives an optical reception signal from an optical fiber 130 and converts the optical reception signal into a reception symbol sequence. The reception symbol sequence is a digital signal representing a time sequence of symbols representing data to be transmitted by the transmitter 110.

[0015] In addition to the above components, the optical communication system 100 includes a distortion compensator 140 that compensates for distortion introduced by the optical transmitting unit 112 into the optical transmission signal. The distortion compensator 140 may be included in the transmitter 110 or the receiver 120. In the former case, the distortion compensator 140 is disposed before the optical transmitting unit 112 and operates as a pre-compensator. The pre-compensator obtains a transmission symbol sequence, imparts pre-distortion to the transmission symbol sequence, and passes the transmission symbol sequence to the optical transmitting unit 112.

[0016] In the latter case, the distortion compensator 140 is disposed after the optical receiver 122 and operates as a post-compensator. The post-compensator receives a received symbol sequence from the optical receiver 122 and applies post-distortion to the received symbol sequence.

[0017] Fig. 2 is a diagram showing an overview of the training device 2000 of embodiment 1. Note that the overview shown in Fig. 2 shows an example of the operation of the training device 2000 to make the training device 2000 easy to understand, and is not intended to limit or narrow the operating range of the training device 2000.

[0018] For calibration of the distortion compensator 140 (in other words, for generating appropriate parameters of the distortion compensator 140 to be used for performing pre-distortion or post-distortion by the distortion compensator 140), the training device 2000 uses a set of machine learning-based models called a model set 400. The model set 400 includes a generator 200 and a classifier 300, and may constitute a Generative Adversarial Network (GAN).

[0019] The training device 2000 trains the model set 400 using a training dataset 10 including input data 12 and ground truth data 14. The training dataset 10 is generated such that conversion of the input data 12 to the ground truth data 14 is substantially equivalent to distortion compensation of the optical transmission signal. To this end, the training dataset 10 is generated using a set of a first transmission symbol sequence and a second transmission symbol sequence. The first transmission symbol sequence is a symbol sequence input to the optical transmitting unit 112. On the other hand, the second transmission symbol sequence is a symbol sequence obtained by demodulating the optical transmission signal output from the optical transmitting unit 112 in response to the first transmission symbol sequence being input to the optical transmitting unit 112. This demodulation of the optical transmission signal is performed by demodulating the training data 10. set Intermediate number provided to generate 10 light This can be performed by the receiving unit 50.

[0020] The training device 2000 performs GAN-type training on the model set 400. Specifically, the generator 200 is configured to acquire input data 12, and is trained to generate data sufficiently similar to the ground truth data 14. Meanwhile, the discriminator 300 is configured to acquire the ground truth data 14 or data generated by the generator 200, and is trained to distinguish between these data. Since the training data set 10 is generated such that the conversion from the input data 12 to the ground truth data 14 is substantially equivalent to the distortion compensation of the optical transmission signal, the above training allows the generator 200 to perform distortion compensation of the optical transmission signal (i.e., pre-distortion of the transmission symbol sequence or post-distortion of the reception symbol sequence).

[0021] Once the training of the model set 400 is completed, the training device 2000 outputs parameter information 20 including at least the trainable parameters of the generator 200. The distortion compensator 140 may be configured with the trainable parameters of the generator 200. This enables the distortion compensator 140 to compensate for distortion introduced by the optical transmitter 112 into the optical transmission signal.

[0022] <Examples of effects> According to the training device 2000, the training data set 10, a pair of a generator 200 and a discriminator 300 constituting a GAN is trained to determine parameters of the distortion compensator 140. Specifically, the generator 200 is configured to acquire input data 12, and is trained to generate data sufficiently similar to the ground truth data 14. Meanwhile, the discriminator 300 is configured to acquire the ground truth data 14 or data generated by the generator 200, and is trained to distinguish between these data. By determining parameters of the distortion compensator 140 through this GAN-type training, the distortion compensator 140 can accurately compensate for the distortion introduced into the optical transmission signal by the optical transmission unit 112.

[0023] A more detailed description of training device 2000 is provided below.

[0024] <Example of functional configuration> 3 illustrates a functional configuration of the training device 2000. The training device 2000 includes an acquisition unit 2020, a generation unit 2040, a training execution unit 2060, and an output unit 2080. The acquisition unit 2020 acquires a plurality of sets of a first transmission symbol sequence and a second transmission symbol sequence. The generation unit 2040 generates a training data set 10 for each set of the first transmission symbol sequence and the second transmission symbol sequence. The training execution unit 2060 executes training of the model set 400 using the training data set 10. The output unit 2080 outputs parameter information 20.

