Coherent Isin machine, model parameter optimization method and device thereof, medium and product

Abstract

The invention discloses a coherent Isin machine and a model parameter optimization method and device thereof, a medium and a product. The method comprises the following steps: constructing an initial training population; the initial training population comprises a plurality of quantum neural network QNN instances configured with different parameter combinations; based on an evolutionary optimization algorithm, performing iterative updating on the initial training population to obtain a corresponding target training population when a convergence condition is reached; and determining an optimal QNN instance in the target training population, and obtaining a parameter combination of the optimal QNN instance. The embodiment of the invention provides a solution combining a population evolutionary optimization algorithm and the characteristics of a coherent Isin machine aiming at the problems of discontinuity and non-differentiability existing in the quantum state of the coherent Isin machine. The method comprises the following steps of: iteratively updating parameters of discrete QNN instances in a training population based on the evolutionary optimization algorithm; and meanwhile, the optimal parameter combination is efficiently solved by utilizing the parallel computing advantage of the coherent Isin machine, so that the model training efficiency is improved.

Description

Technical Field

[0001] The present application relates to the field of quantum neural networks, and in particular to a coherent Ising machine and its model parameter optimization method, device, medium and product. Background Art

[0002] The Coherent Ising Machine (CIM) is a quantum simulator based on the principles of quantum optics. It simulates quantum spin systems through a network of degenerate optical parametric oscillators (DOPOs). By leveraging the superposition and coherence of quantum states, it explores massive candidate solutions in parallel, enabling efficient solutions to combinatorial optimization problems.

[0003] Currently, machine learning models are typically trained using techniques such as gradient descent algorithms. Gradient descent algorithms calculate gradients based on the continuity of the model parameter space and the differentiability of the loss function, and then iteratively optimize the model parameters based on these gradients. However, quantum states are inherently discrete, and quantum state transitions are discontinuous. Furthermore, the probabilistic collapse of quantum states prevents state transitions from being described using continuously differentiable functions. Therefore, due to the discontinuity and non-differentiability of quantum states, coherent Ising machines have not been used in related techniques for model parameter optimization. Summary of the Invention

[0004] In view of this, the embodiments of the present application provide a coherent Ising machine and its model parameter optimization method, device, medium and product, aiming to achieve model parameter optimization using the coherent Ising machine and improve model training efficiency.

[0005] The technical solution of the embodiment of the present application is implemented as follows: In a first aspect, an embodiment of the present application provides a model parameter optimization method applied to a coherent Ising machine, the method comprising: Constructing an initial training population; the initial training population includes multiple quantum neural network (QNN) instances configured with different parameter combinations; Iteratively updating the initial training population based on an Evolutionary Optimization Algorithms (EOAs) algorithm to obtain a target training population corresponding to when a convergence condition is met; An optimal QNN instance in the target training population is determined, and a parameter combination of the optimal QNN instance is obtained.

[0006] In the above solution, the initial training population is iteratively updated based on the evolutionary optimization algorithm to obtain the target training population corresponding to the convergence condition, including: Calculate the fitness results of each QNN instance in the current training population and determine whether the convergence conditions are met; If not, based on the fitness results of each QNN instance in the training population, the training population is updated using an evolutionary optimization algorithm, and the step of calculating the fitness results of each QNN instance in the current training population is returned to; If so, the current training population is determined to be the target training population; Determining the optimal QNN instance in the target training population includes: Determine the QNN instance with the highest fitness result in the target training population as the optimal QNN instance.

[0007] In the above scheme, the fitness results of each QNN instance in the current training population are calculated, including: Obtain the fitness evaluation model and the objective function model of the training task; After the training samples are converted into quantum states, they are input into each QNN instance in the current training population to obtain the prediction results output by each QNN instance; Input the prediction results output by each QNN instance into the objective function model to obtain the objective function results of each QNN instance; The objective function result of each QNN instance is input into the fitness evaluation model to obtain the fitness result of each QNN instance.

[0008] In the above solution, the method further includes: Based on a self-play strategy, determining a competing QNN instance that participates in a competitive task from the training population; Executing the competition task to obtain a fitness adjustment value of the competing QNN instance; The objective function result of each QNN instance is input into the fitness evaluation model to obtain the fitness result of each QNN instance, including: Inputting the objective function result of the QNN instance into the fitness evaluation model to obtain an initial result output by the fitness evaluation model; If the QNN instance is the competing QNN instance, obtaining a fitness result of the QNN instance based on the initial result of the QNN instance and the fitness adjustment amount; If the QNN instance is not the competing QNN instance, determining the initial result of the QNN instance as the fitness result of the QNN instance.

[0009] In the above solution, executing the competition task to obtain the fitness adjustment value of the competing QNN instance includes: Obtain a competitive payoff function model; Calculating a task performance indicator of the competing QNN instance based on the prediction result output by the competing QNN instance; Calculating similarities between the competing QNN instances based on the quantum states of the competing QNN instances; The task performance index and the corresponding similarity of the competing QNN instance are input into the competition benefit function model to obtain the fitness adjustment amount of the competing QNN instance.

[0010] In the above scheme, the construction of the initial training population includes: Build a set number of QNN instances; Based on the set initialization rules, different parameter combinations are assigned to each of the QNN instances, and after the quantum state of each of the QNN instances is generated, an initial training population is obtained; The setting initialization rule includes: a classic random initialization rule or a uniform distribution rule.

