Multi-classifier fusion method based on quantum convolutional neural network

Abstract

The invention provides a multi-classifier fusion method based on a quantum convolutional neural network, and aims to solve the problem of large deep learning computing resource demand and the problem of poor multi-classifier fusion effect, the method is based on a parameterized quantum circuit, the entanglement capability and expressible capability of the circuit are considered, and the multi-classifier fusion effect is improved. Three kinds of quantum convolutional neural networks with different convolutional structures are provided as base classifiers, required quantum bits and training parameters are reduced, and meanwhile, high classification accuracy and certain difference are achieved; according to the method, the output result of each base classifier is converted into an evidence form, and an improved quantum average combination method is used as a fusion method of multiple quantum classifiers, so that conflicts among the base classifiers are solved, and accurate classification of a multi-classifier fusion system is realized.

Description

Technical Field

[0001] The invention belongs to the field of quantum neural networks, and in particular relates to a multi-classifier fusion method based on quantum convolutional neural networks. Background Art

[0002] Multi-classifier fusion has stronger generalization ability than base learners and can reduce the influence of uncertainty factors, that is, it can turn weak learners that are slightly better than random guessing into strong learners that can accurately predict. Therefore, ensemble learning methods have received widespread attention. However, in practical applications, the results of multi-classifier fusion are not good using many existing ensemble learning methods, and are even worse than the results of a single classifier. On the one hand, this is because the base classifiers are not accurate and different; on the other hand, it is because the multi-classifier fusion method is not effective enough. In order to improve the performance of the base classifier, the network used for training is becoming more and more complex, and the demand for computing resources is also increasing. For example, the convolution process in convolutional neural networks is one of the main sources of resource demand in machine learning. Quantum computing has become an important direction for breaking through the bottleneck of computing performance due to its high-speed parallelism; the existing multi-classifier fusion method of quantum convolutional neural network has the problem that the complex network structure of the base classifier leads to large computing resource requirements and low accuracy of the multi-classifier fusion method. Summary of the invention

[0003] In order to overcome the shortcomings of the prior art and address the problem of large demand for computing resources for deep learning, the present invention is based on parameterized quantum circuits. Through the entanglement ability and expressibility of circuits, three quantum convolutional neural networks with different convolutional structures are proposed as base classifiers to reduce the required quantum bits and training parameters while having high classification accuracy and certain differences. In addition, in order to address the problem of poor fusion effect of multiple classifiers, the present invention converts the output results of each base classifier into evidence form, uses the improved quantum average combination method as the fusion method of multiple quantum classifiers, resolves the conflicts between the base classifiers, and realizes accurate classification of the multi-classifier fusion system.

[0004] A multi-classifier fusion method based on quantum convolutional neural network includes the following steps:

[0005] Step S100: constructing a base classifier based on parameterized quantum circuits;

[0006] Step S200: Based on the improved quantum averaging method, multiple base classifiers are fused.

[0007] Furthermore, the steps of constructing a base classifier based on parameterized quantum circuits are:

[0008] Step S101: constructing a quantum state preparation circuit;

[0009] Step S102: constructing a quantum convolutional neural network;

[0010] The quantum convolutional neural network constructed based on parameterized quantum circuits is the base classifier based on parameterized quantum circuits.

[0011] Furthermore, the steps of constructing the quantum state preparation circuit are:

[0012] Step S101: constructing a quantum state preparation circuit;

[0013] The quantum state preparation circuit is a quantum amplitude encoding circuit;

[0014] In quantum networks, input data χ is converted to a different space through quantum feature mapping, i.e., φ:χ→κ, where φ represents the mapping process and κ represents a Hilbert space;

[0015] Quantum amplitude coding circuit is to convert N-dimensional data x=(x1, x2,…, x N ) is encoded into an n-bit quantum state, expressed as follows:

[0016]

[0017] Where U(x) represents the quantum amplitude coding circuit, represents N-dimensional real number space, Represents the encoded quantum state, N = 2 n , i> represents the i-th ground state.

[0018] Furthermore, the steps of constructing the quantum convolutional neural network are:

[0019] Step S102-1: constructing a quantum convolution filter;

[0020] The quantum convolution filter includes two qubits, and the qubits include a combination of a single qubit gate and a double qubit gate circuit containing parameters;

[0021] The parameterized quantum circuits of the quantum convolution filters in the same convolution layer are the same, and the quantum convolution filters are equivalent to the convolution kernels in CNN.