[0025] <Example of hardware configuration> Training device 2000 may be realized by one or more computers, each of which may be a dedicated computer created for implementing training device 2000, or a general-purpose computer such as a personal computer (PC), a server machine, or a mobile device.

[0026] Training device 2000 can be realized by installing an application on a computer. The application is realized by a program that causes the computer to function as training device 2000. In other words, the program implements the functional components of training device 2000.

[0027] 4 is a block diagram showing an example of a hardware configuration of a computer 1000 that realizes a training device 2000. In FIG. 4, the computer 1000 includes a bus 1020, a processor 1040, a memory 1060, a storage device 1080, an input / output interface 1100, and a network interface 1120.

[0028] The bus 1020 is a data communication path for the processor 1040, the memory 1060, the storage device 1080, the input / output interface 1100, and the network interface 1120 to transmit and receive data to and from each other. The processor 1040 is a processor such as a central processing unit (CPU), a graphics processing unit (GPU), or a field-programmable gate array (FPGA). The memory 1060 is a main storage element such as a random access memory (RAM) or a read only memory (ROM). The storage device 1080 is an auxiliary storage element such as a hard disk, a solid state drive (SSD), or a memory card. The input / output interface 1100 is an interface between the computer 1000 and a peripheral device (such as a keyboard, a mouse, or a display device). The network interface 1120 is an interface between the computer 1000 and a network. The network may be a local area network (LAN) or a wide area network (WAN).

[0029] The storage device 1080 may store the programs described above. Processor 1040 executes the program to realize each functional component of the training device 2000.

[0030] The computer 1000 is not limited to the configuration shown in Fig. 4. For example, as described above, the training device 2000 may be realized by multiple computers. In this case, the computers may be connected to each other via a network.

[0031] <Processing flow> 5 is a flowchart showing an example of processing executed by the training device 2000. The acquisition unit 2020 acquires a plurality of sets of a first transmission symbol sequence and a second transmission symbol sequence (S102). The generation unit 2040 generates a training data set for each set of the first transmission symbol sequence and the second transmission symbol sequence (S104). The training execution unit 2060 executes training of the model set 400 (S106). The output unit 2080 outputs the parameter information 20 (S108).

[0032] <Acquisition of the first transmission symbol sequence and the second transmission symbol sequence: S102> The acquiring unit 2020 acquires a plurality of sets of the first transmission symbol sequence and the second transmission symbol sequence. The first transmission symbol sequence may be any digital signal that can be input to the optical transmitting unit 112. The second transmission symbol sequence may be obtained by 1) inputting the first transmission symbol sequence to the optical transmitting unit 112, and 2) converting the optical transmission signal output from the optical transmitting unit 112 into a digital signal. The digital signal obtained by this conversion may be used as the second transmission symbol sequence. As shown in FIG. 2, the conversion of the optical transmission signal into the second transmission symbol sequence may be performed by the intermediate optical receiving unit 50.

[0033] The multiple sets of the first transmission symbol sequence and the second transmission symbol sequence may be stored in advance in a storage device accessible from the training device 2000. In this case, the acquiring unit 2020 may acquire the multiple sets of the first transmission symbol sequence and the second transmission symbol sequence from the storage device. Furthermore, the training device 2000 may receive the multiple sets of the first transmission symbol sequence and the second transmission symbol sequence transmitted from an arbitrary computer.

[0034] <Generating training dataset: S104> Generation section 20 4 0 generates a training data set 10 for each set of the first transmission symbol sequence and the second transmission symbol sequence (S104). As described above, the training data set 10 is generated such that the conversion from the input data 12 to the ground truth data 14 is substantially equivalent to the distortion compensation of the optical transmission signal.

[0035] A first example of such a training data set 10 includes a first transmission symbol sequence as input data 12 and an inverse signal of the distortion included in a second transmission symbol sequence as ground truth data 14. Using this type of training data set 10, the generator 200 is trained to generate from the first transmission symbol sequence an inverse signal of the distortion included in the second transmission symbol sequence. The inverse signal of the distortion can be applied to the first transmission symbol sequence to counteract the distortion introduced by the optical transmitter 112. Thus, the trainable parameters of the generator 200 trained with the training data set 10 of the first example can be applied to the distortion compensator 140 acting as a pre-compensator.

[0036] The generating unit 2020 generates an inverted signal of the distortion as follows.

number

[0037] The second example of the training data set 10 comprises the second transmitted symbol sequence as input data 12 and the first transmitted symbol sequence as ground truth data 14. By using this type of training data set 10, the generator is trained to generate the first transmitted symbol sequence from the second transmitted symbol sequence (in other words, to recover the original digital signal from the distorted digital signal). This means that in this case the generator 200 can act as a post-compensator. Thus, the trainable parameters of the generator 200 trained using the second example of the training data set 10 can be applied to the distortion compensator 140 acting as a post-compensator.

