System optimization method

By receiving multiple candidate parameter value sets in the computer, applying an optimization algorithm to generate candidate update status and evaluate the optimality score, the problems of combination optimization problems in the prior art are solved, and the problem of high resource consumption is realized within the budget Approximate optimal solution generation.

CN118974723APending Publication Date: 2024-11-15MICROSOFT TECHNOLOGY LICENSING LLC
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202380032043.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2022-03-28
Filing Date
2023-02-16
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

Existing combined optimization problem solvers are difficult to effectively find the global optimal solution under the processing budget constraints, and existing methods usually rely on external hyperparameter optimization, which has high efficiency and resource consumption.

Method used

By receiving multiple candidate parameter value sets in the computer, an optimization algorithm is applied to generate candidate update status, and the optimal score is evaluated based on the optimality metric, and the state of the highest score is selected as the approximate optimal solution to balance processing efficiency and resource usage.

Benefits of technology

Generating approximate optimal solutions within the processing budget improves the solution efficiency of combined optimization problems and is suitable for environments with various computing resource limitations.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118974723A_ABST
    Figure CN118974723A_ABST
Patent Text Reader

Abstract

A computer-implemented method includes: receiving a plurality of candidate parameter value sets in a specified ordered sequence, each candidate parameter value set including a respective candidate parameter value for at least one parameter of an optimization algorithm, where the number of candidate parameter value sets is based on a processing budget; for each set of candidate parameter values in the sequence: applying an optimization algorithm to a plurality of initial states of a model representing the system, where at least one parameter is set to a respective candidate parameter value, thereby generating a corresponding candidate update state, and evaluating each candidate update state according to an optimality metric, thereby generating a corresponding optimality score; selecting the candidate update state with the highest optimality score as an estimated optimal state of the model; and outputting the selected estimated optimal state of the model to a user interface, computer interface, or other application.
Need to check novelty before this filing date? Find Prior Art

Description

Background Art

[0001] Combinatorial optimization problems are a group of problems with various applications, including resource scheduling, capacity planning, drug discovery, and building low-density parity-check codes for 5G technology.

[0002] Many combinatorial optimization problems are solved by formulating the problem as an Ising model. The Ising model is formulated as the minimization of a quadratic function: Among them, s i is a variable, usually called spin, taking binary values ​​±1, and G ij are the elements of the matrix G, usually called the interactions between spins. The solution to the Ising model is to assign binary values ​​to the variables s that minimize the above expression i The model can be used to find approximate solutions to real-world optimization problems by mapping the variable interactions and objectives of the real-world problem to the Ising model definition, so that minimizing the energy of the Ising model is equivalent to optimizing the objective of the real system. The Ising problem can alternatively be formulated as a quadratic unconstrained binary optimization problem (QUBO), or when G ij Define nodes i When the edges of the graph between G ij The maximum cut of .

[0003] Many existing solvers applied to combinatorial optimization use a technique called simulated annealing to determine the optimal state. Simulated annealing works by moving between random adjacent states with some randomness based on the energies in those states until an estimated optimal state is reached.

[0004] Some existing Ising solvers use a standard gradient descent update scheme to update the state of the Ising model to the minimum energy using the gradient of the energy function, where a hyperparameter optimization package is used to determine the hyperparameters of the algorithm. Summary of the invention

[0005] Existing solvers for combinatorial optimization problems typically use external hyperparameter optimization methods to determine the optimal parameter set of an optimization algorithm for generating an optimal state. This article describes a technique for determining an approximate optimal solution to such an optimization problem by using a parameter set for a given optimization algorithm, processing multiple different parameter value sets, generating an updated state set as a candidate solution, and selecting an overall optimal solution from the candidate set. The method disclosed herein provides an efficient and flexible optimization method by implementing an optimization algorithm with multiple different parameter value sets. The evaluation of multiple parameterized results of the optimization algorithm is useful in order to find a state close to the optimal state. The number of parameter value sets used in the optimization algorithm depends on the processing budget. For a given processing budget, the present invention outputs an estimated optimal state of the system being modeled by evaluating the candidate states generated in the application of the optimization algorithm for different parameter value sets, wherein the application of the optimization algorithm is defined to remain within the processing budget. This balances the advantages of implementing different parameter value sets for finding candidate optimal states, wherein efficiency needs to be maintained within the budget of other limitations of processing time, energy consumption, or computing resources.

[0006] The number of sets of parameter values ​​depends on the processing budget of the computer on which the optimization algorithm is implemented. As described below, the computer can be a conventional electronic-digital computer including a processor and memory, or an optical computer including a light source and optical hardware components (such as photodetectors, lenses, and diffraction elements). The processing budget can be based on any of the following: processing time, energy consumption of the computer, memory usage of the computer, number of separate operations performed by the computer, or any other resource used by the computer during processing.

[0007] Note that in the following description, reference is made to estimated and approximate optimal solutions. This is because the methods described herein are intended to find optimal solutions to a given problem, but are not guaranteed to find a global optimum. Therefore, these solutions may be globally suboptimal, but provide optimal solutions that are available within the constraints of computing resources and time. Any reference herein to an optimal solution produced by the methods described herein should be interpreted as an approximate or estimated optimal solution.

[0008] More generally, according to one aspect disclosed herein, a computer-implemented method is provided, comprising: receiving a plurality of candidate parameter value sets in a specified ordered sequence relative to each other, each of the plurality of candidate parameter value sets comprising a corresponding candidate parameter value for at least one parameter of an optimization algorithm, wherein the number of received candidate parameter value sets is based on a processing budget; for each candidate parameter value set in the ordered sequence: applying the optimization algorithm to each of a plurality of initial states of a model representing a system, wherein at least one parameter of the optimization algorithm is set to a corresponding candidate parameter value, thereby generating a corresponding plurality of candidate update states of the model, and evaluating each of the candidate update states according to an optimality metric, thereby generating a corresponding optimality score for the candidate update state, such that different of the candidate update states have different optimality scores; selecting the candidate update state with the highest corresponding optimality score as the estimated optimal state of the model; and outputting the selected estimated optimal state of the model to one of: a user interface, a computer storage device, or a destination application.

[0009] The parameter value sets may be received from an application for generating parameter value sets, from a network interface, from a device external to the computer system performing the optimization, or retrieved from a storage device, such as from a database or other data structure stored in a computer memory, or from a user. The candidate parameter value sets are received in order. The received candidate parameter value sets are selected as a subset of the first k parameter value sets in the ordered list of parameter value sets, where k is a predefined number or a number dynamically selected based on a processing budget. Alternatively, the received parameter value sets may be randomly generated.

[0010] The system may be represented by a graph defining the interactions between pairs of variables of the system, where the state of the system is the assignment of binary values ​​to the variables of the system. The model may be an Ising model.

[0011] The optimization algorithm may be a gradient descent algorithm. The gradient descent algorithm may use one of standard momentum or Nesterov momentum. The gradient descent algorithm may alternatively or additionally use adaptive moment estimation.

[0012] The system represented by the model can be one of a computing system, an electronic system, or a mechanical system, wherein the method further comprises implementing the estimated optimal state in the system. In some embodiments, in some embodiments, for drug design applications, the system represented by the model can be a family of compounds or a molecular scaffold defining an underlying structure, wherein a finite number of molecular combinations can be assigned to specific locations within the underlying structure. The system can be a collection of multiple computing resources for implementing a collection of processes, wherein the state is an assignment of the processes to corresponding computing resources in the multiple computing resources.

[0013] The method disclosed herein is implemented in a computer. The computer may be a conventional digital computer including a memory and one or more computer processors that execute the method described herein by executing instructions. In this case, the optimization algorithm described herein may be provided as software, which is interpreted by a computer to generate a set of machine code instructions executed in conventional computer hardware. Alternatively, the computer may be an optical computer including optical hardware for implementing the steps of the algorithm described herein. An optical computer does not necessarily operate by executing instructions like a conventional computer, but uses optical components such as lasers, photodetectors, modulators, etc. to implement arithmetic and logical operations. In an optical computer, a variable is modeled by an optical signal generated, for example, by a laser or other light source, wherein the optimization algorithm is implemented by using one or more optical hardware components (such as photodetectors, lenses, and diffraction gratings) to interact with the optical signal to generate an optical signal set representing the approximate optimal state of the modeled system.

[0014] This summary is provided to introduce in simplified form a selection of concepts that will be further described in the detailed description below. This summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used to limit the scope of the claimed subject matter. The claimed subject matter is also not limited to implementations that solve any or all of the disadvantages noted herein. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] To aid understanding of the present disclosure and to show how the embodiments may be effected, reference is made to the accompanying drawings by way of example, in which:

[0016] Figure 1 is a schematic block diagram of an algorithm for finding the optimal state of a modeled system,

[0017] Figure 2 is a flowchart showing the steps of an algorithm for optimizing the state of a model,

[0018] Figure 3 is a schematic block diagram showing the implementation of an optimization algorithm within a multi-step optimization framework,

[0019] Figure 4A An example scheme for searching a two-dimensional parameter space is shown,

[0020] Figure 4B shows the partitioning of the parameter space into regions based on the observed performance of the relative performance metric,

[0021] Figure 5A A schematic block diagram showing an optical hardware implementation of the optimization method,

[0022] Figure 5BA schematic block diagram showing an optical hardware implementation of an example optimization algorithm, and

[0023] Figure 6 An example wavelength selective switch configuration for performing optical vector-matrix multiplication is shown. DETAILED DESCRIPTION

[0024] Figure 1 A schematic block diagram of a method for finding an optimal state for a model of a real-world system is shown. The system can adopt a large number of possible states, where the optimal state is the state that provides the best performance according to some characteristics of the system (such as the state with the lowest energy). In some embodiments, the model includes a graph or matrix representing the interaction between multiple variables of the system, where a given state s is an assignment of values ​​to a set of variables of the system. In some examples, the variables are binary, taking values ​​1 and -1, or alternatively 1 and 0. The examples described herein mainly relate to binary variables, but the methods disclosed herein can also be applied to states with variables that adopt any discrete values. As described later, continuous variables can also be optimized according to the methods described herein.