[0022] On the basis of the quantum convolution filter, #1 quantum convolution filter, #2 quantum convolution filter and #3 quantum convolution filter are constructed; the input of the quantum convolution filter is the data converted into quantum form;

[0023] #1 The structure of the quantum convolution filter is:

[0024] The data converted into quantum form is input into two quantum bits |0> and |1>, and R x (θ x ), R z(θ Z ) and the controlled NOT gate CNOT act on |0> and |1> to form a basic unit u1; four u1s are connected in series to form a 1# quantum convolution filter; R x (θ x ) is the x revolving door, R z (θ Z ) is the y revolving door; θ x The initial value of is a random value; Z The initial value of is a random value;

[0025] #2 The structure of the quantum convolution filter is:

[0026] The data converted into quantum form is input into the two quantum bits |0> and |1>, and the H gate, the controlled Z gate and the R x (θ x ) acts on |0> and |1> to form a basic unit u2; four u2s are connected in series to form a 2# quantum convolution filter; the H gate is a Hadamard gate;

[0027] #3 The structure of the quantum convolution filter is: 3 u1s and 1 u2 are connected in series to form a quantum convolution filter;

[0028] Step S102-2: constructing a quantum pooling structure;

[0029] The quantum pooling structure is a two-qubit gate circuit. The two-qubit gate circuit consists of a CNOT gate and does not contain free parameters. p express;

[0030] Step S102-3: constructing a quantum convolutional neural network;

[0031] The convolutional neural network based on parameterized quantum circuit of the present invention comprises a quantum state preparation circuit, a first convolutional layer, a first pooling layer, a second convolutional layer and a second pooling layer, and the quantum pooling circuit of the pooling layer has the same structure and parameters;

[0032] The first convolutional layer contains four identical quantum convolutional filters;

[0033] The first pooling layer contains two U p ;

[0034] The second convolutional layer contains a quantum convolution filter;

[0035] The second pooling layer contains a U p ;

[0036] The quantum convolutional neural network based on #1 quantum convolution filter is BC-QCNN1;

[0037] The quantum convolutional neural network based on #2 quantum convolution filter is BC-QCNN2;

[0038] The quantum convolutional neural network based on #3 quantum convolution filter is BC-QCNN3.

[0039] Furthermore, the steps of fusing multiple base classifiers based on the improved quantum averaging method are as follows:

[0040] Step S201: construct the evidence form corresponding to the base classifier;

[0041] For the binary classification task, assuming that the recognition framework is Φ = {H1, H2}, the result obtained by the base classifier is p = [p1, p2], that is, p(H1) = p1, p(H2) = p2, which is converted into the basic probability distribution function m(A) through formula (2):

[0042]

[0043] Where α is the classification accuracy of the base classifier, which is used as a discount factor here; A is the proposition; the evidence corresponding to the classification results of the three base classifiers is obtained through formula (2), m1 is the evidence form corresponding to the classification result output by BC-QCNN1, m2 is the evidence form corresponding to the classification result output by BC-QCNN2, and m3 is the evidence form corresponding to the classification result output by BC-QCNN3;

[0044] For binary classification tasks, the output of the base classifier is a two-dimensional vector in the form of probability, so it is first necessary to convert it into the form of the basic probability distribution function in evidence theory;

[0045] Step S202: fusing multiple base classifiers based on an improved quantum averaging method;

[0046] The optimal phase parameters were obtained using the improved quantum averaging method, and the quantum model of the evidence was established;

[0047] Based on the evidence forms m1, m2 and m3 obtained by the base classifier, the quantum model |φ1> of m1, the quantum model |φ2> of m2 and the quantum model |φ3> of m3 are established. |φ1|, |φ2> and |φ3> have different constraints respectively. According to the improved quantum average combination method, |φ1>, |φ2> and |φ3> are fused, and the quantum model of the fused evidence is obtained as follows: After quantum measurement operation, the quantum state Collapse into different categories with corresponding probabilities, namely And thus determine the category to which the sample belongs.

