Optimization theory and method of quantum neural network coding and evolutionary strategy
Through the initial coding optimization theory and dynamic adjustment strategy of quantum neural network, the convergence speed of quantum neural network algorithms is solved and the problem of local optimality is easily trapped, and a faster and more efficient quantum neural network algorithm is realized, which improves global search capabilities and convergence. speed.
Patent Information
- Application Number
- CN202410971246.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-19
- Publication Date
- 2024-11-12
AI Technical Summary
Quantum neural network algorithms have problems such as slow convergence speed, easy to fall into local optimality, and lack of theoretical support.
The traversal theorem of the initial coding optimization theory of quantum neural network was established, and a quantum neural network optimization algorithm based on real-value input and output coding and dynamic adjustment strategies was proposed. Through mathematics, the optimization of its time and space complexity was proved to achieve faster and more efficient Quantum neural network algorithm.
The quantum neural network structure is effectively optimized, the network convergence speed is improved, the local optimal trap is avoided, and the global search capability is provided and the faster convergence speed is provided.
Smart Images

Figure CN118940848A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of quantum neural network algorithms, and in particular to an optimization theory and method for quantum neural network coding and evolutionary strategies. Background Art
[0002] As a transformative and disruptive emerging technology, artificial intelligence technology is bringing revolutionary changes to human civilization in all aspects. Whether it is speech, image recognition, intelligent control, or autonomous decision-making, mathematical modeling and solution of related problems are carried out. However, with the continuous expansion of the system scale and the increase of data dimensions, the distribution of limited-scale data in space will gradually become sparse, thus losing statistical significance. This requires obtaining the statistical characteristics of the data set based on a very large data scale, but this will bring a serious burden on computing resources. After years of development, the field of artificial intelligence has proposed many methods and tools for dealing with high-dimensional problems. As one of the methods of machine learning, the neural network method has alleviated the difficulties caused by the dimensionality disaster to a certain extent.
[0003] A neural network is composed of a large number of nodes (neurons) and the connections between them. Each node contains a specific output function, called an activation function. The connection between each two nodes represents the weighted value of the signal passing through the connection, called a weight. If a neural network has more than two layers, it is usually called a deep neural network.
[0004] As the amount of data increases, traditional machine learning methods perform poorly, while deep learning's excellent performance in target recognition has quickly gained attention from industry and academia. As research deepens, scholars from various countries have gradually solved the classification, nonlinear, and XOR problems of neural network models, but they still cannot effectively solve the problems of large computational complexity, long network training time, and gradient vanishing, causing the network to fall into local extremes.
[0005] Neural networks have the problem of large amount of calculation and easy to fall into local minimum, and the advantages of quantum computing can make up for the shortcomings of neural networks. Quantum artificial intelligence is an attempt to introduce quantum computing theory into the field of artificial intelligence. It has the following advantages: first, it can represent a large solution space with a small number of individuals; second, it has a strong global search capability; third, it has a fast convergence speed and can search for the optimal solution in a short time. Summary of the invention
[0006] The purpose of the present invention is to provide an optimization theory and method for quantum neural network coding and evolutionary strategy. Aiming at the problems of slow convergence speed, easy to fall into local optimum, and lack of theoretical support in quantum neural network algorithm, the present invention establishes the ergodic theorem of quantum neural network initial coding optimization theory, clarifies the mapping relationship between time complexity and space complexity, and proposes a quantum neural network optimization algorithm based on real-valued input and output coding and dynamic adjustment strategy, thus obtaining a faster and more efficient quantum neural network algorithm.
[0007] The present invention firstly uses the theory of mathematical computational complexity to deduce the time and space complexity of quantum neural network algorithms and traditional neural network algorithms, and gives the mechanism of the mutual influence of time and space complexity on the basis of this theory; then, the superiority of real-valued input and output coding and the universality of normalized coding are proved through rigorous mathematics; finally, the feasibility and rapidity of the proposed dynamic adjustment strategy are proved through the derivation of mathematical formulas.