[0038] A third example of the training data set 10 includes the second transmission symbol sequence normalized by its gain as input data 12 and the first transmission symbol sequence as ground truth data 14. The second transmission symbol sequence is Department The input data 12 is normalized to remove the amplification applied to the optical transmission signal by the optical transmitter 112, thereby bringing the input data 12 into the same range as the ground truth data 14. By using this type of training data set 10, the generator 200 is trained to generate the first transmission symbol sequence from the second transmission symbol sequence normalized by its gain. This means that in this case the generator 200 can act as a pre-compensator. Thus, the trainable parameters of the generator 200 trained with the third example of the training data set 10 can be applied to the distortion compensator 140 placed before the optical transmission unit 112 to act as a pre-compensator.

[0039] <About the generator 200> The generator 200 is implemented as any machine learning based model, such as a neural network or a support vector machine. When the generator 200 is implemented as a neural network, the generator 200 preferably includes a set of layers with linear activations and a set of layers with non-linear activations. This structure allows modeling the combined effect of both linear and non-linear distortions. This allows generating a distortion compensator 140 that can compensate for a combined linear and non-linear distortion. Note that the trainable parameters of the generator 200 may include the weights and biases of each layer when the generator 200 is implemented as a neural network.

[0040] Figure 6 shows an example configuration of a generator 200 that includes a set of layers with linear activation and a set of layers with non-linear activation. In Figure 6, the generator 200 includes an input layer 210, a first layer set 220, a second layer set 230, and an output layer 240. Each layer of the layer sets 220 and 230 includes trainable weights with activation functions. Note that for simplicity, edges between layers are not shown in Figure 6.

[0041] The first layer set 220 is a set of layers with linear activation functions, and the second layer set 230 is a set of layers that include non-linear activation functions, such as a Rectified Linear Unit (ReLU) function.

[0042] The size of the output layer 240 is set to the same size as the symbol. If the optical communication system 100 is a single-polarized, single-channel optical communication system, the size of the symbol is 2 since the symbol contains an in-phase component and a quadrature component. Thus, the size of the output layer 240 is set to 2. On the other hand, in the case of a dual-polarized system, the size of the symbol is 4 since the symbol contains an in-phase component and a quadrature component for each of the two polarizations. Thus, the size of the output layer 240 can be set to 4. Alternatively, the generator 200 can include two separate neural networks, one of which generates a set of in-phase and quadrature components for the X polarization and the other of which generates the set for the Y polarization.

[0043] Regarding the input layer 210, in order to take into account the memory effect in the distortion characteristics, it is preferable to set the size of the input layer 210 larger than that of the output layer 240. Suppose the size of the symbol is 2 and the effect of the past (M-1) symbols is considered. In this case, the size of the output layer 240 is set to 2, while the size of the input layer 210 is set to 2*M since the input layer 210 is set to have two components for each of the current symbol and the past (M-1) symbols (i.e., a total of M symbols).

[0044] Theoretically, the first layer set 220 The output from can be expressed as:

number

[0045] Furthermore, the final output from the generator 200 (i.e., the output from the output layer 240) can be expressed as:

number

[0046] Note that the generator 200 may have a skip connection between the last layer of the first layer set 220 and the last layer of the second layer set 230. In this case, the final output from the generator 200 can be expressed as follows:

number

[0047] In addition, when the generator 200 has the above-mentioned skip connection, in order to connect the first layer set 220 and the second layer set 230, the final layer of the first layer set 220 is configured to have the same number of nodes as the first layer of the second layer set 230.

[0048] X_out indicates the properties of a function that can be modeled by the generator 200. As we will see later, we improve the expressiveness of this function by dynamically adjusting the number of trainable weights and layers of the generator 200.

[0049] <Regarding the classifier 300> The classifier 300 can be implemented as any machine learning based model, such as a neural network or a support vector machine. For example, the classifier 300 is implemented as a neural network having two nodes whose output layer outputs a one-hot vector representing the classification result. Specifically, for example, the output layer outputs a vector (0,1) when the classifier 300 determines that the input data is generated by the generator 200, and outputs a vector (1,0) when the classifier 300 determines that the input data is ground truth data 14. In the following, for easy understanding and concise explanation, the data output from the generator 200 will be described as generated data, and the ground truth data 14 will be described as true data.