[0048] The beneficial effects of the present invention are:

[0049] Compared with existing quantum convolutional neural networks, the BC-QCNN1, BC-QCNN2 and BC-QCNN3 proposed in the present invention have the same classification accuracy while requiring fewer parameters, effectively reducing the demand for computing resources; compared with other methods, the multi-classifier fusion method based on quantum convolutional neural network proposed in the present invention has achieved the highest classification accuracy in multiple classification tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is the overall framework diagram of the method of the present invention;

[0051] Figure 2 Schematic diagram of dimensionality reduction method;

[0052] Figure 3 This is a schematic diagram of the structure of #1 quantum convolution filter;

[0053] Figure 4 This is a schematic diagram of the structure of the #2 quantum convolution filter;

[0054] Figure 5 This is a schematic diagram of the structure of #3 quantum convolution filter;

[0055] Figure 6 Schematic diagram of quantum pooling structure;

[0056] Figure 7 This is the overall structure of the quantum convolutional neural network. DETAILED DESCRIPTION

[0057] The technical solution adopted by the present invention to solve the technical problem is: a multi-classifier fusion method based on quantum convolutional neural network, comprising the following steps:

[0058] Step S100: constructing a base classifier based on parameterized quantum circuits;

[0059] In the noisy medium-scale quantum era, the research on quantum neural networks mainly focuses on the construction of parameterized quantum circuits. Quantum neural networks undergo a quantum state preparation process to convert data into quantum data, which is then input into the network. Different from the weight matrix of a neural network, parameterized quantum circuits use the phase in the quantum gate as a training parameter, and achieve the transformation from the quantum initial state to the desired quantum final state through quantum gate operations, while introducing nonlinearity through quantum measurement operations.

[0060] The steps to construct a base classifier based on parameterized quantum circuits are:

[0061] Step S101: constructing a quantum state preparation circuit;

[0062] The quantum state preparation circuit is a quantum amplitude encoding circuit;

[0063] In quantum networks, input data χ is converted to a different space through quantum feature mapping, i.e., φ:χ→κ, where φ represents the mapping process and κ represents a Hilbert space;

[0064] Quantum amplitude coding circuit is to convert N-dimensional data x=(x1, x2,…, x N ) is encoded into an n-bit quantum state, expressed as follows:

[0065]

[0066] Where U(x) represents the quantum amplitude coding circuit, represents N-dimensional real number space, Represents the encoded quantum state, N = 2 n , i> represents the i-th ground state;

[0067] Step S102: constructing a quantum convolutional neural network;

[0068] The input image is processed by blank segmentation and average pooling downsampling methods in sequence;

[0069] The quantum convolutional neural network includes a quantum convolution layer and a quantum pooling layer;

[0070] The quantum convolution layer includes a quantum convolution filter; the quantum pooling layer includes a quantum pooling structure; the quantum convolution layer includes a #1 quantum convolution filter, a #2 quantum convolution filter, and a #3 quantum convolution filter;

[0071] The steps to construct the quantum convolutional neural network structure are:

[0072] Step S102-1: constructing a quantum convolution filter;

[0073] The quantum convolution filter includes two qubits, which include a combination of single-qubit gates and double-qubit gate circuits containing parameters; this is consistent with the role of the convolution kernel in the traditional convolutional network;

[0074] The parameterized quantum circuits of the quantum convolution filters in the same convolution layer are the same. The quantum convolution filter is equivalent to the convolution kernel in CNN. The design of the quantum convolution filter retains the characteristic of convolution kernel weight sharing.

[0075] On the basis of the quantum convolution filter, #1 quantum convolution filter, #2 quantum convolution filter and #3 quantum convolution filter are constructed; the input of the quantum convolution filter is the data converted into quantum form;

[0076] #1 The structure of the quantum convolution filter is:

[0077] The data converted into quantum form is input into two quantum bits |0> and |1>, and R x (θ x ), R z (θ Z ) and the controlled NOT gate CNOT act on |0> and |1> to form a basic unit u1; four u1s are connected in series to form a 1# quantum convolution filter; R x (θ x ) is the x revolving door, R z (θ Z ) is the y revolving door; θ x The initial value of is a random value; Z The initial value of is a random value;

[0078] #2 The structure of the quantum convolution filter is:

[0079] The data converted into quantum form is input into the two quantum bits |0> and |1>, and the H gate, the controlled Z gate and the R x (θ x ) acts on |0> and |1> to form a basic unit u2; four u2s are connected in series to form a 2# quantum convolution filter; the H gate is a Hadamard gate;

[0080] #3 The structure of the quantum convolution filter is: 3 u1s and 1 u2 are connected in series to form a quantum convolution filter; Figure 5 As shown;