[0011] In the above solution, the updating of the training population using an evolutionary optimization algorithm based on the fitness results of each QNN instance in the training population includes at least one of the following: Deleting the QNN instances in the training population whose fitness results are lower than a set fitness threshold; Performing a crossover operation on two target QNN instances in the training population to construct a child QNN instance, and adding the child QNN instance to the training population; Adding a perturbation to the parameters of the target QNN instance in the training population to construct a variant QNN instance, and adding the variant QNN instance to the training population; The target QNN instance includes a QNN instance whose fitness result is higher than or equal to a set fitness threshold.

[0012] In a second aspect, an embodiment of the present application provides a model parameter optimization device applied to a coherent Ising machine, the device comprising: A construction module is used to construct an initial training population; the initial training population includes multiple quantum neural network QNN instances configured with different parameter combinations; An iterative update module, configured to iteratively update the initial training population based on an evolutionary optimization algorithm to obtain a target training population corresponding to when a convergence condition is met; A determination module is used to determine the optimal QNN instance in the target training population and obtain a parameter combination of the optimal QNN instance.

[0013] In a third aspect, an embodiment of the present application provides a coherent Ising machine, comprising: a processor and a memory for storing a computer program that can be run on the processor, wherein the processor, when running the computer program, executes the steps of the method described in the first aspect.

[0014] In a fourth aspect, an embodiment of the present application provides a storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the method described in the first aspect are implemented.

[0015] In a fifth aspect, an embodiment of the present application provides a computer program product, comprising a computer program, which, when executed by a processor, implements the steps of the method described in the first aspect.

[0016] The technical solution provided in the embodiments of the present application is applied to a coherent Ising machine; an initial training population is constructed; the initial training population includes multiple QNN instances configured with different parameter combinations; based on an evolutionary optimization algorithm, the initial training population is iteratively updated to obtain a target training population corresponding to when convergence conditions are met; the optimal QNN instance in the target training population is determined, and the parameter combination of the optimal QNN instance is obtained. Thus, the embodiments of the present application address the discontinuity and non-differentiability issues of the quantum state of the coherent Ising machine by proposing a solution that combines a population evolutionary optimization algorithm with the characteristics of the coherent Ising machine: by iteratively updating the parameters of the discrete QNN instances in the training population based on the evolutionary optimization algorithm, while utilizing the parallel computing advantages of the coherent Ising machine to efficiently solve for the optimal parameter combination, the efficiency of model training is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 Schematic diagram of the flow of the model training optimization method according to the embodiment of the present application; Figure 2 Schematic diagram of the steps of the evolutionary optimization algorithm; Figure 3 This is a schematic diagram of the structure of the model training optimization device according to an embodiment of the present application; Figure 4 This is a schematic structural diagram of the coherent Ising machine model according to an embodiment of the present application. DETAILED DESCRIPTION

[0018] The present application will be described in further detail below with reference to the accompanying drawings and embodiments.

[0019] In the following description, reference is made to “some embodiments”, which describes a subset of all possible embodiments, but it will be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those explained below or as commonly understood by those skilled in the art to which this application pertains. The terms used herein in the specification of this application are for the purpose of describing specific embodiments only and are not intended to limit this application.

[0021] The present application embodiment provides a model parameter optimization method applied to a coherent Ising machine, such as Figure 1 As shown, the method includes: Step 101: construct an initial training population.

[0022] The initial training population includes multiple QNN instances configured with different parameter combinations.

[0023] Step 102: Based on the evolutionary optimization algorithm, the initial training population is iteratively updated to obtain the target training population corresponding to the convergence condition.

[0024] Step 103: determine the optimal QNN instance in the target training population and obtain the parameter combination of the optimal QNN instance.

[0025] Here, the QNN instance is a model that integrates quantum computing and neural networks, using quantum bits instead of classical bits in traditional neural networks as basic information units. The core is to use quantum operations such as quantum superposition, entanglement, and quantum gates to construct quantum neurons and network structures; the QNN instance can be applied to coherent Ising machines.

[0026] Here, the coherent Ising machine is a quantum simulator based on the principles of quantum optics. It simulates quantum spin systems through a degenerate optical parametric oscillator network, and uses the superposition and coherence of quantum states to explore massive candidate solutions in parallel. Specifically, in the coherent Ising machine, the coupling between spins is globally parallel and interactive. The state update of each spin is directly affected by all other spins. There is no need to wait for serial steps, and a global state update can be completed in microseconds to milliseconds. Therefore, the coherent Ising machine is often used to efficiently solve combinatorial optimization problems.

[0027] It should be noted that although the parallel search capability of the coherent Ising machine has a significant improvement in computing power compared to the serial search capability of classical computers, in related technologies, the coherent Ising machine is generally not used in the field of machine learning model training. The specific reason is: currently, classical computers usually use the gradient descent algorithm for model training. During the training process, it is necessary to define a loss function to measure the gap between the predicted value and the true value. The trend and amplitude of parameter optimization adjustment are determined by calculating the gradient of the loss function with respect to the model parameters. That is, in the gradient descent algorithm, the loss function is required to be continuous and differentiable, and the model parameters need to search for the optimal solution in a continuous parameter space according to the loss function; while the quantum state of the coherent Ising machine can only take discrete values ​​such as ±1 of the Ising spin and the integer state of the number of photons, and cannot take values ​​of the classical model parameters within a continuous parameter range. Moreover, based on the discreteness of the quantum state, the state change of the quantum state cannot be represented by a continuous differentiable function. Therefore, due to the discontinuity and non-differentiability of the quantum state of the coherent Ising machine, the coherent Ising machine cannot be applied to classical optimization algorithms such as gradient descent.