[0008] A. Initial Coding Optimization
[0009] (1) Real-valued coding input and output
[0010] In order to realize the quantum state description of the training samples, the training samples described by real-valued vectors in n-dimensional Euclidean space are Define the following conversion formula:
[0011] |X>=[|x 1 >,|x 2 >, ..., |x n >] T
[0012] in,
[0013]
[0014] is the training sample, |X> is the state after the training sample is transformed; after evolution, the output is the probability amplitude of the state |1> in the quantum bit superposition state;
[0015] (2) Using the phase of the quantum bit as the initial encoding
[0016] The coding entity is composed of qubits, and the phase of the qubit is directly used as the initial coding. The coding scheme is as follows:
[0017] p i =[θ i1 ,θ i2 ,…,θ ij ,…,θ in ]
[0018] Among them, p irepresents the i-th initial code, θ ij =2π×rand, rand is a random number between (0,1), j=1,2,…,n, n represents the number of quantum bits;
[0019] (3) Proof of time and space complexity of initial coding optimization
[0020] Using asymptotic analysis methods, we clarify the factors affecting the algorithm's space complexity; using the parallel structure of the quantum neural network algorithm, we effectively reduce the number of recursions of the algorithm and prove that the space complexity function is different from the traditional quantum neural network algorithm space complexity S(n) = O(n t ) is reduced to S(n)=O(n k ), k <t;
[0021] By using the main theorem and the divide-and-conquer method, and drawing on the idea of Chebyshev's inequality, we establish the valuation inequality in the quantum neural network algorithm, give the upper and lower bounds of the time complexity function in the quantum neural network algorithm, explore the internal mechanism that affects the time complexity of the algorithm, and prove that the time complexity function is different from the traditional quantum neural network algorithm time complexity T(n) = O(n t ) is reduced to T(n)=O(n k );
[0022] B. Intermediate Evolutionary Strategy Optimization
[0023] (1) Update of the rotation angle θ of the quantum rotating gate
[0024] By rotating the i Change to adaptive changes, that is, dynamically adjust the strategy and adopt the optimal retention mechanism to ensure the global convergence of the algorithm;
[0025] (2) Update of quantum NOT gate
[0026] All individuals in the population are reordered. For all individuals after ordering, the k-bit quantum NOT gate of the k-th individual is taken as the k-bit code change of the new individual, and the (k+1)-bit quantum NOT gate of the (k+1)-th individual is taken as the (k+1)-bit code change of the new individual. The following codes are taken in sequence until the new individual has the same code number as the original individual to determine the new individual. All individuals are determined by iterative method and cyclic method.
[0027] (3) Convergence proof of intermediate evolution strategy
[0028] The Schrödinger equation is used to preliminarily model the optimization problem of evolutionary strategies. The assumption of the Delta potential well and the concept of multi-scale are used to gradually reduce the quantum perturbation in the evolutionary strategy and approximate the ground state energy of the optimization system. Finally, the tunneling effect of quantum mechanics is used to more quickly pass through the potential barriers next to the local extreme points when the evolutionary strategy is looking for the global minimum, thereby finding the optimal or suboptimal solution.
[0029] Beneficial Effects
[0030] The above technical solution of the present invention has the following beneficial effects: In view of the problems existing in quantum neural network algorithms, such as slow convergence speed, easy falling into local optimum, and lack of theoretical support, the present invention constructs an improved method for quantum neural network algorithms, which can effectively optimize the quantum neural network structure and accelerate network convergence. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] Figure 1 is a flow chart of the present invention; DETAILED DESCRIPTION
[0032] In order to make the technical solution, advantages and purposes of the present invention more clear, the technical solution of the present invention is further described.
[0033] Spacecraft on orbit are affected by inertial devices, refrigeration mechanisms, flexible components, etc., which will produce micro-vibrations of different frequencies and amplitudes, which will have an adverse effect on the high-precision control of the spacecraft's attitude and the execution of its mission. The existing neural network intelligent control method cannot solve the problem of approximation accuracy under high-bandwidth conditions. Therefore, combining the optimization of quantum neural network structure with the optimization design of the controller provides a new solution to the problem of high-bandwidth vibration suppression. The specific implementation plan is as follows:
[0034] A. Initial Coding Optimization
[0035] (1) Real-valued coding input and output
[0036] In order to realize the quantum state description of the training samples, the training samples described by real-valued vectors in n-dimensional Euclidean space are Define the following conversion formula:
[0037] |X>=[|x 1 >,|x 2 >, ..., |x m >] T
[0038] in,
[0039]
[0040] is the training sample, |X> is the state after the training sample is transformed; after evolution, the output is the probability amplitude of the state |1> in the quantum bit superposition state;
[0041] (2) Using the phase of the quantum bit as the initial encoding
[0042] The coding entity is composed of qubits, and the phase of the qubit is directly used as the initial coding. The coding scheme is as follows:
[0043] p i =[θ i1 ,θ i2 ,…,θ ij ,…,θ in ]
[0044] Among them, p i represents the i-th initial code, θ ij =2π×rand, rand is a random number between (0,1), j=1,2,…,n, n represents the number of quantum bits;
[0045] (3) Proof of time and space complexity of initial coding optimization
[0046] Using asymptotic analysis methods, we clarify the factors affecting the algorithm's space complexity; using the parallel structure of the quantum neural network algorithm, we effectively reduce the number of recursions of the algorithm and prove that the space complexity function is different from the traditional quantum neural network algorithm space complexity S(n) = O(n t ) is reduced to S(n)=O(n k ), k <t;
[0047] By using the main theorem and the divide-and-conquer method, and drawing on the idea of Chebyshev's inequality, we establish the valuation inequality in the quantum neural network algorithm, give the upper and lower bounds of the time complexity function in the quantum neural network algorithm, explore the internal mechanism that affects the time complexity of the algorithm, and prove that the time complexity function is different from the traditional quantum neural network algorithm time complexity T(n) = O(n t ) is reduced to T(n)=O(n k );
[0048] B. Intermediate Evolutionary Strategy Optimization
[0049] (1) Update of the rotation angle θ of the quantum rotating gate
[0050] By rotating the i Change to adaptive changes, that is, dynamically adjust the strategy and adopt the optimal retention mechanism to ensure the global convergence of the algorithm;
[0051] (2) Update of quantum NOT gate
[0052] All individuals in the population are reordered. For all individuals after ordering, the k-bit quantum NOT gate of the k-th individual is taken as the k-bit code change of the new individual, and the (k+1)-bit quantum NOT gate of the (k+1)-th individual is taken as the (k+1)-bit code change of the new individual. The following codes are taken in sequence until the new individual has the same code number as the original individual to determine the new individual. All individuals are determined by iterative method and cyclic method.