[0050] <Training the generator 200 and the discriminator 300: S106> The training execution unit 2060 executes training of the model set 400 (S106). Conceptually, the training execution unit 2060 executes GAN-like training on the model set 400. The generator 200 is trained to generate data that is determined to be true data by the classifier 300, and the classifier 300 is trained to determine that the data generated by the generator 200 is not true data.

[0051] 7 shows an example of a flow of training the model set 400. The training executor 2060 initializes the model set 400 (S202). For example, the training executor 2060 may obtain hyperparameters of the generator 200 (such as predefined input / output sizes, predefined number of nodes in each layer, and predefined number of layers). The training executor 2060 initializes the generator 200 using the obtained hyperparameters. The trainable parameters of the generator 200 are also initialized.

[0052] The same applies to the classifier 300. Specifically, the training execution unit 2060 can acquire hyperparameters of the classifier 300, initialize the classifier 300 with the acquired hyperparameters, and initialize trainable parameters of the classifier 300.

[0053] After initializing the model set 400, the training executor 2060 trains the generator 200 without linking it with the classifier 300 (S204). As described above, the generator 200 is trained using a training data set 10. For example, for each training data set 10, the training executor 2060 inputs input data 12 to the generator 200, obtains an output from the generator 200, calculates a loss by applying the ground truth data 14 and the output from the generator 200 to a predetermined loss function, and updates trainable parameters of the generator 200 based on the calculated loss.

[0054] Next, the training execution unit 2060 executes training of the classifier 300 (S206). 3 00 is trained using ground truth data 14 and the output from the generator 200 trained in step S204.

[0055] Specifically, for each training data set 10, the training execution unit 2060 inputs the input data 12 to the generator 200 and obtains an output from the generator 200. Then, the training execution unit 2060 inputs the data output from the generator 200 to the discriminator 300 and trains the discriminator 300 to determine that the input data is generated data.

[0056] Specifically, for example, the training execution unit 2060 acquires an output from the classifier 300, calculates a loss by applying an actual output and an ideal output from the classifier 300 to a predetermined loss function, and updates trainable parameters of the classifier 300. The ideal output in this case is data indicating that the data input to the classifier 300 is generated data (for example, a one-hot vector (0,1) described in the above description of the classifier 300).

[0057] The training execution unit 2060 also inputs the ground truth data 14 to the classifier 300 and trains the classifier 300 to determine that the input data is true data. The trainable parameters of the classifier 300 can be updated in a manner similar to the manner in which those parameters are updated when the data generated by the generator 200 is input to the classifier 300. In this case, the ideal output used to calculate the loss is data indicating that the data input to the classifier 300 is true data (for example, the one-hot vector (1,0) described in the above description of the classifier 300).

[0058] After the generator 200 and the classifier 300 are individually trained, the training execution unit 2060 trains the generator 200 in cooperation with the classifier 300 (S208). Specifically, for each training data set 10, the training execution unit 2060 operates as follows. The training execution unit 2060 inputs the input data 12 to the generator 200 trained in step 204, and obtains an output from the generator 200. Next, the training execution unit 2060 inputs the output obtained from the generator 200 to the classifier 300 trained in step 206, and obtains an output from the classifier 300. Then, the training execution unit 2060 calculates a loss, and updates the trainable parameters of the generator 200 based on the calculated loss. In this case, the loss is calculated by applying the actual output and the ideal output from the classifier 300 to a predetermined loss function. In this case, the ideal output is data indicating that the data input to the classifier 300 is true data (for example, the one-hot vector (1,0) described in the above description of the classifier 300).

[0059] When step 208 is completed, the training execution unit 2060 determines whether a predetermined end condition is satisfied (S210). Various types of end conditions can be used here. For example, the end condition can be "completion of a predetermined number of repetitive executions of steps S204 to S212."

[0060] Alternatively, for example, the termination condition may be "the performance of the model set 400 has achieved a predetermined goal." The performance of the model set 400 may be calculated using a predetermined loss function. For example, the loss calculated in step 204, 206, or 208 may be used to represent the performance of the model set 400. In this case, for example, the training execution unit 2060 determines that "the performance of the model set 400 has achieved a predetermined goal" if the loss is less than a predetermined threshold.

[0061] If the end condition is satisfied (S210: YES), the training execution unit 2060 ends the training of the model set 400. On the other hand, if the end condition is not satisfied (S210: NO), the training execution unit 2060 executes step 212.

[0062] In step 212, the training execution unit 2060 changes the configuration of the model set 400 based on a predetermined rule (hereinafter, the change rule). The configuration of the model set 400 to be changed may be the number of layers in the layer set 220, the number of layers in the layer set 230, or both.