[0028] It should be noted that, in response to the discontinuity and non-differentiability problems of the quantum state of the coherent Ising machine, the embodiment of the present application provides a technical solution that combines the evolutionary optimization algorithm of the population with the characteristics of the coherent Ising machine, so that the coherent Ising machine can be used to optimize the parameters of the deep learning model and utilize the parallel search capability of the coherent Ising machine to improve the model training efficiency.

[0029] Here, the evolutionary optimization algorithm is a type of random optimization method that simulates biological evolution theory. The core idea is to simulate the biological evolution process of "natural selection and genetic variation" to find the optimal solution. Specifically, Figure 2 As shown in the figure, the basic process of the evolutionary optimization algorithm usually includes the steps of initializing the population - evaluating the fitness - selecting the operation - genetic operation - iterative updating - obtaining the optimal solution. In the initializing the population step, a population including multiple individuals is initialized and generated, and each individual represents a random candidate solution; in the fitness evaluation step, the fitness of each individual is calculated to evaluate the quality of the candidate solution for the processing task; in the selection operation and genetic operation steps, by simulating the biological evolution process of "natural selection and genetic variation", individuals with low fitness are eliminated from the population, and new individuals are generated and added to the population to maintain the diversity of candidate solutions in the population; by repeatedly executing the above-mentioned fitness evaluation, selection operation and genetic operation steps, the population is continuously iterated and updated, and finally the optimal solution is obtained after the population converges.

[0030] Here, when using the evolutionary optimization algorithm to optimize model parameters, various candidate model parameter combinations are used as individuals in the population, and the optimal individual obtained after the population converges corresponds to the final optimized model parameter combination.

[0031] It can be understood that the evolutionary optimization algorithm is based on the iterative search for the optimal solution in the population. It only needs to evaluate the fitness of individuals in the population. The fitness of individuals are all discrete quantities, and there is no need to calculate the gradient or rely on the differentiability of the function. The evolutionary optimization algorithm searches for the optimal solution in the discrete solution space. The population iteration process is essentially to delete or add discrete solutions to the discrete solution space. The process of searching for the optimal solution does not involve optimizing the solution in the continuous solution space. Therefore, even if the coherent Ising machine quantum state has characteristics such as discontinuity and non-differentiability that conflict with the gradient descent algorithm, in the embodiment of the present application, the coherent Ising machine uses the evolutionary optimization algorithm to quickly calculate the QNN instance corresponding to the optimal model parameter combination, thereby achieving the technical effect of model parameter optimization.

[0032] In addition, evolutionary optimization algorithms achieve global optimization through population diversity, and coherent Ising machines can simultaneously process the evaluation tasks of all individuals in the population. Compared with using classical computers to execute evolutionary optimization algorithms, using coherent Ising machines to execute evolutionary optimization algorithms can not only circumvent the limitations of discontinuity and non-differentiability of the quantum state of the coherent Ising machine, but also greatly improve the iterative efficiency of the evolutionary optimization algorithm.

[0033] The following describes in detail the model parameter optimization method applied to the coherent Ising machine in an embodiment of the present application.

[0034] Exemplarily, constructing an initial training population includes: constructing a set number of QNN instances; assigning different parameter combinations to each QNN instance based on a set initialization rule, generating a quantum state of each QNN instance, and then obtaining the initial training population.

[0035] Here, each QNN instance serves as an individual in the training population. In the embodiment of the present application, the structure and parameters of the QNN instance are encoded through quantum states.

[0036] Here, the structure of a QNN instance refers to the connection method of quantum neurons, the type and arrangement order of quantum gates, and is used to determine how the QNN instance processes the input quantum state; in some embodiments, the QNN instances in the training population have the same structure.

[0037] Here, the parameters of the QNN instance are adjustable variables, which are equivalent to the weights in the classic neural network model. That is, the process of optimizing the model parameters in the embodiment of the present application is essentially the process of searching for the optimal QNN instance parameter combination in the training population.

[0038] Here, since the parameter combination of the optimal QNN instance determined in the embodiment of the present application is ultimately applied to the machine learning model for performing the actual task, the structure of the QNN instance is determined based on the machine learning model used to perform the actual task.

[0039] Here, after the QNN instance is generated, different parameter combinations are assigned to each QNN instance in the initial training population based on the set initialization rules. In order to ensure the individual diversity of the initial training population, the set initialization rules include but are not limited to: classic random initialization rules or uniform distribution rules.

[0040] Here, the classical random initialization rule is a distribution rule based on a random number generator that randomly selects values ​​within a preset parameter range; the uniform distribution rule is a distribution rule that generates parameters based on the equal probability of selecting each value within the parameter range; it should be noted that the embodiment of the present application does not specifically limit the training population initialization method.