[0053] (3) Convergence proof of intermediate evolution strategy
[0054] The Schrödinger equation is used to preliminarily model the optimization problem of evolutionary strategies. The assumption of the Delta potential well and the concept of multi-scale are used to gradually reduce the quantum perturbation in the evolutionary strategy and approximate the ground state energy of the optimization system. Finally, the tunneling effect of quantum mechanics is used to more quickly pass through the potential barriers next to the local extreme points when the evolutionary strategy is looking for the global minimum, thereby finding the optimal or suboptimal solution.
[0055] The present invention realizes the quantum neural network structure and parameter optimization design, solves the approximation accuracy and speed problems of traditional neural networks, effectively improves the robustness and control accuracy of the spacecraft attitude control system, and is of great significance for solving the problem of high-precision attitude control of spacecraft in the future.
[0056] The contents not described in detail in the specification of the present invention belong to the prior art known to the professional and technical personnel in this field.
Claims
1. A quantum neural network coding and evolution strategy optimization theory and method, characterized by: A. Initial Coding Optimization (1) Real-valued coding input and output In order to realize the quantum state description of the training samples, the training samples described by real-valued vectors in n-dimensional Euclidean space are Define the following conversion formula: |X〉=[|x1>,|x2>,...,|x n 〉] T in, is the training sample, |X> is the state of the training sample after transformation; after evolution, the output is the probability amplitude of the state |1> in the quantum bit superposition state; (2) Using the phase of the quantum bit as the initial encoding The coding entity is composed of qubits, and the phase of the qubit is directly used as the initial coding. The coding scheme is as follows: p i =[θ i1 ,i i2 ,…,θ ij ,…,θ in ] Among them, pi represents the i-th initial code, θ ij =2π×rand, rand is a random number between (0,1), j=1,2,…,n, n represents the number of quantum bits; (3) Proof of time and space complexity of initial coding optimization Using asymptotic analysis methods, we clarify the factors affecting the algorithm's space complexity; using the parallel structure of the quantum neural network algorithm, we effectively reduce the number of recursions of the algorithm and prove that the space complexity function is different from the traditional quantum neural network algorithm space complexity S(n) = O(n t ) is reduced to S(n)=O(n k ), k <t; By using the main theorem and the divide-and-conquer method, and drawing on the idea of Chebyshev's inequality, we establish the valuation inequality in the quantum neural network algorithm, give the upper and lower bounds of the time complexity function in the quantum neural network algorithm, explore the internal mechanism that affects the time complexity of the algorithm, and prove that the time complexity function is different from the traditional quantum neural network algorithm time complexity T(n) = O(n t ) is reduced to T(n)=O(n k ); B. Intermediate Evolutionary Strategy Optimization (1) Update of the rotation angle θ of the quantum rotating gate By rotating the i Change to adaptive changes, that is, dynamically adjust the strategy and adopt the optimal retention mechanism to ensure the global convergence of the algorithm; (2) Update of quantum NOT gate All individuals in the population are reordered. For all individuals after ordering, the k-bit quantum NOT gate of the k-th individual is taken as the k-bit code change of the new individual, and the (k+1)-bit quantum NOT gate of the (k+1)-th individual is taken as the (k+1)-bit code change of the new individual. The following codes are taken in sequence until the new individual has the same code number as the original individual to determine the new individual. All individuals are determined by iterative method and cyclic method. (3) Convergence proof of intermediate evolution strategy The Schrödinger equation is used to preliminarily model the optimization problem of evolutionary strategies. The assumption of the Delta potential well and the concept of multi-scale are used to gradually reduce the quantum perturbation in the evolutionary strategy and approximate the ground state energy of the optimization system. Finally, the tunneling effect of quantum mechanics is used to more quickly pass through the potential barriers next to the local extreme points when the evolutionary strategy is looking for the global minimum, thereby finding the optimal or suboptimal solution.
Citation Information
Cited By
Optimizer design method and device using quantum tunneling effect, and medium
CN119783840A