[0063] It is preferable to pre-set appropriate modification rules that allow the model set 400 to converge to an appropriate configuration (eg, the generator 200 has an optimal number of layers in the layer set 220 and the layer set 230). change An example rule is "iteratively increase the number of layers in layer set 230 until the loss is below a threshold, then iteratively decrease the number of layers in layer set 220 until the point where decreasing the number of layers in layer set 220 leads to increased loss."

[0064] The modification of the generator 200 (step 212) is performed to simplify and improve the accuracy of the distortion compensator 140 as much as possible. Specifically, since each optical transmitter may have its own unique characteristics, there is no optimal configuration (number of weights and layers) of the distortion compensator 140 for all optical transmitters. It is also possible to statically configure the generator 200 with many weights and layers without considering the complexity of the distortion compensator 140. However, the higher the complexity of the generator 200, the more computation time and resources are required to train the generator 200. In addition, it also reduces the efficiency of the distortion compensator 140.

[0065] By repeatedly modifying the generator 200, the generator 200 is gradually adjusted to be as simple as possible while maintaining its high accuracy. This allows the configuration of the distortion compensator 140 to be customized according to the characteristics of the optical transmitter 112, and allows accurate distortion compensation to be performed with less computation time and resources.

[0066] After step S212, step 204 is executed again. Specifically, steps 204 to S212 are repeatedly executed until it is determined in step 210 that the end condition is satisfied.

[0067] <Parameter information output: S108> The output unit 2080 outputs the parameter information 20 (S108). The parameter information 20 includes trainable parameters of the generator 200. As described above, the trainable parameters of the trained generator 200 can be applied to the distortion compensator 140.

[0068] The parameter information 20 may also include hyperparameters of the generator 200. In particular, as described above, when the configuration of the generator 200 is changed based on a modification rule, the output unit 2080 also includes the parameters (e.g., the number of layers in the layer sets 220 and 230) changed based on the modification rule in the parameter information 20. In this case, the distortion compensator 140 is configured to have the number of layers specified in the parameter information 20.

[0069] There are various methods for outputting the parameter information 20. For example, the parameter information 20 is stored in a storage device accessible from the training device 2000. In another example, the parameter information 20 is transmitted to another computer (for example, a computer that configures the distortion compensator 140 using the parameter information 20).

[0070] <Another use of parameter information 20> The parameter information 20 can be used not only to configure the distortion compensator 140 but also to generate a mathematical model of the optical transmitter 112. In this case, the training data set 10 includes the first transmit symbol sequence as input data 12, and the second transmit symbol sequence or an arbitrarily scaled version of the second transmit symbol sequence as input data 13. symbol Based on this type of input data 12, the generator 200 can generate a sequence of first transmitted symbols column to the second transmitted symbol column (in other words, to generate a distorted signal from the original signal). Thus, the trainable parameters of the trained generator 200 are used to Department 112 models can be generated.

[0071] The training device 2000 performs GAN-like training on the model set 400, so that the generator 200 can be accurately trained. Ta By using it as a parameter of the mathematical model of the optical transmitting unit 112, the mathematical model of the optical transmitting unit 112 can be generated accurately.

[0072] After generating the mathematical model of the optical transmitting unit 112, the training device 2000 can generate the second transmission symbol sequence without using the optical transmitting unit 112. Specifically, the training device 2000 can obtain the second transmission symbol sequence by inputting the first transmission symbol sequence into the mathematical model of the optical transmitting unit 112 instead of the actual optical transmitting unit 112.

[0073] The second transmission symbol sequence output from the mathematical model can be used to generate parameter information 20 including parameters suitable for the distortion compensator 140. Specifically, the training device 2000 generates a training data set 10 based on the first transmission symbol sequence input to the mathematical model and the second transmission symbols output from the mathematical model in response to the first transmission symbol sequence being input to the mathematical model. At this time, the training data set 10 suitable for training the model set 400 is generated, and parameters of the distortion compensator 140 are generated. Then, the training device 2000 executes training of the model set 400 using the training data set 10 generated here, and generates and outputs parameter information 20 including parameters of the trained generator 200.

[0074] Although the present invention has been described above with reference to the embodiment, the present disclosure is not limited to the above embodiment. Various modifications that can be understood by a person skilled in the art can be made to the configuration and details of the present disclosure within the scope of the invention.