[0041] Here, after assigning different parameter combinations to each QNN instance, a training population consisting of a set number of QNN instances is obtained; wherein the i-th QNN instance in the training population consists of the quantum state It is expressed as follows:

[0042] in, is the weight coefficient, As the basic state.

[0043] Exemplarily, based on the evolutionary optimization algorithm, the initial training population is iteratively updated to obtain the target training population corresponding to when the convergence condition is met, including: calculating the fitness results of each QNN instance in the current training population, and judging whether the convergence condition is met; if not, based on the fitness results of each QNN instance in the training population, the evolutionary optimization algorithm is used to update the training population, and the step of calculating the fitness results of each QNN instance in the current training population is returned; if so, the current training population is determined to be the target training population.

[0044] Here, the convergence condition is a pre-set condition for stopping iteration. After each iterative update of the training population based on the evolutionary optimization algorithm, it will be determined whether the convergence condition is met. If the convergence condition is met, the iteration is stopped and the current training population is used as the target training population after convergence; if the convergence condition is not met, the next training population update is performed.

[0045] In some embodiments, the convergence condition includes but is not limited to: reaching a set number of iterations, or the change amplitude of the fitness result is lower than a set amplitude threshold.

[0046] Here, the fitness result represents the degree of adaptation of the individual (i.e., QNN instance) to this training task. The higher the fitness result of the QNN instance, the closer the parameter combination of the QNN instance is to the optimal parameter combination.

[0047] Here, updating the training population specifically refers to updating some QNN instances in the training population; an evolutionary optimization algorithm is used to update the training population, that is, the fitness results of the QNN instances are used as a reference for updating the training population, QNN instances with higher fitness results in the training population are retained, and QNN instances with lower fitness results in the training population are eliminated, so that the training population gradually converges to the optimal parameter combination with continuous iterations.

[0048] Exemplarily, determining the optimal QNN instance in the target training population includes: determining the QNN instance with the highest fitness result in the target training population as the optimal QNN instance.

[0049] It can be understood that after reaching the convergence condition, the coherent Ising machine stops iteratively updating the training population and outputs the QNN instance and its parameter combination with the highest fitness result in the current training population. The parameter combination of the optimal QNN instance can be used as the model parameters after optimization to apply to the machine learning model that performs actual tasks.

[0050] Exemplarily, calculating the fitness result of each QNN instance in the current training population includes: obtaining a fitness evaluation model and an objective function model of the training task; after the training sample is converted into a quantum state, inputting it into each QNN instance in the current training population to obtain a prediction result output by each QNN instance; inputting the prediction result output by each QNN instance into the objective function model respectively to obtain an objective function result of each QNN instance; inputting the objective function result of each QNN instance into the fitness evaluation model to obtain a fitness result of each QNN instance.

[0051] Here, the training samples carry the corresponding true labels representing the expected values, which are used to compare with the predicted results output by the QNN instance.

[0052] It should be noted that since the training samples obtained are usually classical data (that is, data processed on a classical computer), before the training samples are input into the QNN instance as input data, the sample data needs to be subjected to quantum state conversions such as base state encoding and angle encoding to map the sample data into a quantum superposition state or coherent state.

[0053] In one example, after the training sample is input into the QNN instance, the prediction result is output As shown in the following formula:

[0054] in, is the quantum state of the i-th QNN instance, For parameters Constructed evolution operator.

[0055] Here, before executing the fitness evaluation step, it is necessary to pre-construct the objective function model of this training task. The objective function model can be determined based on the expected goal of this training task, such as classification accuracy or regression error. In one example, the expected goal of the training task is to minimize the error, and the constructed objective function model is shown in the following formula:

[0056] in, is the objective function result of the i-th QNN instance, is the prediction result output by the i-th QNN instance based on the j-th training sample, is the true label corresponding to the jth training sample, and M is the number of training samples.

[0057] Here, before executing the fitness evaluation step, it is necessary to pre-build a fitness evaluation model for this training task. The fitness evaluation model is determined based on the expected goal of this training task and supports discrete value input. In one example, the fitness evaluation model is shown as follows:

[0058] in, is the initial result representing the fitness of the i-th QNN instance. In some embodiments, the initial result is the final fitness result of the i-th QNN instance; N is the number of QNN instances in the training population.

[0059] It can be understood that the input and output quantities in the above fitness evaluation step are both discrete values, so they can be executed on the coherent Ising machine, and the fitness results of the corresponding evolutionary optimization algorithm based on the coherent Ising machine can be evaluated, which can greatly improve the computing efficiency.

[0060] Exemplarily, based on the fitness results of each QNN instance in the training population, an evolutionary optimization algorithm is used to update the training population, including at least one of the following: deleting QNN instances in the training population whose fitness results are lower than a set fitness threshold; performing a crossover operation on two target QNN instances in the training population to construct a child QNN instance, and adding the child QNN instance to the training population; adding a perturbation amount to the parameters of the target QNN instance in the training population to construct a variant QNN instance, and adding the variant QNN instance to the training population; wherein the target QNN instance includes a QNN instance whose fitness result is higher than or equal to the set fitness threshold.