[0081] #1 The quantum convolution filter contains 16 adjustable parameters and has strong expressive power;

[0082] #2 Quantum convolution filter contains 8 adjustable parameters and has strong entanglement ability;

[0083] #3 Quantum convolution filter comprehensively parameterizes the expressiveness and entanglement capabilities of quantum circuits, with 14 adjustable parameters;

[0084] Step S102-2: constructing a quantum pooling structure;

[0085] The quantum pooling structure is a two-qubit gate circuit. The two-qubit gate circuit consists of a CNOT gate and does not contain free parameters. p Indicates; Figure 6 As shown, the two-qubit gate circuit performs unitary transformation operations on the active qubits, thereby removing redundant information and reducing the complexity of the network;

[0086] The quantum convolutional neural network designed by the present invention uses a quantum pooling circuit to perform dimensionality reduction processing on the obtained quantum state characteristics;

[0087] Step S102-3: constructing a quantum convolutional neural network;

[0088] The convolutional neural network based on parameterized quantum circuit of the present invention comprises a quantum state preparation circuit, a first convolutional layer, a first pooling layer, a second convolutional layer and a second pooling layer, and the quantum pooling circuit of the pooling layer has the same structure and parameters;

[0089] The first convolutional layer contains four identical quantum convolutional filters;

[0090] The first pooling layer contains two U p ;

[0091] The second convolutional layer contains a quantum convolution filter;

[0092] The second pooling layer contains a U p ;

[0093] like Figure 7 As shown;

[0094] After preprocessing, the input image data is compressed and the feature vector is input into the quantum convolutional neural network. First, the data is unitarily transformed according to the quantum amplitude encoding circuit U(x) of the quantum state preparation circuit, and the feature vector is encoded into the corresponding quantum state. The first convolutional layer and the second convolutional layer and the first pooling layer and the second pooling layer are used to extract features and reduce the dimension of the quantum state data respectively; the same convolutional layer is composed of the same convolutional filter and has the same parameter θ; and the convolutional filters of different layers have the same structure but different parameters θ;

[0095] The quantum convolutional neural network based on #1 quantum convolution filter is BC-QCNN1;

[0096] The quantum convolutional neural network based on #2 quantum convolution filter is BC-QCNN2;

[0097] The quantum convolutional neural network based on #3 quantum convolution filter is BC-QCNN3;

[0098] The constructed BC-QCNN1, BC-QCNN2 and BC-QCNN3 are three base classifiers based on parameterized quantum circuits;

[0099] Step S200: fusing multiple base classifiers based on an improved quantum averaging method;

[0100] Based on the classification results of the three base classifiers BC-QCNN1, BC-QCNN2 and BC-QCNN3, the steps to achieve the fusion of multiple quantum base classifiers are as follows:

[0101] Step S201: construct the evidence form corresponding to the base classifier;

[0102] For the binary classification task, assuming that the recognition framework is Φ = {H1, H2}, the result obtained by the base classifier is p = [p1, p2], that is, p(H1) = p1, p(H2) = p2, which is converted into the basic probability distribution function m(A) through formula (2):

[0103]

[0104] Where α is the classification accuracy of the base classifier, which is used as a discount factor here; A is the proposition; the evidence corresponding to the classification results of the three base classifiers is obtained through formula (2), m1 is the evidence form corresponding to the classification result output by BC-QCNN1, m2 is the evidence form corresponding to the classification result output by BC-QCNN2, and m3 is the evidence form corresponding to the classification result output by BC-QCNN3;

[0105] For binary classification tasks, the output of the base classifier is a two-dimensional vector in the form of probability, so it is first necessary to convert it into the form of the basic probability distribution function in evidence theory;

[0106] Step S202: fusing multiple base classifiers based on an improved quantum averaging method;

[0107] The optimal phase parameters were obtained using the improved quantum averaging method, and the quantum model of the evidence was established;

[0108] Based on the evidence forms m1, m2 and m3 obtained by the base classifier, the quantum model |φ1> of m1 is established, the quantum model |φ2> of m2 is established, and the quantum model |φ3> of m3 is established. |φ1>, |φ2> and |φ3> have different constraints respectively. According to the improved quantum average combination method, |φ1>, |φ2> and |φ3> are fused, and the quantum model of the fused evidence is obtained as follows: After quantum measurement operation, the quantum state Collapse into different categories with corresponding probabilities, namely And thus determine the category to which the sample belongs.