[0025] The algorithm is used to apply multiple possible parameterizations of a predefined optimization algorithm to multiple initial states of a system to generate a set of candidate updated states as possible solutions to the optimization problem. The optimization algorithm updates the initial state to generate the candidate state. The optimality score is calculated by applying an optimality metric as a measure of the candidate state. For example, this can be an energy function representing the energy of a physical system as described below, or some other measure representing the goal of a given optimization problem. For example, in the case where the problem is the allocation of computing tasks to computing resources, the optimality score can depend on the time spent on performing all tasks or a similar measure. Similarly, for drug discovery, the optimality score can depend on the toxicity and / or effectiveness of the candidate compound represented by the candidate state. In this case, the optimality metric is defined so that a high optimality score is assigned to a candidate compound with low toxicity and high efficiency. The optimality metric can be a function of multiple factors on which the optimal state depends. The optimality metric is applied to each candidate state to generate an optimality score for the state. The state with the highest optimality score can then be selected as the estimated optimal state. This paper uses "optimality metric" to refer to the function applied to a state to measure the optimality of the state for a given optimization, while "optimality score" is used to refer to the value obtained when the optimality metric is applied to a given state.

[0026] The method can be applied to determine an approximately optimal state for a real-world system. This may be an electronic system, such as a collection of computer processors, to which a collection of processes is optimally allocated, wherein the optimality metric in this case is a measure of the overall use of computing resources used to handle the processes.

[0027] Another possible real-world application of the methods described herein is in the process of drug discovery. In this case, the system can be a structure or scaffold that defines the type of molecules that are desired to be tested and synthesized in order to develop a drug, wherein the scaffold provides a set of possible positions in which each of a finite set of atoms and / or molecules is placed. For this problem definition, the state of the system is represented by assigning an element or compound to each position of the scaffold. The optimality of the solution can be evaluated by defining an optimality metric that is maximization for the desired properties of the drug (such as effectiveness for the desired function of the drug, ease of manufacture, etc.) and minimization for undesirable properties (such as toxicity). In this case, the state may include a binary variable A collection of positions R i and compound j is indexed so that when compound j is placed at position R i The variable has the value 1 when , otherwise it has the value 0. In addition to the optimal properties of the drugs described above, an optimality measure is defined (which is based on the objective function to be maximized) so that two compounds can be assigned to the same position by penalizing such a state with a heavy penalty term. The optimality measure and the optimization algorithm are selected based on the requirements of the optimal state for a given drug design problem. For example, in the case where it is desired that the drug has a low total weight, a term can be added to the optimality measure: Among them, w j is the weight of compound j in position R1, where additional terms related to other positions are not written but are also included in the function F. As required, when the weight w j When , ...

[0028] The method is applied by processing an ordered sequence 100 of parameter value sets in a given order. As described in more detail later with respect to a specific embodiment of the present invention, the ordered sequence can be selected as a subset from a sorted list of parameters. For example, a sorted list can be generated from a predetermined range of each parameter based on the expected performance of an optimization algorithm for those parameters. The expected performance can be measured by performing a search of the parameter space and applying a modified optimization algorithm with each parameter value set of the search parameters to an exploratory initial state set of the system, and measuring a performance metric on the optimality score of the update state generated by the modified optimization algorithm. The performance metric is a representative metric of the optimality score on the exploratory update state set for a given parameter value set. Examples of performance metrics include the mean, median, minimum, or maximum of the optimality scores aggregated over all exploratory update states. For example, the performance metric can be the average, median, minimum energy of the exploratory state set, where the energy of the system at a given state is used as an optimality metric to determine the optimality score (where energy is inversely related to optimality). The performance measure may also be any other statistical measure of the optimality score, or any other measure related to the optimality score, such that the performance measure represents the expected performance of the optimization function with that set of parameter values.

[0029] The parameter search may be performed by applying a Bayesian approach. Alternatively, a grid search of a region of the parameter space may be performed to generate a set of parameters to be processed in the exploration phase. In another alternative, the parameters to be processed may be randomly generated.

[0030] The sequence may be fully determined before processing begins, or the sequence to be processed may be a subset of a longer list, where the actual sequence processed is determined from the ordered list based on a processing budget (i.e., the resources available for processing) and the sequence is completed once that budget is reached.

[0031] Where the system is represented by a matrix, a linear stability analysis of the eigenvalues ​​of the matrix can be used to determine the range to explore for each parameter, as described in more detail later. Alternatively, the ranges can be manually initialized and dynamically updated based on the optimality scores calculated for the updated states in the exploration phase.

[0032] The number of processed candidate parameter value sets is the number that can be processed within the processing budget. The processing budget is defined according to one or more of the following: processing time, number of processing cycles, amount of memory resources consumed, and / or energy consumption. For example, in an embodiment, the processing budget may include a limit on processing time, and the number of processed candidate parameter value sets is the number that can be processed within the processing time limit. In another alternative embodiment, the processing budget may include a limit on the amount of memory resources consumed, and the number of processed candidate parameter value sets is the number that can be processed within the limit on memory. In other embodiments, the processing budget may include a limit on the number of processing cycles executed, and the number of processed candidate parameter value sets is the number that can be processed within a limited number of processing cycles. Alternatively, in an embodiment, the processing budget may include a limit on processing energy consumption, and the number of processed candidate parameter value sets is the number that can be processed within the limit on energy consumption. In another embodiment, the processing budget may include constraints on two or more such factors. For example, when any one of two or more (e.g., time limit or energy consumption) or constraints can be defined according to a combined metric combining two or more factors, processing can stop.

[0033] Each parameter value set of the sequence contains values ​​for parameterizing the optimization algorithm 102. In some embodiments, the optimization algorithm takes two parameters, and each item in the sequence is a pair of values ​​corresponding to those parameters. Other optimization algorithms take three values, and the sequence is a sequence of triplets. This example is in Figure 1 , where each parameter value set is shown as a 3-tuple (P1, P2, P3).

[0034] The optimization algorithm may be an iterative algorithm that updates the state of the system from an initial state within a predetermined number of iterations by repeatedly applying an update rule. Some such optimization algorithms apply iterative gradient descent updates over a predetermined number of iterations, where the gradient function is parameterized by two parameters. Other example algorithms use gradient descent with a momentum term parameterized by a third momentum parameter γ. These examples will be described in more detail later. The modified optimization algorithm for generating the sorted list may be the same iterative algorithm implemented over a smaller number of iterations. The gradient descent method uses the gradient of a function to determine the direction of the update toward a local minimum, where the gradient is a vector of first-order partial derivatives of a function with respect to each of its variables. Other possible iterative optimization algorithms may use higher-order derivatives in order to determine the direction of the update.

[0035] Methods based on non-derivatives (or derivative-free) can also be used. For example, as mentioned above, simulated annealing is a common algorithm applied to solve combinatorial optimization problems. Another family of optimization algorithms that can be used to solve this optimization problem is evolutionary algorithms. These algorithms include particle swarm optimization and genetic algorithms. These techniques are well known in the field of optimization and will not be discussed in detail in this article. The method for optimizing the system described herein can be used in combination with any of these optimization methods. In other words, a process can be applied, wherein for any of the mentioned optimization algorithms, an ordered list of parameter value sets defining the optimization algorithm can be processed in a given order, and the corresponding optimization algorithm can be applied to the initial state of the system to generate a candidate solution.

[0036] Each value set is selected in a defined order and the optimization algorithm parameters are set to the corresponding values ​​P1, P2, ... for the given set. The optimization algorithm is applied to a plurality of initial states 104 of the system, the initial states Represents the first initial state, the initial state represents the Nth initial state. This generates N updated states The process is repeated for each parameter value set until a full set of candidate update states 106 is generated. An initial state may be randomly generated for each parameter value set. Alternatively, a single set of initial states may be generated and reused with each parameter value set. Note that for the implementation of the described technology in software, it is computationally efficient to generate a single set of initial states and reuse them for each parameter value set, whereas for the optimization algorithm implemented in optical hardware, the initial state of the system may be different for each parameter value set, which may be modeled as the intensity or phase of a set of one or more optical signals, as described in more detail later. When setting the initial state in optical hardware, one option is to set all states to have an initial value of all zeros. Alternatively, random initialization may be applied, in which random states are initially generated in the digital domain, and these random states are passed to a spin generator to generate a modulated optical signal representing these random initial states. Having different randomly sampled initial states allows a wider coverage of the solution space to be searched. More complex sampling strategies may be used to provide a set of initial states to which the optimization algorithm is applied. For example, some strategies use centrality measures for a given problem to identify the relative importance of variables. Common centrality measures include betweenness, degree, PageRank, etc. These measures are defined with respect to the nodes of a graph, which is one possible representation of the variables for a combinatorial optimization problem. The selection of the initial state can then be performed so that the direction of the most important nodes (i.e., variables) is fixed.

[0037] An optimization algorithm is defined as updating the state of a system in the direction of an optimal solution according to some goal or objective of the optimization problem represented by an optimality score, such as by minimizing an energy function representing the total energy of the system, where the optimal state is the lowest energy state of the system.

[0038] Once the set of candidate update states 106 is generated, an optimality metric is applied to each candidate state, which may represent the energy of the physical system or some other optimality measure. The highest scoring state s according to the given optimality metric is * is chosen as the estimated optimal state of the system. The optimality measure can be defined in terms of some cost function or objective function that represents the goal of optimizing a given system.