[0075] The program includes a set of instructions (or software code) that, when loaded into a computer, causes the computer to perform one or more functions described in the embodiments. The program may be stored in a non-transitory computer-readable medium or a tangible storage medium. By way of example and not limitation, the computer-readable medium or tangible storage medium includes random-access memory (RAM), read-only memory (ROM), flash memory, solid-state drive (SSD) or other memory technology, CD-ROM, digital versatile disc (DVD), Blu-ray® disk or other optical disk storage, magnetic cassette, magnetic tape, magnetic disk storage or other magnetic storage device. The program may be transmitted on a transitory computer-readable medium or communication medium. By way of example and not limitation, the transitory computer-readable medium or communication medium includes electrical, optical, acoustic, or other forms of propagated signals.

[0076] A part or all of the above-described embodiments can be described as, but is not limited to, the following supplementary notes. <Additional Notes> (Appendix 1) At least one processor; a storage element for storing instructions; The at least one processor executes the instructions to: obtaining a first transmission symbol sequence which is a sequence of symbols to be input to an optical transmission unit included in a transmitter, the optical transmission unit obtaining the first transmission symbol sequence, converting the first transmission symbol sequence into an optical transmission signal, and transmitting the optical transmission signal to a receiver; obtaining a second transmission symbol sequence which is a sequence of symbols obtained by demodulating the optical transmission signal; generating a training data set including input data and ground truth data based on the first transmitted symbol sequence and the second transmitted symbol sequence; training a set of generators and classifiers using the training dataset, the generators being trained to generate data to be determined by the classifier to be the ground truth data, and the classifiers being trained to distinguish between the ground truth data and the data generated by the generators; A training device that outputs parameter information including trainable parameters of the generator. (Appendix 2) The training device described in Appendix 1, wherein the trainable parameters of the generator included in the parameter information are output as parameters to be applied to a pre-compensator located before the optical transmitting unit in the transmitter or to a post-compensator located after the optical receiving unit in the receiver, and the optical receiving unit receives the optical transmission signal transmitted by the transmitter and converts the received signal into a string of symbols. (Appendix 3) The generation of the training data includes: generating an inverted signal of the distortion included in the optical transmission signal based on the second transmission symbol sequence; 3. The training apparatus of claim 1 or 2, further comprising generating the training data set including the first transmitted symbol sequence as the input data and an inverted signal of the distortion as the ground truth data. (Appendix 4) 3. The training device of claim 1, wherein generating the training data includes generating the training data set in which the second transmission symbol sequence is included as the input data and the first transmission symbol sequence is included as the ground truth data. (Appendix 5) 3. The training device of claim 1, wherein generating the training data includes generating the training data set in which the second transmission symbol sequence normalized by its gain is included as the input data and the first transmission symbol sequence is included as the ground truth data. (Appendix 6) 6. The training apparatus of any one of claims 1 to 5, wherein the generator includes a first set of layers with linear activation functions and a second set of layers with non-linear activation functions. (Appendix 7) 7. The training apparatus of claim 6, wherein the training of the set of generators and classifiers includes changing a number of layers in the first set of layers, changing a number of layers in the second set of layers, or both. (Appendix 8) 8. The training apparatus of claim 6 or 7, wherein the generator includes a skip connection connecting a last layer of the first set of layers and a last layer of the second set of layers. (Appendix 9) The at least one processor further comprises: obtaining a mathematical model of the optical transmitter to which the trainable parameters included in the parameter information should be applied; generating the second transmission symbol sequence by inputting the first transmission symbol sequence into the mathematical model; generating a second training data set including the input data and the ground truth data based on the first transmission symbol sequence input to the mathematical model and the second transmission symbol sequence output from the mathematical model; performing the training of the set of generators and classifiers using the second training data set; 9. The training apparatus of any one of claims 1 to 8, wherein the training apparatus outputs the parameter information including trainable parameters of the generator trained using the second training data set. (Appendix 10) obtaining a first transmission symbol sequence which is a sequence of symbols to be input to an optical transmission unit included in a transmitter, the optical transmission unit obtaining the first transmission symbol sequence, converting the first transmission symbol sequence into an optical transmission signal, and transmitting the optical transmission signal to a receiver; obtaining a second transmission symbol sequence which is a sequence of symbols obtained by demodulating the optical transmission signal; generating a training data set including input data and ground truth data based on the first transmitted symbol sequence and the second transmitted symbol sequence; training a set of generators and classifiers using the training dataset, the generators being trained to generate data to be determined by the classifier to be the ground truth data, and the classifiers being trained to distinguish between the ground truth data and the data generated by the generators; and outputting parameter information including trainable parameters of the generator. (Appendix 11) The control method of claim 10, wherein the trainable parameters of the generator included in the parameter information are output as parameters to be applied to a pre-compensator located before the optical transmitting unit in the transmitter or to a post-compensator located after the optical