[0061] It can be understood that deleting QNN instances in the training population whose fitness results are less than the set fitness threshold is a selection (elimination) operation step of the evolutionary optimization algorithm; performing a crossover operation on two target QNN instances in the training population to construct an offspring QNN instance, and adding the offspring QNN instance to the training population is a crossover operation step of the evolutionary optimization algorithm; adding a perturbation amount to the parameters of the target QNN instance in the training population to construct a mutant QNN instance, and adding the mutant QNN instance to the training population is a genetic operation step of the evolutionary optimization algorithm; in some embodiments, based on the fitness results of each QNN instance in the training population, an evolutionary optimization algorithm is used to update the training population, including the above-mentioned selection operation step, crossover operation step and genetic operation step.

[0062] It can be understood that after obtaining the fitness results of each QNN instance in the current training population, the QNN instances in the training population are sorted in descending order based on the fitness results, and the QNN instances with fitness results lower than the set fitness threshold are deleted from the training population, that is, the QNN instances with low adaptability to this training task are eliminated, and the parameter combination of the eliminated QNN instance configuration deviates greatly from the optimal parameter combination.

[0063] Here, the set fitness threshold may be a preset fixed value, or a dynamic fitness value obtained based on the sorting results of the QNN instances and the fitness results of each QNN instance.

[0064] It can be understood that in order to add QNN instances configured with new parameter combinations to the training population so that the training population converges to the optimal parameters after continuous iteration, the embodiment of the present application randomly selects two QNN instances from the target QNN instance as parent QNN instances, performs a crossover operation on the two parent QNN instances, and adds the child QNN instances constructed by the crossover operation to the training population; wherein the parameters of the child QNN instances are determined based on the following formula:

[0065] in, are the parameters of the descendant QNN instance, and are the parameters of the parent QNN instance, is the cross-proportional coefficient.

[0066] It can be understood that in order to add QNN instances configured with new parameter combinations to the training population, so that the training population converges to the optimal parameters after continuous iteration, the embodiment of the present application randomly selects a QNN instance from the target QNN instance and adds a perturbation amount to the parameters of the QNN instance, constructs a variant QNN instance and adds the variant QNN instance to the training population; in some embodiments, the target QNN instance used to construct the variant QNN instance includes: a newly constructed offspring QNN instance and / or a target QNN instance that is not used to construct the offspring QNN instance.

[0067] Here, the parameters of the mutated QNN instance obtained based on the newly constructed descendant QNN instance can be determined based on the following formula:

[0068] in, are the parameters of the mutated QNN instance, is the disturbance amount, the disturbance amount Normally distributed.

[0069] In some embodiments, based on the fitness results of each QNN instance in the training population, an updated training population is obtained by deleting, crossing, and adding disturbances to some QNN instances in the training population, wherein the updated training population includes a set number of QNN instances, and each QNN instance is configured with a different parameter combination.

[0070] It can be understood that after the QNN instances whose fitness results are lower than the set fitness threshold are deleted from the training population, in order to ensure the individual diversity of the training population, the embodiment of the present application will also construct new QNN instances through crossover operations and / or mutation operations and add them to the deleted training population, so that after each iteration, the number of QNN instances in the training population is maintained at the set number set in the population initialization step.

[0071] In some embodiments, the method further includes: determining a competing QNN instance participating in a competitive task from a training population based on a self-game strategy; and executing the competitive task to obtain a fitness adjustment value of the competing QNN instance.

[0072] Accordingly, the objective function result of each QNN instance is input into the fitness evaluation model to obtain the fitness result of each QNN instance, including: inputting the objective function result of the QNN instance into the fitness evaluation model to obtain the initial result output by the fitness evaluation model; if the QNN instance is a competing QNN instance, then based on the initial result of the QNN instance and the fitness adjustment amount, the fitness result of the QNN instance is obtained; if the QNN instance is not a competing QNN instance, then determining the initial result of the QNN instance as the fitness result of the QNN instance.

[0073] It should be noted that, in order to further improve the model parameter optimization effect, under the framework of coherent Ising machine + evolutionary optimization algorithm, the embodiment of the present application also introduces a self-play strategy to improve the generalization ability.

[0074] Here, the self-game strategy, also known as the self-game mechanism, is an iterative learning paradigm. By allowing individuals in the population to act as opponents and compete and interact in specific tasks through random pairing, disadvantaged individuals in the competition need to optimize their performance to avoid elimination, and advantaged individuals also need to iterate to cope with the evolution of their opponents, thereby continuously improving the overall adaptability and task performance of the population.

[0075] It can be understood that under the framework of the coherent Ising machine + evolutionary optimization algorithm + self-game strategy in the embodiment of the present application, a dual-drive mechanism of "competition + evolution" is formed, in which the evolutionary optimization algorithm realizes population iteration through selection operations, crossover operations and mutation operations, and the self-game strategy provides an accurate selection basis for population iteration, and the competition score obtained based on the self-game strategy can be used as a fitness adjustment amount to adjust the initial result representing the fitness output based on the fitness evaluation model, so as to avoid the optimal individual search from falling into local optimality and enhance the ability of the population to solve complex tasks.

[0076] Here, before updating the current training population, the embodiment of the present application determines the number of competitive tasks based on the self-game strategy and the QNN instances participating in the competitive tasks, wherein the competitive QNN instances participating in each competitive task are two QNN instances randomly selected from the training population.