[0039] The estimated optimal state of the system is then output. The state can be output to the user via a user interface (such as a graphical user interface for display on a screen or as an audible output). Alternatively, the state can be directly output to another system or application to be used directly in the configuration of the system state, such as by passing the state to a computer-aided design tool, or to a controller of a real-world system that can interpret the output state and implement the output state in the system. The estimated optimal state can also be output to a device outside the computer system performing the optimization, or to a network interface. In other embodiments, the state is output to a computer storage device, which can be a temporary (such as a RAM) or a non-volatile storage device (such as a hard disk drive, a disk storage device, etc.), wherein it can be accessed by a user or a computer application (such as a control application of a modeling tool or a system).

[0040] The modeling system may be a real-world system. The real-world system may be any of the following: a computing system, an electronic system, an electrical or electromagnetic system, a mechanical system, a chemical system, or an application for treating a biological being, or any other real-world system that may be expressed as a combinatorial optimization problem. The estimated optimal state output by the optimization algorithm as described above may be achieved by manufacturing, constructing, or programming a state in a real-world system, or by modifying the existing state of the system through a physical, chemical, mechanical, or electrical process. The user may directly affect the estimated optimal state in the system, such as by moving, modifying, or constructing a component of a physical system. Alternatively, the user may provide input to a computer application or other tool to act on the state in the system, such as by configuring a computer-aided design tool and / or manufacturing equipment to create an optimal system state or modify an existing state to an optimal state. Similarly, a computer application such as a computer-aided design or modeling application may receive an estimated optimal state and control a manufacturing device or other machinery to affect the state in a real-world system, which may be a physical, chemical, mechanical, or electrical system. As mentioned above, combinatorial optimization problems have a wide range of real-world applications, including resource scheduling, building low-density parity-check codes for 5G, and drug discovery, among other applications known to those skilled in the art of optimization methods.

[0041] Figure 2 Shows Figure 1 Flowchart of the method shown in FIG. 1 . At step S200, an ordered sequence of parameter value sets is received. At step S202, the first value set in the sequence is assigned to the parameters of the optimization algorithm. At step S204, the optimization algorithm is executed with the given parameter values ​​at a plurality of initial states to generate a set of candidate update states at step S206. These are added to the set of candidate update states as shown in FIG. Figure 1 The entire set of candidate states 106 described above is presented. At step S212, an optimality score is calculated for each candidate update state using an optimality metric. As described above, an optimality metric is a measure of a given state as a solution to a given system being optimized, and is defined in terms of the system and the optimization objective. At this point, if it is determined that there are other parameter value sets to be processed, the next parameter value set is assigned to the parameters of the optimization algorithm, and steps S206 and S206 are repeated until there are no other parameter value sets to be processed. An optimality score is calculated for each parameter.

[0042] Although Figure 2The steps of executing the optimization algorithm, generating candidate update states, calculating optimality scores, and assigning the next parameter value set as a continuous sequence are shown, but these steps can be performed in parallel on multiple processors so that the initial state set can be processed by the optimization algorithm using a first value set, and before that processing is completed, a second parameter value set can be assigned to the parameters of the optimization algorithm for processing on a second processor while the optimization algorithm is executed on the first processor. If not all parameter value sets have been assigned to the parameters of the optimization algorithm for processing, then the condition is not satisfied. Figure 2 However, the step of selecting the highest scoring candidate as the estimated optimal solution only occurs when the full set of candidate update states has been generated, i.e., after all processing of the optimization algorithm has been completed, or a processing time limit has been reached. While the number of parameter value sets to be processed can be determined in advance based on the processing budget, the number actually processed can be adjusted dynamically by ending the processing once a given time limit, memory, etc. has been reached.

[0043] Although Figure 2 A step of calculating an optimality score for each candidate state immediately after computing the state is shown, but this step can be performed at any time after the state is generated. This calculation can also be performed in parallel on multiple processors and can be performed for different update states at different times. The set of candidate states and their optimality scores are maintained in memory and updated as processing continues because the algorithm outputs more candidate update states and scores until all processing is complete. At this point, S214, the candidate update state corresponding to the highest optimality score is selected as the estimated optimal state for the system.

[0044] The present invention uses a sequence of parameter value sets for a predefined optimization algorithm and generates estimated optimal states of the system by applying the optimization algorithm parameterized by the ordered sequence of value sets to various initial states of the system. This provides a method for generating near-optimal solutions within a processing budget, since the number of parameter value sets processed depends on the processing budget. Furthermore, since the optimization algorithm processes states for different parameter values ​​independently of other parameter values, the method is suitable for parallel processing implementations.

[0045] Existence according to the above reference Figure 1 and Figure 2 The steps outlined generate a number of possible methods for an ordered list of parameters for processing. Different forms of parameterized optimization algorithms may also be used in accordance with the described methods. Example embodiments of the invention will now be described in which these features are discussed in more detail.

[0046] Figure 3 A high-level overview of the algorithm's structure is given to first generate the ordered sequence of parameters used to define the optimization algorithm, then perform the Figure 1 and Figure 2 A method is described, also referred to herein as "deep search", for generating a set of candidate solutions to an optimization problem and selecting an estimated optimal state.

[0047] The optimization algorithm 102 used to generate candidate update states to be considered as possible solutions to the problem is parameterized by a set of parameters. These parameters can take a wide range of values, and one problem with the method described for performing the optimization function with multiple sets of parameter values ​​is that it is intractable to perform a complete optimization for a large number of sets of parameter values ​​covering a large parameter space. Figure 3 The method shown in introduces two additional steps to reduce the parameter set to a tractable and useful parameter set before the "deep search" of the algorithm for a large number of initial states is implemented. In the preprocessing stage 302, an input problem 300 is provided, which in this example includes a graph or matrix G defining the interactions between the variables of the system. ij In the preprocessing stage 302, the problem is analyzed, and given the parameters of the optimization algorithm, ranges are identified for searching each parameter to avoid wasting resources when applying the optimization algorithm with significantly suboptimal parameter values. The preprocessing stage 302 identifies the eigenvalues ​​and eigenvectors of the graph or matrix defining the problem, and performs a linear stability analysis to identify suitable ranges for the parameters of the optimization function. This analysis is described in further detail below. The output of the preprocessing stage 302 is a set of ranges for each parameter to narrow the parameter search space for the next stage of processing.

[0048] In the exploration phase 304, the optimization algorithm 102 is applied to a set of parameter combinations covering the space defined by the ranges identified in the preprocessing phase 302. As described in more detail later, the optimization algorithm 102 in this implementation uses gradient descent-based updates to generate a near-optimal state from an initial state over multiple iterations. In order to reach a state that is close to the optimal from a given initial state, the optimization algorithm 102 should run a large number of iterations to allow it to converge to the optimum. However, in the exploration phase, there is a possible range of parameters for the optimization function, and it is not possible to perform a deep optimization process for every set of parameter values.

[0049] The goal of the exploration phase 304 is not to find the overall optimal solution to the input problem, but to identify a subset of parameter value combinations that may provide good optimization performance and are easy to explore in the deep optimization or "deep search" phase. In order to evaluate the larger parameter set for its expected performance in the optimization algorithm searched in the exploration phase, a "shallow" optimization is performed, using a small number of iterations to update a small number of initial states for each parameter value combination searched. The number of initial states and the number of iterations are defined before the processing in the exploration phase. The number of iterations may also be referred to as the "depth" of optimization in this article. The depth of optimization applied in the exploration phase 304 is less than the depth of optimization applied in the deep search phase 306. Once a shallow optimization is run for a set of suitable parameter value combinations within the range determined in the preprocessing step, the parameter value combination can be evaluated based on the metric applied to the updated state generated by the shallow optimization. In the case where the goal of the optimization is to minimize the total energy of the Ising system (as described in more detail below), the metric for each parameter value combination can be, for example, the average energy of the updated state generated by the shallow optimization algorithm for the parameter value combination. Alternatively, the minimum or median or any similar metric on the energy generated by the shallow optimization can be evaluated for each parameter value combination. This evaluation is used to identify the best parameter set, i.e., those parameters with the lowest average (or median, minimum, etc.) energy are considered the "best" parameter value combination and therefore the combination that should be evaluated first in the deep search stage. The subset of parameter value combinations with the best performance relative to a given metric is output from the exploration stage as an ordered sequence 100 for processing in the deep search stage 306.

[0050] In the deep search phase 306, an ordered sequence of sets of parameter values ​​(also referred to above as parameter value combinations) is processed by assigning parameter values ​​of each sequence set to parameters of the optimization algorithm and performing the optimization algorithm update over multiple iterations for multiple initial states. Although the optimization algorithm 102 applies the same update scheme in the exploration phase and the deep search phase, the number of iterations and initial states for each parameter value set in the deep search phase is much larger than the number applied for each "shallow" optimization performed in the exploration phase because the number of parameter value sets to be processed is smaller. For each updated state generated by the optimization algorithm in the deep search phase 306, an optimality metric is applied to evaluate the updated state based on the objective of the optimization. The updated state s with the highest optimality score is selected. * As an estimated solution to problem 300 .

[0051] As mentioned above, the goal of the Ising model is to minimize the total energy of the system, so the energy of the system can be calculated for each updated state and regarded as a cost measurement. In this article, "optimality metric" is used to describe the measurement for maximizing the optimal solution, wherein the term "cost" is used to describe the measurement for minimizing the optimal solution. The cost and optimality metrics as used herein have the same effect in this application, and the application evaluates the solution based on the degree to which the optimization goal is met. The cost value can be derived from the score, and vice versa. For example, by setting the optimality metric to the negative of the energy function, the highest optimality score corresponds to the lowest energy. Any one of them can be used to represent the measurement of optimality in this article.