receiving unit in the receiver, and the optical receiving unit receives the optical transmission signal transmitted by the transmitter and converts the received signal into a string of symbols. (Appendix 12) The generation of the training data includes: generating an inverted signal of the distortion included in the optical transmission signal based on the second transmission symbol sequence; 12. The control method of claim 10 or 11, further comprising generating the training data set including the first transmission symbol sequence as the input data and an inverted signal of the distortion as the ground truth data. (Appendix 13) 12. The control method according to claim 10, wherein generating the training data includes generating the training data set including the second transmission symbol sequence as the input data and the first transmission symbol sequence as the ground truth data. (Appendix 14) 12. The control method according to claim 10, wherein generating the training data includes generating the training data set in which the second transmission symbol sequence normalized by its gain is included as the input data and the first transmission symbol sequence is included as the ground truth data. (Appendix 15) 15. The method of claim 10, wherein the generator includes a first set of layers with linear activation functions and a second set of layers with non-linear activation functions. (Appendix 16) 16. The control method of claim 15, wherein the training of the set of generators and classifiers includes changing a number of layers in the first set of layers, changing a number of layers in the second set of layers, or both. (Appendix 17) 17. The control method of claim 15 or 16, wherein the generator includes a skip connection connecting a last layer of the first set of layers and a last layer of the second set of layers. (Appendix 18) obtaining a mathematical model of the optical transmitter to which the trainable parameters included in the parameter information should be applied; inputting the first transmission symbol sequence into the mathematical model to generate the second transmission symbol sequence; generating a second training data set including the input data and the ground truth data based on the first transmission symbol sequence input to the mathematical model and the second transmission symbol sequence output from the mathematical model; performing the training of the set of generators and classifiers using the second training data set; and outputting the parameter information including trainable parameters of the generator trained using the second training data set. (Appendix 19) obtaining a first transmission symbol sequence which is a sequence of symbols to be input to an optical transmission unit included in a transmitter, the optical transmission unit obtaining the first transmission symbol sequence, converting the first transmission symbol sequence into an optical transmission signal, and transmitting the optical transmission signal to a receiver; obtaining a second transmission symbol sequence which is a sequence of symbols obtained by demodulating the optical transmission signal; generating a training data set including input data and ground truth data based on the first transmitted symbol sequence and the second transmitted symbol sequence; training a set of generators and classifiers using the training dataset, the generators being trained to generate data to be determined by the classifier to be the ground truth data, and the classifiers being trained to distinguish between the ground truth data and the data generated by the generators; A non-transitory computer-readable medium storing a program for causing a computer to execute the program to output parameter information including trainable parameters of the generator. (Appendix 20) The storage medium of claim 19, wherein the trainable parameters of the generator included in the parameter information are output as parameters to be applied to a pre-compensator located before the optical transmitting unit in the transmitter or to a post-compensator located after the optical receiving unit in the receiver, and the optical receiving unit receives the optical transmission signal transmitted by the transmitter and converts the received signal into a string of symbols. (Appendix 21) The generation of the training data includes: generating an inverted signal of the distortion included in the optical transmission signal based on the second transmission symbol sequence; 21. The storage medium of claim 19 or 20, further comprising generating the training data set including the first transmitted symbol sequence as the input data and an inverted signal of the distortion as the ground truth data. (Appendix 22) 21. The storage medium of claim 19, wherein generating the training data includes generating the training data set in which the second transmission symbol sequence is included as the input data and the first transmission symbol sequence is included as the ground truth data. (Appendix 23) 21. The storage medium of claim 19, wherein generating the training data includes generating the training data set in which the second transmission symbol sequence normalized by its gain is included as the input data and the first transmission symbol sequence is included as the ground truth data. (Appendix 24) 24. The storage medium of any one of claims 19 to 23, wherein the generator includes a first set of layers with linear activation functions and a second set of layers with non-linear activation functions. (Appendix 25) 25. The storage medium of claim 24, wherein the training of the set of generators and classifiers includes changing a number of layers in the first set of layers, changing a number of layers in the second set of layers, or both. (Appendix 26) 26. The storage medium of claim 24 or 25, wherein the generator includes a skip connection connecting a last layer of the first set of layers and a last layer of the second set of layers. (Appendix 27) The program is obtaining a mathematical model of the optical transmitter to which the trainable parameters included in the parameter information should be applied; generating the second transmission symbol sequence by inputting the first transmission symbol sequence into the mathematical model; generating a second training data set including the input data and the ground truth data based on the first transmission symbol sequence input to the mathematical model and the second transmission symbol sequence output from the mathematical model; performing the training of the set of generators and classifiers using the second training data set; and outputting the parameter information including trainable parameters of the generator trained using the second training data set. [Explanation of symbols]