[0109] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0110] The method of the present invention is described by taking the application of the method of the present invention on the MNIST data set as an example:

[0111] 1. Construction of base classifier based on parameterized quantum circuit;

[0112] Select two different digital categories in the MNIST dataset as different binary classification tasks, such as {0, 1} classification, {1, 2} classification, {7, 9} classification, etc.

[0113] According to step 1-1, three convolution filters with different entanglement capabilities and expression capabilities are constructed, which are represented as convolution filter #1, convolution filter #2 and convolution filter #3 respectively.

[0114] According to step 1-2, a quantum pooling structure is constructed, denoted as U p .

[0115] Based on steps 1-1 and 1-2, the overall structure of the quantum convolutional neural network is established according to step 1-3, denoted as BC-QCNN1, BC-QCNN2 and BC-QCNN3, with different entanglement capabilities and expressibility, respectively serving as different base classifiers for the binary classification task. Assume that the recognition framework corresponding to the binary classification task is Φ = {H1, H2}, and the output of BC-QCNN1 is denoted as p a =[p a1 ,p a2 ], and the output of BC-QCNN2 is recorded as p b =[p b1 ,p b2 ], and the output of BC-QCNN3 is recorded as p c =[p c1 ,p c2 ].

[0116] 2. Multi-classifier fusion based on improved quantum averaging method;

[0117] Through step 2-1, p a 、p b and p c Converted into basic probability distribution function, represented as m a 、m b and m c , where α a , α b and α c is the classification accuracy of each base classifier BC-QCNN1, BC-QCNN2 and BC-QCNN3.

[0118] Obtain m through step 2-2 a 、m b and m c The corresponding quantum model is represented by |φ a >,|φ b > and |φ c >, respectively with constraints C a , C b and C c . For |φ a >,|φ b > and |φ c >Perform quantum average fusion, and the fused model is expressed as And it has constraint condition C. Through quantum measurement, the probabilities p(H1) and p(H2) of samples in the MNIST data set belonging to different categories are obtained, and the category to which the sample belongs is determined according to the size of p(H1) and p(H2).

[0119] The present invention conducted comparative experiments on the Fashion-MNIST dataset to verify the classification performance of the base classifier and the effectiveness of the overall multi-classifier fusion method. Two different commodity categories in the Fashion-MNIST dataset were selected as different binary classification tasks. The evaluation indicators were the classification accuracy and the number of parameters θ. The comparative experimental results with other base classifiers are shown in Table 1, and the comparative experimental results of the overall multi-classifier fusion method are shown in Table 2. Compared with the existing quantum convolutional neural networks, the BC-QCNN1, BC-QCNN2 and BC-QCNN3 proposed in this patent have the same classification accuracy while requiring fewer parameters, which can reduce the demand for computing resources. Compared with other methods, the multi-classifier fusion method based on quantum convolutional neural network proposed in this patent has achieved the highest classification accuracy in multiple classification tasks.

[0120] Table 1 Comparison results of base classifiers on different binary classification tasks in Fashion-MNIST dataset

[0121]

[0122] Table 2 Comparison results of multi-classifier fusion method based on quantum convolutional neural network and other methods

[0123]

[0124]

Claims

1. It is characterized by: The steps include: Step S100: constructing a base classifier based on parameterized quantum circuits; Step S200: Based on the improved quantum averaging method, multiple base classifiers are fused.

2. The multi-classifier fusion method based on quantum convolutional neural network according to claim 1 is characterized in that: The steps of constructing a base classifier based on parameterized quantum circuits are: Step S101: constructing a quantum state preparation circuit; Step S102: constructing a quantum convolutional neural network; The quantum convolutional neural network constructed based on parameterized quantum circuits is the base classifier based on parameterized quantum circuits.

3. The multi-classifier fusion method based on quantum convolutional neural network according to claim 1 is characterized in that: The steps of constructing the quantum state preparation circuit are: Step S101: constructing a quantum state preparation circuit; The quantum state preparation circuit is a quantum amplitude encoding circuit; In quantum networks, input data χ is converted to a different space through quantum feature mapping, i.e., φ:χ→κ, where φ represents the mapping process and κ represents a Hilbert space; Quantum amplitude coding circuit is to convert N-dimensional data x=(x1, x2,…, x N ) is encoded into an n-bit quantum state, expressed as follows: Where U(x) represents the quantum amplitude coding circuit, represents N-dimensional real number space, Represents the encoded quantum state, N = 2 n , |i> represents the i-th ground state.