[0052] Some details of an example optimization algorithm 102 used in one implementation of the present invention will now be described in more detail. The example optimization algorithm will be referred to as an accelerated optimization algorithm. This is an accelerated gradient descent-based optimization algorithm parameterized by multiple parameters (α, β, γ) according to the methods already described. Figure 3 The preprocessing and exploration phases described in are used to generate a sequence 100 of parameter values ​​for the optimization algorithm.

[0053] The accelerated optimization algorithm is configured to solve a problem provided in the form of an Ising model. The Ising model is formulated as the minimization of a quadratic function: Among them, s i is a variable taking binary values ​​±1, and G ij is the element of the matrix G for the interactions between spins. The solution to the Ising model is to assign binary values ​​to the variables s that minimize the above expression i .

[0054] To address this problem, an iterative optimization algorithm 102 may be used. This applies an iterative update of the form: where x is a continuous real-valued vector of states that is mapped to a binary variable s, and is the gradient of a function F that depends on the state vector x and the interaction G and is designed to encourage updates to the vector x towards a set of values ​​corresponding to lower energy states of the system. The above update scheme is often referred to as gradient descent and is known in the field of optimization.

[0055] The above algorithm can be used to solve different formulations of combinatorial optimization problems, where function F is defined according to the requirements for a given problem. F is defined for a given problem so that the state x that minimizes F is optimal for the given problem. For example, for a problem of drug discovery, a function F can be defined that is minimized when the state corresponds to a candidate with desired drug properties, for example, for high efficacy and low toxicity. The following description provides further details on how to apply the algorithm of equation (2) to an optimization problem in the form of an Ising problem. However, as will be clear to those skilled in the art, the same algorithms and techniques described herein for selecting parameters for an optimization algorithm can be directly applied to other combinatorial optimization problems by defining a suitable difference function F for the problem.

[0056] By defining an appropriate function F, the above iterative update algorithm can be applied to the vector x corresponding to the initial state of the system over multiple iterations to find an estimated solution to the Ising problem. The exact form of F can be defined differently for different problems. The inventors have recognized that for the above definition of the Ising problem, a particularly effective gradient Takes the following general form: where α and β are parameters of the optimization algorithm, and where H1, H2, and H3 refer to three vectors of linear or nonlinear functions that can be implemented in hardware. H1, H2, and H3 can be applied element-wise: for each element of H1, H2, and H3, a different type of linear or nonlinear function can be used. The above expression specifies The general form of , but no specific function to be used is defined, as a variety of computer implementable functions can be effectively used in functions of the above form. The flexibility of using linear or nonlinear functions allows solving combinatorial problems that also include continuous variables that need to be optimized.

[0057] For example, a value having the form specified in Equation 3 can be implemented The specific expression is: where sin and tan are well-known trigonometric functions, and sign is defined as:

[0058] The parameter α characterizes the system dissipation rate and controls how much the state x is reduced at each time step. α itself depends on time in a linear or nonlinear manner. In one example, α is linear in time as follows: Where T is the final time after all iterations are completed.

[0059] Given the above The expression for state update can be written as equation (2): x t+1 =x t -Δt.(α·H1(x t )-β·H2(G·H3(x t ))). (5)

[0060] As the parameters of the optimization function, according to the reference Figure 1 and Figure 2 The method described above processes an ordered sequence of possible values ​​of α and β. Δt defines the size of the update at each iteration, which may also be referred to as the "learning rate" in this article. Δt can also be regarded as a parameter of the optimization algorithm, or alternatively a fixed value can be set for Δt. If the standard gradient descent defined in equation 5 above is used, then Figure 3 The illustrated pre-processing 302 and exploration phase 304 may generate an ordered sequence of parameter values ​​for α and β.

[0061] However, many improved gradient descent techniques use variations of the above update schemes to speed up the algorithm's convergence to a near-optimal solution. The function F over which the gradient is computed at different states may have a complex structure, e.g. with many local minima. Variations of gradient descent have been developed that adjust the updates with the goal of converging faster on an estimated minimum.

[0062] One family of gradient descent methods uses the concept of momentum. Momentum is used in gradient descent to overcome the problem that depending on the exact form of the gradient, the update rule in equation (2) may converge slowly due to fluctuations from one iteration to the next, and cause the state to "bounce" around the optimal value rather than moving continuously towards it. The momentum gradient descent method is designed to overcome this problem by introducing a "momentum" term into the update rule in equation (2), resulting in the following update rule:

[0063] Substituting into the above equation 3 given Expression, the update rule can be expressed as follows: x t+1 =x t -Δt·(α·H1(x t )-β·H2(G·H3(x t )))+γ(x t -x t-1 ). (6a)

[0064] This corresponds to the dynamics of the differential equation:

[0065] The additional term γ(x t -x t-1 ) reduces bouncing around the state by pushing updates in the direction of recent updates, with the effect that updates have a tendency to continue along the path they have already taken, and thus become less prone to "bounce" due to fluctuations in the gradient. The parameter γ is typically less than 1, since the dependence of a given state on previous states should weaken with the number of iterations between the current state and the earlier state. In other words, when the above update rule is applied iteratively, the relationship between the current state and the earlier state depends on γ n , γ n should decrease with n so that a given state depends less on states that are further away in time than on the most recent state.

[0066] The second variation of momentum is the Nesterov momentum, where the update rule is given by the following equation:

[0067] This also corresponds to the differential equation (7), however, the Nesterov momentum update evaluates the point x where the momentum term has been applied t +γ(x t -x t-1 ), rather than evaluating the gradient at the most recent state x t It should be understood that either the standard momentum defined in Equation 6 or the Nesterov momentum update in Equation 8 can be used as the optimization algorithm for a given combinatorial optimization problem, and different choices may be applicable to different optimization problems and associated functions F.

[0068] All elements of the vector x are updated at each iteration of the update scheme in Equations 5, 6, and 8 using a learning rate defined by the time step Δt (i.e., the size of the update at each iteration). A good value for Δt is one that is large enough so that the update moves relatively quickly toward the optimum, but small enough so that the step does not "overshoot" the minimum due to large updates.

[0069] When the expression of equation (3) is substituted into equation 6 above, the accelerated optimization algorithm update becomes: x t+1 =x t -Δt·(α·H1(x t )-β·H2(G·H3(x t )))+γ(x t -x t-1 ). (9)

[0070] The expression is similar for the Nesterov momentum, except that, as above, the gradient is in the adjusted state xt +γ(x t -x t-1 ) is assessed at:

[0071] Either or both of the above-described accelerated optimization algorithms may be used with the methods described herein to generate an optimal system state.

[0072] Other accelerated gradient descent algorithms can be used for optimization algorithm 102. For example, a family of algorithms use adaptive learning rates to adjust updates over time and accelerate the state to the optimal state faster than for standard gradient descent methods.

[0073] The adaptive moment estimation method (Adam) is an extension of gradient descent that uses gradients that decay exponentially in time and squared gradients:

[0074] Here, the exponential moving average is used to estimate the average gradient m t and the mean squared gradient v t The effective learning rate varies in the iterative update rule (11) and depends on the estimated squared gradient magnitude. Compared to the momentum gradient descent methods considered previously, two additional parameters ξ1 and ξ2 need to be optimized for the Adam method. Therefore, in the case where the optimization algorithm uses the Adam method, each parameter value set includes the values ​​of ξ1, ξ2 and the function Parameters such as α and β.

[0075] Each of the above optimization algorithms will At least α and β and optionally additional parameters γ, Δt, etc. are used as parameters. These parameters are the set of parameters for which ranges are determined and values ​​are searched in the preprocessing phase and the exploration phase, respectively. These steps will now be described.

[0076] The following describes a preprocessing step for determining the range and parameter values ​​of α, β, γ, where the optimization algorithm uses a momentum update scheme such as Equation 6 or 8. However, as will be clear to those skilled in the art, the same method can be applied to search the space of any number of parameters. However, it should be noted that for each additional parameter of the optimization algorithm, the space increases exponentially, and therefore it becomes more difficult to identify a sequence of good sets of parameter values, so the inventors have recognized that a preferred embodiment uses an optimization algorithm with a small number of parameters (such as three or four parameters). In the preprocessing stage, the goal is to determine the range of values ​​to be searched, which may result in good performance of the optimization algorithm for at least some of the parameter values ​​selected from the range.

[0077] As mentioned above, G ijThe proposed method for determining the range involves the use of a graph or matrix G representing the problem ij Perform eigenvector analysis. An eigenvector of a matrix A is a vector v that satisfies the following equation: Av=λv For some scalar value λ, where λ is called the eigenvalue of A corresponding to the eigenvector v. An eigenvector is changed only by a scalar factor when transformed by the matrix A. Eigenvalues ​​and eigenvectors are well-known concepts in linear algebra and will not be described in further detail herein.

[0078] The k largest eigenvalues ​​of the graph G of the given Ising problem are calculated, where k is a predetermined number. The eigenvectors x corresponding to these largest eigenvalues ​​are mapped to binary valued states s, and the energies of these states are calculated according to Equation 1, where the lowest energy of these states is used as a trivial energy threshold.

[0079] To determine the bounds from the extraordinary energy thresholds, a linear stability analysis is performed on the optimization algorithm update formula, where the eigenvalues ​​correspond to the energy thresholds determined from the eigenvector analysis. Linear stability is used to determine that the static solution of the differential equation is the relationship between the α and β parameters, i.e., corresponding to all variables x t is zero or a non-trivial stationary state, i.e., produces a non-zero variable x t This can be applied to determine regions in parameter space where the energy is below the trivial energy threshold associated with the eigenvalues, which is the region of interest in Ising model optimization since the goal is to find the lowest energy state. The above analysis is for parameters α, β in a similar optimal unit range (e.g., ∈[0,1]). new Any input graph G of γ and γ provides the appropriate scaling terms in the iterative update rule.

