122. As a result, a large amount of measurement data can be collected, but the data is data having variations such as a difference in type of the m wireless terminal, a difference in antenna performance, and how the terminal is held at the time of measurement. Therefore, an error occurs between the wireless quality estimated by the model trained by such data and the actual wireless quality of the terminal A, which is the target of the wireless quality estimation.

Therefore, in the present embodiment, transfer learning is performed on a model on which the preliminary learning has been performed using measurement data obtained from terminals of a large number of users to create a transfer trained model (a retrained model).

In the transfer learning, the transfer learning control unit 123 selects measurement data of a terminal close to the terminal A, which is the target of the wireless quality estimation, from among the measurement data stored in the database unit 122, and sets a range of a layer in a model of a neural network that is a retraining portion in the transfer learning, in the learning unit 121.

In the example of FIG. 5(b), three intermediate layers before the output layer are set as the retraining portion of the transfer learning. As described above, the setting of the retraining portion may be performed in units of layers, or the retraining portion may be set more finely in units of neurons in the layers. In any case, the range of neurons in the neural network is designated as the retraining portion.

Regarding the portion in the pre-trained model to be set as the retraining portion, it may be a predetermined portion (for example, three intermediate layers before the output layer), an optimal retraining portion may be determined by performing an experiment, or it may be determined by another method.

In the transfer learning (retraining), the data (the position information and the terminal information) of the terminal close to the terminal A, which is the target of the wireless quality estimation, is input to the pre-trained model, and the weight parameter is adjusted such that the output becomes the correct data (the wireless quality) of the terminal.

More specifically, the weight parameter of the pre-trained model becomes an initial value, only the weight parameter of the portion set as the retraining portion is adjusted, and the weight parameter of the portion other than the portion set as the retraining portion is maintained at the initial value. Alternatively, the learning rate of the portion other than the portion set as the retraining portion is set low to reduce the update amount of the weight parameter.

Note that the above example is an example of a learning method in retraining such as transfer learning, and is not limited to the above example.

In addition, in the transfer learning, a transfer trained model for each piece of terminal information (terminal type) may be generated. For example, from the same pre-trained model, a transfer trained model A1 is generated using data A1 of a terminal close to a terminal A1, and a transfer trained model A2 is generated using data A2 of a terminal close to a terminal A2. Then, for example, in a case where the terminal A2 is used as the terminal to be actually used, the wireless quality is estimated using the transfer trained model A2.

Examples of the terminal close to the terminal A, which is the target of the wireless quality estimation include the terminal A itself, a terminal of a device model the same as the terminal A, a terminal having a shape the same as the terminal A (a smartphone type, a mobile router type, or the like), a terminal mounted with a wireless communication chip having a model number the same as the model number of the wireless communication chip mounted on the terminal A, a terminal using wireless standards the same as the wireless standards used by the terminal A, a terminal whose antenna performance (gain, half-value width, and the like) falls under a difference within a certain range or less in comparison with the terminal A, and a terminal used under a use condition the same as the use condition assumed by the terminal A (walking, in-vehicle, or the like).

The measurement position of the terminal close to the terminal A is not necessarily geographically close to the position of the terminal A (the position to be estimated), and the measurement may be performed by the terminal close to the terminal A at a place different from the place where the terminal A exists.

The wireless quality of the terminal A can be more accurately estimated by using the transfer trained model obtained by performing learning again with the measurement data of the terminal close to the terminal A.

It is assumed that the data for retraining is less than other data. By performing retraining using data based on a terminal close to the terminal A with respect to the pre-trained model based on the measurement data of another terminal, wireless quality for the terminal A can be estimated with high accuracy even if there is little data based on a terminal close to the terminal A.

Operation Flow Example

An operation flow example of the wireless quality estimation device 100 in the present embodiment will be described with reference to FIGS. 6 to 8. FIG. 6 illustrates a flow at the time of pre-training of all terminal data.

In S101, the “location information, terminal information, and wireless quality information” for each terminal is input to the wireless quality estimation unit 120 as learning data and recorded in the database unit 122. Here, the measurement data obtained from various terminals is recorded in the database unit 122.

In S102, the learning unit 121 reads the data group from the database unit 122 and performs learning of the model using a machine learning algorithm. As a result, the pre-trained model is generated. The pre-trained model is held in the learning unit 121.

FIG. 7 illustrates a flow at the time of learning using specific terminal data.

In S201, the transfer learning control unit 123 extracts data of the target specific terminal (that is, the data of the terminal close to the terminal A) from the data recorded in the database unit 122.

In S202, the transfer learning control unit 123 designates the range of the layer of the neural network to be the retraining portion for the pre-trained model for the learning unit 121.

In S203, the transfer learning control unit 123 inputs the data extracted in S201 to the learning unit 121 in which the retraining portion is set. The learning unit 121 performs retraining on the pre-trained model using the data extracted in S201. As a result, a model on which the transfer learning has been performed is generated, and the learning unit 121 retains the model.

FIG. 8 illustrates a flow at the time of estimating the wireless quality for the terminal A.

In S301, the position information acquisition unit 111 acquires the current position information of the terminal A, and the estimated position designation unit 112 selects a position where the wireless quality is desired to be estimated for the terminal A based on the current position information.