[0077] Here, since only some QNN instances in the training population participate in this competition task, and the rest of the QNN instances do not participate in this competition task, for the competing QNN instances participating in this competition task, it is necessary to adjust the initial result output by the fitness evaluation model based on the fitness adjustment amount representing the competition score to obtain the fitness result of the QNN instance for comparison with the set fitness threshold; for the competing QNN instances that do not participate in this competition task, the initial result output by the fitness evaluation model is directly used as the fitness result for the QNN instance for comparison with the set fitness threshold.

[0078] Exemplarily, executing a competition task and obtaining a fitness adjustment amount of a competition QNN instance includes: obtaining a competition benefit function model; calculating a task performance indicator of the competition QNN instance based on a prediction result output by the competition QNN instance; calculating a similarity between the competition QNN instances based on the quantum state of each competition QNN instance; and inputting the task performance indicator and the corresponding similarity of the competition QNN instance into the competition benefit function model to obtain a fitness adjustment amount of the competition QNN instance.

[0079] Here, to execute the competitive task step based on the self-game strategy, it is necessary to obtain the task performance indicators of each competing QNN instance, the similarity between competing QNN instances, and the interaction-based competitive profit function model.

[0080] Among them, the task performance indicators are similar to the objective function results, and are also obtained based on the prediction results output by the competing QNN instances, and are both determined by the expected goals of this training task; for example, if this task is a classification task, the prediction results output by the QNN instance are the predicted classification results. By comparing each predicted classification result with the true label representing the actual classification and statistically comparing the results, the task performance indicators representing the classification accuracy are obtained, and the predicted classification results and the true labels are input into the objective function model to obtain the quantitative objective function results representing the classification error.

[0081] Here, the similarity between competing QNN instances is specifically the inner product metric similarity between quantum states, as shown in the following formula:

[0082] in, The closer the similarity is to 1, the more similar the quantum states and parameter strategies of the competing QNN instances participating in the competition task are.

[0083] It is understandable that the higher the similarity between QNN instances in the training population, the easier it is to fall into a local optimum. Therefore, the similarity between competing QNN instances can be used as a penalty mechanism to adjust the fitness results of competing QNN instances to ensure that QNN instances with too high similarity are eliminated during the iterative update of the training population.

[0084] Here, when executing the competitive task steps based on the self-game strategy, it is necessary to pre-build an interactive competitive benefit function model; the competitive benefit function model provides a reward and punishment mechanism for fitness results. The competitive QNN instance participating in the competitive task inputs the task performance indicator that characterizes the performance of processing the training task into the competitive benefit function model to obtain the competitive score of the competitive QNN instance, that is, the fitness adjustment amount used to adjust the initial result.

[0085] Here, the competition profit function model of the embodiment of the present application is specifically shown as follows:

[0086] in, are the quantum states of the i-th and j-th QNN instances participating in the competition task, is the competition score, i.e., the fitness adjustment amount, is the task performance index of the i-th QNN instance, and is a hyperparameter, Used to control the reward weight, The larger it is, the more obvious the competitive score gain of the dominant QNN instance in the competitive task is; Used to control the penalty weight, The larger it is, the greater the downward adjustment of the fitness results of QNN instances with high similarity.

[0087] Here, after obtaining the fitness adjustment of the competing QNN instance, the fitness result of the competing QNN instance is calculated, as shown in the following formula:

[0088] in, is the adjusted fitness result of the i-th QNN instance, is the learning rate, which is used to control the influence of the competition score on the fitness result.

[0089] In some embodiments, the fitness result of the QNN instance is obtained based on the initial result and the fitness adjustment amount of the QNN instance, including: adjusting the initial result based on the fitness adjustment amount of the QNN instance, and then normalizing it to obtain the fitness result of the QNN instance.

[0090] Here, the normalization method is specifically shown as follows:

[0091] in, is the i-th QNN instance after normalization, which is used to instruct the training population to perform iterative updates.

[0092] It can be understood that based on the normalization process, the total fitness of the training population is made to be 1.

[0093] In order to implement the method of the embodiment of the present application, the embodiment of the present application also provides an image rendering device, which corresponds to the above-mentioned image rendering method, and each step in the above-mentioned image rendering method embodiment is also fully applicable to the embodiment of this device.

[0094] like Figure 3As shown, the model parameter optimization device of the embodiment of the present application is applied to a coherent Ising machine and includes a construction module 301, an iterative update module 302, and a determination module 303. The construction module 301 is used to construct an initial training population; the initial training population includes multiple QNN instances configured with different parameter combinations; the iterative update module 302 is used to iteratively update the initial training population based on an evolutionary optimization algorithm to obtain a target training population corresponding to when the convergence condition is met; the determination module 303 is used to determine the optimal QNN instance in the target training population and obtain the parameter combination of the optimal QNN instance.

[0095] In some embodiments, the iterative update module 302 is specifically used to: calculate the fitness results of each QNN instance in the current training population, and determine whether the convergence condition is met; if not, based on the fitness results of each QNN instance in the training population, use the evolutionary optimization algorithm to update the training population, and return to the step of calculating the fitness results of each QNN instance in the current training population; if so, determine the current training population as the target training population.

[0096] In some embodiments, the determination module 303 is specifically configured to determine the QNN instance with the highest fitness result in the target training population as the optimal QNN instance.