[0080] Once ranges of parameter values ​​have been determined for each parameter to be optimized, the values ​​within these ranges are implemented as parameters of the optimization algorithm 102 in the exploration phase 304. Using an accelerated optimization algorithm with parameters α, β, and γ, the ranges from the preprocessing phase 302 define a three-dimensional space within which any given point defines a triplet of parameter values ​​to be replaced as parameters of the optimization algorithm. The goal of the exploration phase is to efficiently explore this three-dimensional space and evaluate the expected performance of the optimization for each point sampled from the space so that a subset of "good" parameters can be passed on for more extensive investigation in the "deep search" phase 306.

[0081] To illustrate one possible approach to exploring parameter space, FIG4a shows an example method for searching parameter space in 2D (e.g., where the value of γ is fixed and the optimization algorithm depends only on parameters α and β). As shown on the left hand side of FIG4, parameter α∈[α min ,α max ] and β∈[β min ,β max ] is shown as a rectangular region 400 in a two-dimensional space. Although the figure shows the parameter β, this is equivalent to searching for β within an appropriate range. new , which is scaled by the eigenvalue λ as described above to constrain the search to an appropriate range. Any point sampled from the region defines a pair of values ​​α and β within the determined range. In the exploration phase, a set of parameter value combinations covering the region can be sampled by first sampling the center point 402 of the region, and then by splitting the region into a grid of four sub-regions 404 and sampling the center point of each of these sub-regions 404. Each sub-region can be divided into four additional sub-regions, and the center point of each additional sub-region can be sampled. For each sampled point in the parameter space (i.e., each sampled parameter value pair), the optimization algorithm is run on a small number of iterations (e.g., 100-200) and a small number of initial states (e.g., 10-20) to generate a set of updated states. A performance metric is measured for the set of updated states generated at each sampling point (e.g., the mean, median, or minimum energy of the system for the updated states). This allows each sampling point to be labeled with an estimated performance of the optimization for that set of parameter values. This process is repeated until enough parameter combinations have been sampled to get a reasonable idea of ​​the performance across region 400 of parameter space.

[0082] As shown in Figure 4b, once the parameter value set has been evaluated, the area can be marked with estimated performance based on the results of the performance metric in the area. This can take the form of a heat map, where the shading of each pixel of the heat map is based on the performance metric of nearby points. A simple diagram is shown in Figure 4b, where the area is divided into two based on a certain threshold of the performance metric, where the area 408 with parameter scores above the threshold is shown as a horizontal line and the area where the parameter value set score is below the threshold is shown as a diagonal line. As described above, the deep search stage 306 is applied to a subset of the best parameter value sets identified in the exploration stage 304. The size of the subset may be constrained to a size L due to limitations in computing resources or time. The L parameters with the highest performance scores are selected from the area 400 and passed in descending order from the highest performance as an ordered sequence to be processed in the deep search stage.

[0083] The deep search phase is as follows Figure 1The optimization algorithm 102 is performed as described in , where the optimization algorithm 102 is applied to each parameter value set of the sequence designated as parameters of the algorithm. The optimization is run for a predetermined number of iterations on a predetermined number of initial states. These quantities are determined based on the desired processing budget for the deep search phase, for example a typical number of iterations, i.e. the number of times the state is updated for a given initial state and parameter value set, may be in the range of 1000 to 10000, and for each combination of parameters a typical number of initial states may be in the range of 100-1000.

[0084] In some embodiments, the deep search phase can be ended once the processing budget has been reached. It should be noted that "processing budget" as used herein includes restrictions on computing resources as well as restrictions on time. For example, the deep search phase can be ended once a certain number of operations have been performed, or after a predetermined amount of time is required to provide a solution. In this case, although the complete parameter sequence passed from the exploration phase has not yet been processed, an estimated optimal state is selected from the states that have been calculated in the deep search phase. In this case, since the deep search phase processes each parameter value set of the sequence in a given order, the sequence is preferably sorted according to performance as described above. However, in other embodiments without the above-mentioned exploration phase, parameter value sets can be randomly generated from within the parameter search space and passed to the deep search phase in a random order.

[0085] It should be noted that Figure 3 4 and the corresponding description relate to specific embodiments of the present invention, and other alternatives are possible. In some embodiments, instead of performing as for Figure 3 In the spectral analysis described in the preprocessing stage 302 of the present invention, in order to limit the range to be searched for each parameter, an initial set of boundaries can be initially defined and dynamically updated based on the computational performance of the solutions within the boundaries compared to the solutions for the parameters outside the current boundaries. In other words, for a given α when the boundaries are [α min ,α max ], by increasing the margin to, for example, 10*α maxAnd evaluating solutions for increasing parameters within the bounds, it can be determined that a solution outside the current bounds is better than a solution within the current bounds, and in this case the bounds are increased. This process can be repeated in multiple iterations until a suitable region of parameter space is found. An ordered sequence of sets of parameter values ​​can be randomly generated by randomly sampling values ​​from within the range of each parameter using existing random sampling techniques (e.g., Monte Carlo random sampling within a region), where the region or range of each parameter is determined, for example, by spectral analysis or dynamic adjustment of a set of bounds. Alternatively, a brute force grid search can be performed on all parameters within a given range. The grid search performs an exhaustive search through the parameter region defined by the range of each parameter, exploring each point of a grid constructed within the region of the parameter space to be searched.

[0086] The Bayesian approach creates a probabilistic model for the objective function and focuses on the most promising parameter values ​​for evaluating the objective function. The Bayesian approach performs a more targeted search than either the random search or grid search approaches, as both random search and grid search techniques, while being direct and fast to implement, ignore results achieved for previously seen parameters, and may spend significant computational resources evaluating poor parameter regions. In another alternative, the Bayesian approach may be used alone or in addition to the parameter search process described in FIG. 4 to determine a sequence of sets of parameter values ​​to provide further investigation of the deep search phase. However, the random and grid search techniques are faster than the more targeted search of the Bayesian approach, and therefore in a preferred embodiment, random search and / or grid search are used alone or in combination with the Bayesian approach.

[0087] In addition, the optimization algorithm can take various forms, including non-gradient based optimization algorithms. As mentioned above, derivative-free methods such as simulated annealing and particle swarm optimization can be used. Any parameterized optimization algorithm can be used in conjunction with the methods described herein for searching parameter space and determining the optimal state of the system.

[0088] Furthermore, while the above example describes a method for solving the Ising problem, the above techniques can be used to solve other combinatorial optimization problems where the optimization algorithm is appropriate for the given problem. The β term in represents the gradient of the new optimization problem. Implementation in hardware

[0089] The above method can be implemented on one or more hardware processors. In some embodiments, the optimization method can be provided via a cloud service, where a user can input a given optimization problem via a web interface, which is communicated to a set of one or more hardware processors located remotely from the user via a network, and the set of one or more hardware processors is configured to implement the above optimization method and generate an estimated solution that is communicated back to the user via the network. The processor may include a CPU and / or a GPU. As described above, either or both of the exploration phase and the deep search phase can be performed simultaneously and in parallel for multiple parameter value sets.

[0090] In some embodiments, the above methods are implemented using an optical computer comprising optical hardware. The optical computer is configured to implement arithmetic and logic operations using optical components, including but not limited to lasers, modulators, photodetectors, lenses, and diffraction gratings to process optical signals, as described below with respect to Figure 5A , Figure 5B and Figure 6 Described. The optical computer may additionally include analog and / or digital electronic hardware to convert signals between the optical domain, the analog domain, and the digital domain. In the optical solver, optical signals are used to represent input variables (e.g., "spin" in the Ising problem), and optical elements are used to combine signals in a manner that models the interactions between the variables (e.g., the above-mentioned matrix G). Optical elements that perform vector-vector multiplication in the optical domain (such as liquid crystal displays or ring resonators) are known in the art. Optical hardware for performing arithmetic operations such as vector multiplication is described in European patents with application numbers 21155439.9, 21155447.2, and 21155428.2.

[0091] Figure 5A A schematic block diagram of the functions performed by an optical computer for implementing the combinatorial optimization method described above is shown. In particular, optical hardware is used to model the values ​​of the variables of a given combinatorial optimization problem in numerical properties of optical signals (such as phase or amplitude), and one or more optical components are used to apply an optimization algorithm parameterized by a set of parameter values, such as those described above, to the variables in order to generate an optical representation of the optimal state of the system modeled by the given optimization problem. The optical signal may be converted between the optical domain and the digital domain one or more times during the calculation, as will be described below with reference to Figure 5A and Figure 5B Further described.

[0092] As described above, the present invention provides a method for determining an approximate solution to a combinatorial optimization problem using a parameterized optimization algorithm, by sequentially processing a list of possible parameter values, applying an optimization algorithm having parameters of an optimization algorithm set to the corresponding parameter values ​​of the list to generate candidate solutions before evaluating the candidate solutions according to an optimality metric, and selecting the solution with the highest optimality score as the best solution.

[0093] Figure 5A The implementation of the above-mentioned optimization method in an optical computing system for a given optimization algorithm and a list of possible parameter value sets to be applied in the optimization algorithm to generate candidate solutions are shown. The spin generator 502 generates an optical signal representing a value set of variables for the problem, i.e., an initial state. The spin generator includes a light source and may include one or more modulators that encode numerical values ​​in the properties of the light. The optical signal may be phase modulated, for example, so that each variable has an associated signal whose phase is related to the value of the variable. Alternatively, the optical signal may be amplitude modulated so that the amplitude of the light wave for each variable is directly related to the value of the variable. In a preferred embodiment, amplitude modulation is used to model the variable values ​​because the detection of phase modulated signals requires more hardware complexity. Amplitude and phase modulators are well known in the field of optical computing and will not be further described in this article. The spin generator 502 outputs a set of modulated optical signals representing the initial state of the system modeled by the given combinatorial optimization problem. This is received by an optical optimization update component 504 of optical hardware including arithmetic logic for implementing the optimization algorithm update step. Each signal generated by the spin generator may have a different wavelength. In a preferred embodiment, a different initial state for each parameter value set is used. These different states (e.g., randomly generated states) are passed to a spin generator in the digital domain, and these states are converted into a modulated optical signal that encodes the initial value. In other embodiments, the same initial state can be used for multiple parameter sets.