4. The multi-classifier fusion method based on quantum convolutional neural network according to claim 1 is characterized in that: The steps of constructing a quantum convolutional neural network are: Step S102-1: constructing a quantum convolution filter; The quantum convolution filter includes two qubits, and the qubits include a combination of a single qubit gate and a double qubit gate circuit containing parameters; The parameterized quantum circuits of the quantum convolution filters in the same convolution layer are the same, and the quantum convolution filters are equivalent to the convolution kernels in CNN. On the basis of the quantum convolution filter, #1 quantum convolution filter, #2 quantum convolution filter and #3 quantum convolution filter are constructed; the input of the quantum convolution filter is the data converted into quantum form; #1 The structure of the quantum convolution filter is: The data converted into quantum form is input into two quantum bits |0> and |1>, and R x (θ x ), R z (θ Z ) and the controlled NOT gate CNOT act on |0> and |1> to form a basic unit u1; four u1s are connected in series to form a 1# quantum convolution filter; R x (θ x ) is the x revolving door, R z (θ Z ) is the y revolving door; θ x The initial value of is a random value; Z The initial value of is a random value; #2 The structure of the quantum convolution filter is: The data converted into quantum form is input into the two quantum bits |0> and |1>, and the H gate, the controlled Z gate and the R x (θ x ) acts on |0> and |1> to form a basic unit u2; four u2s are connected in series to form a 2# quantum convolution filter; the H gate is a Hadamard gate; #3 The structure of the quantum convolution filter is: 3 u1s and 1 u2 are connected in series to form a quantum convolution filter; Step S102-2: constructing a quantum pooling structure; The quantum pooling structure is a two-qubit gate circuit. The two-qubit gate circuit consists of a CNOT gate and does not contain free parameters. p express; Step S102-3: constructing a quantum convolutional neural network; The convolutional neural network based on parameterized quantum circuit includes a quantum state preparation circuit, a first convolution layer, a first pooling layer, a second convolution layer and a second pooling layer, and the quantum pooling circuit of the pooling layer has the same structure and parameters; The first convolutional layer contains four identical quantum convolutional filters; The first pooling layer contains two U p ; The second convolutional layer contains a quantum convolution filter; The second pooling layer contains a U p ; The quantum convolutional neural network based on #1 quantum convolution filter is BC-QCNN1; The quantum convolutional neural network based on #2 quantum convolution filter is BC-QCNN2; The quantum convolutional neural network based on #3 quantum convolution filter is BC-QCNN3.

5. The multi-classifier fusion method based on quantum convolutional neural network according to claim 1, characterized in that: The steps of fusing multiple base classifiers based on the improved quantum averaging method are as follows: Step S201: construct the evidence form corresponding to the base classifier; For the binary classification task, assuming that the recognition framework is Φ = {H1, H2}, the result obtained by the base classifier is p = [p1, p2], that is, p(H1) = p1, p(H2) = p2, which is converted into the basic probability distribution function m(A) through formula (2): Where α is the classification accuracy of the base classifier, which is used as a discount factor here; A is the proposition; the evidence corresponding to the classification results of the three base classifiers is obtained through formula (2), m1 is the evidence form corresponding to the classification result output by BC-QCNN1, m2 is the evidence form corresponding to the classification result output by BC-QCNN2, and m3 is the evidence form corresponding to the classification result output by BC-QCNN3; For binary classification tasks, the output of the base classifier is a two-dimensional vector in the form of probability, so it is first necessary to convert it into the form of the basic probability distribution function in evidence theory; Step S202: fusing multiple base classifiers based on an improved quantum averaging method; The optimal phase parameters were obtained using the improved quantum averaging method, and the quantum model of the evidence was established; Based on the evidence forms m1, m2 and m3 obtained by the base classifier, the quantum model |φ1> of m1 is established, the quantum model |φ2> of m2 is established, and the quantum model |φ3> of m3 is established. |φ1>, |φ2> and |φ3> have different constraints respectively. According to the improved quantum average combination method, |φ1>, |φ2> and |φ3> are fused, and the quantum model of the fused evidence is obtained as follows: After quantum measurement operation, the quantum state Collapse into different categories with corresponding probabilities, namely And thus determine the category to which the sample belongs.

Images