[0097] In some embodiments, the iterative update module 302 is specifically used to: obtain a fitness evaluation model and an objective function model of a training task; after performing quantum state conversion on the training sample, input it into each QNN instance in the current training population to obtain a prediction result output by each QNN instance; input the prediction result output by each QNN instance into the objective function model respectively to obtain an objective function result of each QNN instance; input the objective function result of each QNN instance into the fitness evaluation model to obtain a fitness result of each QNN instance.

[0098] In some embodiments, the iterative update module 302 is further configured to: determine a competing QNN instance participating in a competitive task from a training population based on a self-game strategy; and execute the competitive task to obtain a fitness adjustment value of the competing QNN instance.

[0099] In some embodiments, the iterative update module 302 is specifically used to: input the objective function result of the QNN instance into the fitness evaluation model to obtain the initial result output by the fitness evaluation model; if the QNN instance is a competing QNN instance, then based on the initial result of the QNN instance and the fitness adjustment amount, obtain the fitness result of the QNN instance; if the QNN instance is not a competing QNN instance, then determine the initial result of the QNN instance as the fitness result of the QNN instance.

[0100] In some embodiments, the iterative update module 302 is specifically used to: obtain a competitive profit function model; calculate the task performance index of the competitive QNN instance based on the prediction results output by the competitive QNN instance; calculate the similarity of the competing QNN instance based on the quantum state of each competing QNN instance; input the task performance index and similarity of the competing QNN instance into the competitive profit function model to obtain the fitness adjustment amount of the competing QNN instance.

[0101] In some embodiments, the construction module 301 is specifically used to: construct a set number of QNN instances; Based on the set initialization rules, different parameter combinations are assigned to each QNN instance, and after the quantum state of each QNN instance is generated, the initial training population is obtained; wherein the set initialization rules include: classical random initialization rules or uniform distribution rules.

[0102] In some embodiments, the iterative update module 302 is specifically used to: delete QNN instances in the training population whose fitness results are lower than a set fitness threshold; and / or, perform a crossover operation on two target QNN instances in the training population to construct a child QNN instance, and add the child QNN instance to the training population; and / or, add a perturbation to the parameters of the target QNN instance in the training population to construct a variant QNN instance, and add the variant QNN instance to the training population; wherein the target QNN instance includes a QNN instance whose fitness result is higher than or equal to the set fitness threshold.

[0103] It should be noted that the model parameter optimization device provided in the above embodiment only uses the division of the above program modules as an example to illustrate when performing model parameter optimization. In actual applications, the above processing can be assigned to different program modules as needed, that is, the internal structure of the device can be divided into different program modules to complete all or part of the processing described above. In addition, the model parameter optimization device provided in the above embodiment and the model parameter optimization method embodiment are based on the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0104] Based on the hardware implementation of the above program modules, and in order to implement the model parameter optimization method of the embodiment of the present application, the embodiment of the present application also provides a coherent Ising machine, such as Figure 4 As shown, the coherent Ising machine 400 includes: at least one processor 401, a memory 402, a user interface 403 and at least one network interface 404. The various components in the coherent Ising machine 400 are coupled together via a bus system 405. It can be understood that the bus system 405 is used to achieve connection and communication between these components. In addition to including a data bus, the bus system 405 also includes a power bus, a control bus and a status signal bus. However, for the sake of clarity, the bus system 405 is not shown in FIG. Figure 4Various buses are labeled as bus system 405 .

[0105] The user interface 403 may include a display, a keyboard, a mouse, a trackball, a click wheel, keys, buttons, a touch pad or a touch screen.

[0106] The memory 402 in the embodiment of the present application is used to store various types of data to support the operation of the coherent Ising engine 400. Examples of such data include: any computer program used to operate on the coherent Ising engine 400.

[0107] The model parameter optimization method disclosed in the embodiments of this application can be applied to or implemented by processor 401. Processor 401 may be an integrated circuit chip with signal processing capabilities. During implementation, each step of the model parameter optimization method can be completed by hardware integrated logic circuits or software instructions in processor 401. The processor 401 may be a general-purpose processor, a digital signal processor (DSP), or other programmable logic device, discrete gate or transistor logic device, discrete hardware component, etc. Processor 401 can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of this application. A general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly implemented and executed by a hardware decoding processor, or by a combination of hardware and software modules in the decoding processor. The software module may be located in a storage medium located in memory 402. Processor 401 reads information from memory 402 and, in conjunction with its hardware, completes the steps of the model parameter optimization method provided in the embodiments of this application.

[0108] In an exemplary embodiment, the coherent Ising machine 400 can be implemented by one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), FPGAs, general-purpose processors, controllers, microcontrollers (MCUs), microprocessors, or other electronic components to perform the aforementioned model parameter optimization method.

[0109] It is understood that memory 402 can be volatile memory or non-volatile memory, or can include both volatile and non-volatile memory. Non-volatile memory can include read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), EEPROM, ferromagnetic random access memory (FRAM), flash memory, magnetic surface storage, optical disk, or compact disc read-only memory (CD-ROM); magnetic surface storage can include magnetic disk storage or magnetic tape storage. Volatile memory can include random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static random access memory (SRAM), synchronous static random access memory (SSRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct rambus random access memory (DRRAM). The memory 402 described in the embodiments of the present application is intended to include, but is not limited to, these and any other suitable types of memory.