[0094] Figure 5A The configuration shown in can be used to generate a "spin" or signal representing a single state at any given time, or alternatively, multiple states can be processed in parallel by the optimization algorithm as long as the hardware is configured to provide enough signals to represent more than one state at a time. Although the following description outlines the process for updating a single state, it will be clear to those skilled in the art how to implement the same technique for a hardware configuration capable of processing multiple states in parallel.

[0095] As described above, for various update algorithms, the algorithm is parameterized by a set of parameter values. These are provided by a parameter controller 500, which passes the parameter value set to the optimization algorithm hardware in the digital domain. Digital-to-analog conversion is performed to convert a digital signal representing the value of the parameter for the algorithm into an optical signal that can be processed by the optical hardware of the update component 504. The hardware includes one or more components for performing arithmetic operations on the optical signal, such as a vector multiplier and / or detector for performing addition of a set of modulated signals. The hardware used to implement the algorithm update depends on the form of the update itself. Reference is made below to Figure 5B An example implementation including an optical component for implementing an update of the form described above in Equation 6a is shown and described in more detail. Updater 504 performs iterative updates according to a specified algorithm, such as gradient descent or gradient descent with momentum, wherein a signal output at each iteration is converted by optical analog converter 506 to an analog signal that is passed back to the spin generator as a feedback signal, thereby providing a signal representing the input state for the next iteration. As described above, this is repeated over multiple iterations to allow the solver to converge to an optimal solution. Optical analog converter 506 can be, for example, a photodetector that converts an optical signal into an electric current, wherein there is a known relationship between the light property in which the variable value is encoded and the electric current.

[0096] The signal output by the optical analog converter 506 can be periodically sampled and converted to a digital signal by an analog-to-digital converter. The solution is sampled so that the solution is periodically evaluated relative to the optimality metric (e.g., by calculating the total energy of each sampled solution) and the best solution is determined. Once the optimization is terminated, the candidate solution 510 takes the form of a digital representation of the optimal state generated by the algorithm, which can be interpreted based on the definition of the problem and used to configure the real world system. Each parameter value set of the parameter value set list provided to the parameter controller 500 is applied to a plurality of initial states generated by the spin generator to generate a plurality of candidate solutions. Once a candidate solution set is generated for a given parameter value set, the parameter controller 500 selects the next parameter value set and applies the optimization algorithm to a new initial state set generated by the spin generator 502 using the new parameter values.

[0097] For example, in the case where the model is a collection of chemical compounds, where the optimization problem is to discover new drugs that are effective and safe, the numerical output of the system may be a collection of numbers representing the positioning of specific molecules or atoms in the candidate compounds. It should be noted that the variables processed by the algorithm may not be direct variables of the model, for example, algorithm updates may be applied to real-valued numbers, while the state of the model is binary, with each variable taking only the values ​​0 and 1. In this case, the values ​​output by the algorithm are mapped to the variable types of the given model before outputting the solution. Note that there is a direct mapping from the values ​​processed by the algorithm and the variables of the model, and references to values ​​in this article refer to both the values ​​processed by the algorithm and the variables of the system.

[0098] Figure 5B An optical computer for implementing an optimization algorithm of the form described above in Equation 6a is shown. Figure 5A As described, the spin generator 512 generates an optical signal representing a vector x of values ​​that maps to the initial state of the system being modeled. This is processed by the optical optimization algorithm updater 504, which in this example includes three components: an optical momentum component 512, a spin-spin interaction component 514, and an optical annealer 516.

[0099] The optical momentum component implements the momentum term γ(x t -x t-1 ). This is the signal vector x from the current iteration t Subtract the vector x from the signal of the previous iteration t-1 , followed by a scalar multiplication. The addition of the optical signals can be performed by detecting the aggregate signal at a photodetector before converting the resulting analog signal back to the optical domain. For amplitude modulated light of different wavelengths, for example, multiple signals are simply added by detecting the signals together at a photodetector, the intensity of the resulting combined signal being equal to the sum of the intensities of the individual signals. Subtraction can be performed by adding the positive and negative signals at separate detectors to generate two electrical signals corresponding to the positive and negative terms of the subtraction, and subtracting the two terms in the electrical domain. A scalar multiplication can be performed on the resulting optical signal using an attenuator or amplifier to multiply the amplitude (or intensity) by the corresponding factor given by the optimization algorithm.

[0100] The spin-spin interaction component implementation contains the matrix G (i.e.: -Δt·β·H2(G·H3(x t ))) of Equation 6a. Different optical hardware can be implemented to evaluate different possible functions H2, H3. For example, where H2 and H3 are linear functions, these functions can be implemented as scalar multiplication and addition operations on the corresponding vectors of the signals to which they are applied. Nonlinear functions are performed by specific optical components or combinations of optical components (e.g., photodetectors) that are known to produce or approximate nonlinear functions. Such optical components are known to those skilled in the art of optical computing. The operation G·H3() is a matrix-vector multiplication and can be implemented in various ways. Referring to Figure 6 In one configuration described and further detailed in European Patent Application Nos. 21155439.9, 21155447.2 and 21155428.2, a wavelength selective switch configuration may be implemented to perform a vector-matrix multiplication, wherein each element of the input vector is associated with an optical signal having a different wavelength.

[0101] Finally, the optical annealer 516 uses the above optical components to achieve the remaining x t -αΔtH1(x t ) terms for addition when necessary and for scalar and vector multiplication.

[0102] The signals generated by the three components of the optical optimization algorithm updater are combined to generate a signal representing the next state of the system, consistent with Equation 6a. These signals can be combined into a single beam, as indicated by the summation symbol, and this can be detected at a photodetector that performs the conversion from the optical domain to the analog domain. As described above with reference to Figure 5A As described, the algorithm is iteratively executed, and the output state is passed back to the spin generator 502 as a feedback signal after the optical analog conversion, and the spin generator 502 generates an optical signal representing the updated state to be processed again by another algorithm update step. The output state is periodically sampled and converted to the digital domain by the converter 508. The parameter controller 500 sets the values ​​of the parameters of the optimization algorithm, which are α, β and γ in this example. As described above, the parameter controller selects each parameter value set from the parameter value set list in turn and inputs the parameter value to the corresponding optical element implementing the algorithm. The algorithm is executed for multiple initial states to generate a set of multiple candidate solutions. Once the algorithm is executed on a sufficient number of initial states, the parameter controller 500 selects the next parameter value set from the list and sets the parameter value for the optical component implementing the corresponding item of the update step of the algorithm. The number of initial states can be determined in advance, or the number of initial states can be dynamically determined based on available resources. For example, after a certain amount of time for implementing the optimization algorithm for a given set of parameter values, parameter controller 500 may be configured to update the algorithm parameters with the values ​​for the next set of values ​​in the list and begin processing the initial state for that set of parameter values.

[0103] Note that the above description of the optical computer system describes its use of a deep search phase, i.e., a list of parameter values ​​with which an optimization algorithm is to be applied has been determined, and wherein a best solution is output from a set of solutions generated by the optimization algorithm parameterized by one of the parameter value sets on the list. However, as described above, the optimization algorithm may be implemented in an exploratory phase in order to determine, from a wide range of possible parameter values, a shorter list of parameter value sets suitable for further exploration in the deep search phase. In this embodiment, the optimization algorithm is first implemented for a large number of possible parameter values ​​within the parameter space, but is implemented on a smaller number of initial states for each parameter value set and for a smaller number of iterations, such that a "solution" indicates the quality of the algorithm for that parameter value set, and the results of this exploratory optimization phase are used to generate a list of parameter value sets to be implemented on a large number of initial states and for more iterations of the optimization algorithm. In this embodiment, in Figure 5A and Figure 5BThe hardware configuration shown in is used in the exploration phase to generate and sample initial solutions 508 evaluated relative to optimality metrics (such as energy functions). The parameter controller 500 uses the evaluated solutions to narrow the set of parameter value sets to determine a smaller set of candidate parameter value sets for the optimization algorithm, and the controller sets these values ​​in hardware 504, which is configured to implement the optimization algorithm for deeper exploration in the deep search phase, as described above. Therefore, Figure 5B The use of an optimization algorithm over multiple stages is shown, with the results of the exploration stage fed back into the implementation of parameters in the deep search stage, as shown by the arrows between the solution sampler 508 and the parameter controller 500. However, as mentioned above, the parameter list can be obtained by other methods, and in this case, the hardware is configured to implement a fixed set of parameter values ​​so that the solution output for a given set of parameter values ​​is not used to inform future parameter values ​​of the optimization algorithm. This situation is not shown in FIG. Figure 5A and can be used to implement specific examples of optimization algorithms (such as Figure 5B The configuration 504 shown in FIG. 5 is implemented by any specific hardware configuration.

[0104] Note that the same optical computer can be used to implement both the "deep search" phase and the "exploration phase" described above, where parameters are selected randomly or by some search technique over a predefined parameter space, and the optimization algorithm is run for each larger set of parameter value combinations, but each optimization is performed for fewer iterations and a smaller number of initial states than when processing a determined list of sets of parameter values ​​in the "deep search" phase described above.