[0077] 10 Training data set 12 Input Data 14 Ground Truth Data 20 Parameter Information 50 Intermediate optical receiver 100 Optical communication system 110 Transmitter 112 Optical transmitter 120 Receiver 122 Optical receiver 130 Optical Fiber 140 Distortion Compensator 200 generator 210 Input Layer 220 1st Layer Set 230 2nd Layer Set 240 Output Layers 300 Classifier 400 Model Set 1000 Computers 1020 Bus 1040 Processor 1060 Memory 1080 Storage Device 1100 Input / Output Interface 1120 Network Interface 2000 training equipment 2020 Acquisition Department 2040 Generation part 2060 Training Execution Department 2080 Output section

Claims

1. An acquisition unit that acquires a first transmission symbol sequence, which is a sequence of symbols to be input to an optical transmission unit included in a transmitter; the optical transmitter obtains the first transmission symbol sequence, converts the first transmission symbol sequence into an optical transmission signal, and transmits the optical transmission signal to a receiver; the acquiring unit further acquires a second transmission symbol sequence which is a sequence of symbols obtained by demodulating the optical transmission signal; a generator that generates a training data set including input data and ground truth data based on the first transmission symbol sequence and the second transmission symbol sequence; a training execution unit that executes training of a set of a generator and a classifier using the training data set; the generator is trained to generate data to be determined by the classifier to be the ground truth data; the classifier is trained to distinguish between the ground truth data and the data generated by the generator; A training device having an output for outputting parameter information including trainable parameters of the generator.

2. The trainable parameters of the generator included in the parameter information are output as parameters to be applied to a pre-compensator arranged before the optical transmission unit in the transmitter, or output as parameters to be applied to a post-compensator arranged after the optical reception unit in the receiver; 2. The training device of claim 1, wherein the optical receiver receives the optical transmission signal sent by the transmitter and converts the received signal into a string of symbols.

3. The generation unit is generating an inverted signal of the distortion included in the optical transmission signal based on the second transmission symbol sequence; 3. A training apparatus according to claim 1 or 2, configured to generate the training data set, the training data set including the first transmitted symbol sequence as the input data and an inverse signal of the distortion as the ground truth data.

4. The training device according to claim 1 , wherein the generation unit generates the training data set in which the second transmission symbol sequence is included as the input data and the first transmission symbol sequence is included as the ground truth data.

5. 3. The training device according to claim 1, wherein the generation unit generates the training data set in which the second transmission symbol sequence normalized by a gain of the generation unit is included as the input data and the first transmission symbol sequence is included as the ground truth data.

6. 6. The training apparatus of claim 1, wherein the generator comprises a first set of layers with linear activation functions and a second set of layers with non-linear activation functions.

7. 7. The training apparatus of claim 6, wherein training the generator and the classifier includes varying a number of layers in the first set of layers, varying a number of layers in the second set of layers, or both.

8. 8. A training apparatus according to claim 6 or 7, wherein the generator includes a skip connection connecting the last layer of the first set of layers and the last layer of the second set of layers.

9. obtaining a first transmission symbol sequence which is a sequence of symbols to be input to an optical transmission unit included in a transmitter; the optical transmitter obtains the first transmission symbol sequence, converts the first transmission symbol sequence into an optical transmission signal, and transmits the optical transmission signal to a receiver; obtaining a second transmission symbol sequence, the second transmission symbol sequence being a symbol sequence obtained by demodulating the optical transmission signal; generating a training data set including input data and ground truth data based on the first transmitted symbol sequence and the second transmitted symbol sequence; and performing training of a set of generators and classifiers using the training data set; the generator is trained to generate data to be determined by the classifier to be the ground truth data; the classifier is trained to distinguish between the ground truth data and the data generated by the generator; and outputting parameter information including trainable parameters of the generator.

10. A method for obtaining a first transmission symbol sequence, which is a sequence of symbols to be input to an optical transmission unit included in a transmitter, comprising: the optical transmitter obtains the first transmission symbol sequence, converts the first transmission symbol sequence into an optical transmission signal, and transmits the optical transmission signal to a receiver; obtaining a second transmission symbol sequence, the second transmission symbol sequence being a symbol sequence obtained by demodulating the optical transmission signal; generating a training data set including input data and ground truth data based on the first transmitted symbol sequence and the second transmitted symbol sequence; training a set of generators and classifiers using the training dataset; the generator is trained to generate data to be determined by the classifier to be the ground truth data; the classifier is trained to distinguish between the ground truth data and the data generated by the generator; A program that causes the computer to output parameter information including trainable parameters of the generator.

Images