[0110] In an exemplary embodiment, the present application also provides a storage medium, namely, a computer storage medium, which can be a computer-readable storage medium, for example, including a memory 402 storing a computer program. The computer program can be executed by the processor 401 of the coherent Ising machine 400 to complete the steps of the model parameter optimization method described in the embodiment of the present application. The computer-readable storage medium can be a memory such as ROM, PROM, EPROM, EEPROM, Flash Memory, magnetic surface storage, optical disk, or CD-ROM.

[0111] In an exemplary embodiment, the present application further provides a computer program product, including a computer program, which can be executed by the processor 401 of the coherent Ising machine 400 to complete the steps described in the method of the embodiment of the present application.

[0112] It should be noted that: "first", "second", etc. are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0113] In addition, the technical solutions described in the embodiments of the present application can be arbitrarily combined without conflict.

[0114] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A model parameter optimization method, characterized in that: Applied to a coherent Ising machine, the method comprises: Constructing an initial training population; the initial training population includes multiple quantum neural network (QNN) instances configured with different parameter combinations; Based on the evolutionary optimization algorithm, the initial training population is iteratively updated to obtain the target training population corresponding to the convergence condition; An optimal QNN instance in the target training population is determined, and a parameter combination of the optimal QNN instance is obtained.

2. The method according to claim 1, characterized in that The iterative updating of the initial training population based on the evolutionary optimization algorithm to obtain the target training population corresponding to the convergence condition includes: Calculate the fitness results of each QNN instance in the current training population and determine whether the convergence conditions are met; If not, based on the fitness results of each QNN instance in the training population, the training population is updated using an evolutionary optimization algorithm, and the step of calculating the fitness results of each QNN instance in the current training population is returned to; If so, the current training population is determined to be the target training population; Determining the optimal QNN instance in the target training population includes: Determine the QNN instance with the highest fitness result in the target training population as the optimal QNN instance.

3. The method according to claim 2, characterized in that The calculation of the fitness results of each QNN instance in the current training population includes: Obtain the fitness evaluation model and the objective function model of the training task; After the training samples are converted into quantum states, they are input into each QNN instance in the current training population to obtain the prediction results output by each QNN instance; Input the prediction results output by each QNN instance into the objective function model to obtain the objective function results of each QNN instance; The objective function result of each QNN instance is input into the fitness evaluation model to obtain the fitness result of each QNN instance.

4. The method according to claim 3, characterized in that The method further comprises: Determining, from the training population, a competing QNN instance that participates in a competitive task based on a self-playing strategy; Executing the competition task to obtain a fitness adjustment value of the competing QNN instance; The objective function result of each QNN instance is input into the fitness evaluation model to obtain the fitness result of each QNN instance, including: Inputting the objective function result of the QNN instance into the fitness evaluation model to obtain an initial result output by the fitness evaluation model; If the QNN instance is the competing QNN instance, obtaining a fitness result of the QNN instance based on the initial result of the QNN instance and the fitness adjustment amount; If the QNN instance is not the competing QNN instance, determining the initial result of the QNN instance as the fitness result of the QNN instance.

5. The method according to claim 4, characterized in that The executing the competition task to obtain the fitness adjustment value of the competing QNN instance includes: Obtain a competitive payoff function model; Calculating a task performance indicator of the competing QNN instance based on the prediction result output by the competing QNN instance; Calculating similarities between the competing QNN instances based on the quantum states of the competing QNN instances; The task performance index and the corresponding similarity of the competing QNN instance are input into the competition benefit function model to obtain the fitness adjustment amount of the competing QNN instance.

6. The method according to any one of claims 1 to 5, characterized in that The constructing of the initial training population includes: Build a set number of QNN instances; Based on the set initialization rules, different parameter combinations are assigned to each of the QNN instances, and after the quantum state of each of the QNN instances is generated, an initial training population is obtained; The setting initialization rule includes: a classic random initialization rule or a uniform distribution rule.

7. The method according to any one of claims 2 to 5, characterized in that: The updating of the training population using an evolutionary optimization algorithm based on the fitness results of each QNN instance in the training population includes at least one of the following: Deleting the QNN instances in the training population whose fitness results are lower than a set fitness threshold; Performing a crossover operation on two target QNN instances in the training population to construct a child QNN instance, and adding the child QNN instance to the training population; Adding a perturbation to the parameters of the target QNN instance in the training population to construct a variant QNN instance, and adding the variant QNN instance to the training population; The target QNN instance includes a QNN instance whose fitness result is higher than or equal to a set fitness threshold.

8. A model parameter optimization device, characterized in that: Applied to a coherent Ising machine, the device comprises: A construction module is used to construct an initial training population; the initial training population includes multiple QNN instances configured with different parameter combinations; An iterative update module, configured to iteratively update the initial training population based on an evolutionary optimization algorithm to obtain a target training population corresponding to when a convergence condition is met; A determination module is used to determine the optimal QNN instance in the target training population and obtain a parameter combination of the optimal QNN instance.

9. A coherent Ising machine, characterized in that The coherent Ising machine comprises: a processor and a memory for storing a computer program that can be run on the processor, wherein the processor is configured to execute the steps of the method according to any one of claims 1 to 7 when running the computer program.

10. A storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by the processor, the steps of the method according to any one of claims 1 to 7 are implemented.

11. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images