[0105] Figure 6 An example wavelength selective switch configuration that can be implemented in an optical computing system for performing vector-matrix multiplication is shown, which can be used to implement the spin-spin interaction (also called variable interaction) term of the optimization algorithm. The wavelength selective switch includes an input array 908 of light sources generating a set of signals of different wavelengths, a lenslet array 900, a modified spatial light modulator 902, a dispersive element 814 (e.g., a diffractive element), and an output array of optical signals detected at a photodetector array, not shown in FIG. 9 .

[0106] A spatial light modulator (SLM) includes a set of modulators configured to modulate incident light by different corresponding factors. Typically, a spatial light modulator includes multiple vertical columns with different factors associated with each column, so that the optical signal received at any point along a given column of the SLM is modulated by the same factor. In order to use the spatial light modulator for vector-matrix multiplication in this example solver architecture, the vertical axis of the SLM needs to provide different weights even for the same wavelength so that the entire function of the vector-matrix multiplication is realized. This is because, for matrix multiplication, the input vector needs to be multiplied by each row of the matrix Q to generate a complete output vector. SLM 902 is a modified version of a typical SLM (each column of which includes a modulation factor), in which the modulator array is arranged in an array, in which the loss applied by each modulator reflects the weight of the matrix to be applied to the input, that is, the rows of the modified SLM encode the weights in the rows of the matrix. Therefore, each input signal needs to be processed to be spread vertically so that they hit each row of the SLM 902, corresponding to a series of vector-vector multiplications.

[0107] In this example, input array 908 includes vector H3(x t ) of the signal. This vector is passed through a lenslet array 900 with a specific geometry that spreads the signal vertically while collimating the beam corresponding to the wavelength of the signal in the horizontal direction of the SLM 902. This allows more input signals of different wavelengths to be processed at a single SLM. The lenslet array improves the collimation properties of the beam in both directions.

[0108] In contrast to a typical SLM, SLM 902 comprises a 2D array of modulators, each element of the array applying a corresponding weight of the matrix G to the received input signal, whereas a typical SLM does not require elements in the same "column" but at different vertical positions to have different values. Each weighted signal modulated at a particular wavelength but with a different weight (i.e., each signal of a column of SLM 902) needs to bounce off dispersive element 814 at a different vertical position to ensure that the combination of the WDM signals appears at the right photodetector in the photodetector array 904. The dispersive element may be implemented as a diffraction element, such as a diffraction grating or a prism. The output signal may be directed from element 814 via one or more lenses to direct the signal into a beam at the correct vertical height to be monitored using incoherent addition at the photodetector corresponding to the output vector element represented by that beam.

[0109] Figure 6The wavelength selective switch configuration shown in is only one example of an optical hardware configuration suitable for implementing the variable interaction term of the optimization algorithm. Other components such as ring resonators and Mach-Zehnder interferometers may alternatively be used to implement vector multiplications. Matrix-vector multiplications may be decomposed into a collection of vector-vector multiplications, each of which is implemented by one of the above components.

[0110] It should be understood that the above embodiments are disclosed only as examples.

[0111] More generally, according to one aspect disclosed herein, a computer-implemented method is provided, comprising: receiving a plurality of candidate parameter value sets of an ordered sequence of candidate parameter value sets in an order specified in the ordered sequence, each of the plurality of candidate parameter value sets comprising a corresponding candidate parameter value for at least one parameter of an optimization algorithm, wherein the number of received candidate parameter value sets is based on a processing budget; for each candidate parameter value set in the sequence: applying the optimization algorithm to a plurality of initial states of a model representing a system, wherein at least one parameter of the optimization algorithm is set to a corresponding candidate parameter value, thereby generating a corresponding plurality of candidate update states of the model, and evaluating the plurality of candidate update states according to an optimality metric, the evaluation comprising evaluating each candidate update state according to the optimality metric to generate a corresponding optimality score for the candidate update state, such that different candidate update states among the candidate update states have different optimality scores; selecting the candidate update state with the highest corresponding optimality score as the estimated optimal state of the model; and outputting the selected estimated optimal state of the model to one of: a user interface, a computer interface, a computer storage device, or a destination application.

[0112] The candidate parameter value sets may be received from user input, computer memory, a network interface, or from an external application or device configured to generate parameter value sets.

[0113] Outputting the selected estimated optimal state to a user interface can include providing the state in a graphical user interface implemented on a computer display, or by generating an auditory output of the state interpretable by a user or other system. The computer storage device can be one of the following: non-volatile computer memory, such as a hard drive, a solid-state drive, or a flash drive; or a temporary computer storage device, such as a random access memory. The destination application can be a computer-implemented application, such as a modeling application or a computer-aided design application. The user or application can issue controls or actions to implement the state in the real-world system directly or by controlling a tool that is operable to create or modify the state in the real-world system. The estimated optimal state can also be output to a device external to the system on which the optimization method is executed.

[0114] In an embodiment, at least one parameter of the optimization algorithm is a plurality of parameters of the optimization algorithm, wherein each candidate parameter value set in the plurality of candidate parameter value sets includes a plurality of candidate parameter values, each candidate parameter value in the candidate parameter value sets is a corresponding value of a corresponding parameter of the optimization algorithm, and wherein the step of applying the optimization algorithm to a plurality of initial states includes applying the optimization algorithm having the plurality of parameters of the optimization algorithm set to the corresponding plurality of candidate parameter values.

[0115] In an embodiment, the received candidate parameter value sets are selected as a subset of the top k parameter value sets of the ranked list of parameter value sets, where k is a number selected based on a processing budget.

[0116] In an embodiment, the sorted list of parameter value sets is sorted according to a performance metric of the optimization algorithm, wherein for each respective parameter value set in the sorted list, the performance metric measures an expected performance of the optimization algorithm if parameters of the optimization algorithm are set to the respective parameter value set.

[0117] In an embodiment, the optimization algorithm is an iterative optimization algorithm comprising an update rule for generating a state for a next iteration when applied to a state for a current iteration, wherein for each set of parameter values ​​in a sequence, the application of the optimization algorithm comprises iteratively performing an update by applying the update rule to an initial state at a predefined full depth to obtain a candidate updated state, wherein the full depth is the number of iterations to which the update rule is applied.

[0118] In an embodiment, the step of determining a ranked list of parameter value sets includes: sampling a plurality of parameter value combinations from within a predetermined range for each parameter; for each combination: applying a modified optimization algorithm to a set of exploration initial states of the system to generate a set of exploration update states, evaluating each exploration update state according to an optimality metric to generate a corresponding optimality score; and determining a performance metric, the performance metric comprising a representative measure of the optimality scores aggregated over the exploration update states; sorting the combinations according to the calculated performance metric; wherein the modified optimization algorithm comprises iteratively applying an update rule at a partial depth, the partial depth being the number of iterations at which the update rule is applied by the modified optimization algorithm, wherein the partial depth is less than the full depth and / or the set of exploration initial states is less than a plurality of initial states.

[0119] In an embodiment, the model comprises a matrix, and wherein the range for each parameter is determined by performing a linear stability analysis on selected eigenvalues ​​of the matrix.

[0120] In an embodiment, the eigenvalues ​​are selected by calculating a predetermined number of maximum eigenvalues ​​and calculating optimality scores for corresponding eigenvectors, and selecting the eigenvalue corresponding to the eigenvector yielding the highest optimality score from the calculated optimality scores.

[0121] In an embodiment, the processing budget includes constraints on one or more of: processing time, memory consumed, number of processing cycles executed, or processing energy consumption; the number of candidate parameter value sets received is the number that can be processed within the constraints of the processing budget.

[0122] In an embodiment, a system is represented by a graph of interactions between pairs of variables that define the system, and each state of the system is an assignment of values ​​to the variables of the system.

[0123] In an embodiment, the optimality score is calculated by evaluating an energy function of the model, which represents the total energy of the physical system represented by the model.

[0124] In an embodiment, the optimization algorithm is a gradient descent algorithm with momentum.

[0125] In an embodiment, the system represented by the model is one of: a computing system, an electronic system, an electrical or electromagnetic system, a mechanical system, a chemical system, or a therapeutic application to a biological being, and wherein the method further comprises implementing the estimated optimal state in the system.

[0126] Another aspect disclosed herein provides a method for optimizing a state of a system represented by a matrix G, the method comprising: a. Receive an input vector x representing the state of the system at an initial time step t = 0 o , and the vector is iteratively updated at each successive time step t according to an update formula of the following form: bx t+1 =x t -Δt·[α·H1(x t )-β·H2(G·H3(x t ))]-γ(x t -x t-1 ) where H1, H2, H3 are linear or nonlinear functions (the functions are applied independently to the vector x t Each element of x t Different functions can be used for different elements of ), α is the dissipation rate, γ is the momentum parameter, and β is a parameter that defines the relative contribution of the system energy function to the update.

[0127] According to another aspect of the present disclosure, a computer-implemented method is provided, comprising: receiving a plurality of candidate parameter value sets in an order specified in an ordered sequence of candidate parameter value sets, each candidate parameter value set comprising corresponding one or more candidate parameter values ​​for one or more parameters of an optimization algorithm, each candidate parameter value being a value of a corresponding one of the one or more parameters, wherein the number of candidate parameter value sets is based on a processing budget; for each response parameter value set in the sequence: applying the optimization algorithm to a plurality of initial states of a model representing a system, wherein one or more parameter sets are set to corresponding one or more candidate parameter values, thereby generating a corresponding plurality of candidate update states of the model, and evaluating each candidate update state according to an optimality metric, thereby generating an optimality score; selecting the candidate update state corresponding to the highest optimality score as an estimated optimal state of the model; and outputting the selected estimated optimal state of the model to one of: a user interface, a computer storage device, or a destination application.

[0128] In an embodiment, the method may further comprise steps according to any of the methods described herein.

[0129] According to one aspect of the disclosure, there is provided a computer-implemented method comprising: receiving a plurality of candidate parameter value sets in a specified ordered sequence relative to one another, each candidate parameter value set of the plurality of candidate parameter value sets comprising a respective candidate parameter value for at least one parameter of an optimization algorithm, wherein the number of the candidate parameter value sets received is based on a processing budget; For each candidate parameter value set in the ordered sequence: applying an optimization algorithm to each of a plurality of initial states of a model representing the system, wherein at least one parameter of the optimization algorithm is set to the respective candidate parameter value, thereby generating a corresponding plurality of candidate updated states of the model, and evaluating each of the candidate update states according to the optimality metric, thereby generating a corresponding optimality score for the candidate update state, such that different of the candidate update states have different optimality scores; selecting the candidate updated state with the highest corresponding optimality score as the estimated optimal state of the model; and The selected estimated optimal state of the model is output to one of: a user interface, a computer storage device, or a destination application.

[0130] The received candidate parameter value sets may be selected as a subset of the first k parameter value sets in the ranked list of parameter value sets, where k is a number selected based on the processing budget.

[0131] The sorted list of parameter value sets is sorted according to a performance metric of the optimization algorithm, wherein for each respective parameter value set in the sorted list, the performance metric measures an expected performance of the optimization algorithm if parameters of the optimization algorithm are set to the respective parameter value set.

[0132] The optimization algorithm may be an iterative optimization algorithm comprising an update rule for generating a state for a next iteration when applied to a state for a current iteration, wherein for each candidate parameter value set in an ordered sequence, the application of the optimization algorithm comprises: iteratively performing an update by applying the update rule to an initial state at a predefined full depth, thereby obtaining a candidate updated state, wherein the full depth is the number of iterations to which the update rule is applied.

[0133] The step of determining a ranked list of sets of parameter values ​​includes: sampling a plurality of parameter value combinations from within a predetermined range for each parameter of the optimization algorithm; For each combination: applying a modified optimization algorithm to each of a plurality of exploration initial states of the system to generate a plurality of exploration update states, wherein a parameter of the modified optimization algorithm is set to the combination, evaluating each of the exploration update states according to an optimality metric to generate a corresponding optimality score, and determining a performance metric comprising a representative measure of the optimality score aggregated over a plurality of exploration update states; and ranking the combinations according to the calculated performance measure; The modified optimization algorithm comprises iteratively applying the update rule at a partial depth, wherein the partial depth is the number of iterations at which the update rule is applied by the modified optimization algorithm, wherein the partial depth is less than the full depth and / or the set of explored initial states is less than a plurality of initial states.

[0134] The model may include a matrix, and wherein the range for each parameter is determined by performing a linear stability analysis on selected eigenvalues ​​of the matrix.

[0135] The eigenvalues ​​are selected by calculating a predetermined number of maximum eigenvalues ​​and calculating the optimality scores for corresponding eigenvectors, and selecting the eigenvalue corresponding to the eigenvector yielding the highest optimality score from the calculated optimality scores.

[0136] The processing budget may include constraints on one or more of: processing time, memory consumed, number of processing cycles performed, or processing energy consumption; the number of candidate parameter value sets received is a number that can be processed within the constraints of the processing budget.

[0137] The system may be represented by a graph of interactions between pairs of variables defining the system, and each state of the system is an assignment of values ​​to a variable of the system.

[0138] The optimality score may be calculated by evaluating an energy function of the model, which represents the total energy of the physical system represented by the model.

[0139] The optimization algorithm may be a gradient descent algorithm with momentum.

[0140] The system represented by the model is one of: a computing system, an electronic system, an electrical or electromagnetic system, a mechanical system, a chemical system, or a therapeutic application to a biological being, and wherein the method further comprises implementing the estimated optimal state in the system.

[0141] Another aspect disclosed herein provides a computer-implemented method for optimizing a state of a system represented by a matrix G, the method comprising: An input vector x representing the state of the system is received at an initial time step t=0 o , and the vector is iteratively updated at each successive time step t according to an update formula of the following form: x t+1 =x t -Δt·[α·H1(x t )-β·H2(G·H3(x t ))]-γ(x t -x t-1 ) where H1, H2, H3 are linear or nonlinear functions (the functions are applied independently to the vector x t Each element of x t Different functions may be used for different elements of ), α is the dissipation rate, γ is the momentum parameter, and β is a parameter defining the relative contribution of the system energy function to the update.

[0142] Another aspect of the present disclosure provides a computer system, including: a memory and one or more processors, or optical hardware, which is configured to implement any of the methods disclosed herein.

[0143] Another aspect disclosed herein provides a computer program for programming a computer system to perform steps according to any method or function disclosed herein.

[0144] Other variations or use cases may become apparent to those skilled in the art once given the disclosure herein.The scope of the present disclosure is not limited by the above-described embodiments but is only defined by the appended claims.

Claims

1. A computer-implemented method comprising: receiving a plurality of candidate parameter value sets in a specified ordered sequence relative to one another, each candidate parameter value set of the plurality of candidate parameter value sets comprising a respective candidate parameter value for at least one parameter of an optimization algorithm, wherein the number of the candidate parameter value sets received is based on a processing budget; For each candidate parameter value set in the ordered sequence: - applying the optimization algorithm to each of a plurality of initial states of a model representing a system, wherein the at least one parameter of the optimization algorithm is set to the respective candidate parameter value, thereby generating a corresponding plurality of candidate updated states of the model, and - evaluating each of the candidate update states according to an optimality metric, thereby generating a corresponding optimality score for the candidate update state, such that different ones of the candidate update states have different optimality scores; selecting the candidate updated state with the highest corresponding optimality score as the estimated optimal state of the model; as well as The selected estimated optimal state of the model is output to one of: a user interface, a computer storage device, or a destination application. 2 . The method of claim 1 , wherein the received candidate parameter value sets are selected as a subset of top k parameter value sets in a ranked list of parameter value sets, where k is a number selected based on the processing budget.

3. A method according to claim 2, wherein the sorted list of parameter value sets is sorted according to a performance metric of the optimization algorithm, wherein for each corresponding parameter value set in the parameter value sets in the sorted list, the performance metric measures the expected performance of the optimization algorithm when the parameters of the optimization algorithm are set to the corresponding parameter value set.

4. The method of any one of the preceding claims, wherein the optimization algorithm is an iterative optimization algorithm comprising an update rule to generate a state for a next iteration when applied to a state for a current iteration, wherein for each set of candidate parameter values ​​in the ordered sequence, the applying of the optimization algorithm comprises: The candidate updated state is obtained by iteratively performing an update by applying the update rule to the initial state at a predefined full depth, wherein the full depth is the number of iterations for which the update rule is applied.

5. A method according to claim 4 when appended to claim 3, wherein the step of determining the ordered list of sets of parameter values ​​comprises: sampling a plurality of parameter value combinations from within a predetermined range for each parameter of the optimization algorithm; For each combination: - applying a modified optimization algorithm to each of a plurality of exploration initial states of the system to generate a plurality of exploration update states, wherein the parameters of the modified optimization algorithm are set to the combination, - evaluating each of said exploration update states according to said optimality metric to generate a corresponding optimality score, and - determining a performance metric comprising a representative measure of the optimality score aggregated over the plurality of exploration update states; as well as sorting the combinations according to the calculated performance metric; The modified optimization algorithm comprises iteratively applying the update rule at a partial depth, wherein the partial depth is the number of iterations at which the update rule is applied by the modified optimization algorithm, wherein the partial depth is less than the full depth and / or the set of explored initial states is less than the plurality of initial states.

6. The method of claim 5, wherein the model comprises a matrix, and wherein the range for each parameter is determined by performing a linear stability analysis on selected eigenvalues ​​of the matrix.

7. The method of claim 6, wherein the eigenvalues ​​are selected by calculating a predetermined number of maximum eigenvalues ​​and calculating the optimality scores for corresponding eigenvectors, and selecting the eigenvalue corresponding to the eigenvector that produces the highest optimality score from the calculated optimality scores.

8. A method according to any of the preceding claims, wherein the processing budget includes constraints on one or more of: processing time, memory consumed, number of processing cycles executed, or processing energy consumption; and the number of the received candidate parameter value sets is a number that can be processed within the constraints of the processing budget.

9. A method according to any one of the preceding claims, wherein the system is representable by a graph of interactions between pairs of variables defining the system, and each state of the system is an assignment of values ​​to the variables of the system.

10. A method according to any one of the preceding claims, wherein the optimality score is calculated by evaluating an energy function of the model, the energy function representing the total energy of the physical system represented by the model.

11. The method according to any of the preceding claims, wherein the optimization algorithm is a gradient descent algorithm with momentum.

12. A method according to any of the preceding claims, wherein the system represented by the model is one of the following: a computing system, an electronic system, an electrical or electromagnetic system, a mechanical system, a chemical system, or an application to the treatment of a biological being, and wherein the method further comprises achieving the estimated optimal state in the system.

13. A computer-implemented method for optimizing the state of a system that can be represented by a matrix G, the method comprising: An input vector x representing the state of the system is received at an initial time step t=0 o , and the vector is iteratively updated at each successive time step t according to an update formula of the following form: x t+1 =x t -Δt·[α·H1(x t )-β·H2(G·H3(x t ))]-γ(x t -x t-1 ) where H1, H2, H3 are linear or nonlinear functions (the functions are applied independently to the vector x t Each element of x t Different functions may be used for different elements of ), α is the dissipation rate, γ is the momentum parameter, and β is a parameter defining the relative contribution of the system energy function to the update.

14. A computer system comprising: memory and one or more processors, or Optical hardware, It is configured to implement a method according to any of the preceding claims.

15. A computer program for programming a computer system to perform the steps of any one of claims 1 to 13.

Citation Information

Cited By

  • Noise-robust correction update method and system for optical Ising machine and medium

    CN123030035A