A feedback-type optimization system and method using an Ising model with machine learning-based target characteristic learning capabilities.
The feedback-type optimization system with Ising model and machine learning dynamically adapts to environmental changes and improves control accuracy by dynamically adjusting weights and learning target system characteristics, addressing flexibility and accuracy issues in conventional systems.
Patent Information
- Application Number
- JP2026000626
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2026-01-05
- Publication Date
- 2026-11-19
- Estimated Expiration
- 2046-01-05
AI Technical Summary
Conventional optimization systems struggle to flexibly respond to dynamic factors such as rapid environmental changes and equipment deterioration, leading to suboptimal performance, and lack mechanisms for systematically analyzing feedback data to improve accuracy.
A feedback-type optimization system using Ising model optimization with machine learning, which includes state monitoring, deviation determination, weight adjustment, and target characteristic learning to dynamically adapt and improve optimization accuracy.
Enables rapid adaptation to environmental changes, continuous improvement of control accuracy, and automatic learning of target system characteristics, enhancing robustness and cost-effectiveness.
Smart Images

Figure 0007953871000001_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a feedback optimization system that optimizes a target system based on data collected by various sensor groups, and in particular, dynamically adjusts optimization parameters based on the deviation between the current state and the target state, leverages the fast search capability of Ising model optimization to perform optimization efficiently, analyzes accumulated feedback data using a machine learning model to learn the unique characteristics of the target system, and continuously improves the accuracy of weighting parameters in subsequent optimization, relating to a feedback-type Ising model optimization system and method.
[0002] The present invention is widely applicable to energy management, manufacturing process optimization, logistics and supply chain optimization, traffic control, communication network optimization, investment portfolio optimization, material design, and combinatorial optimization in various fields.
[0003] As used herein, "Ising model optimization" refers to a general term for methods that convert a combinatorial optimization problem into the Ising model or QUBO (Quadratic Unconstrained Binary Optimization) format, and search for solutions at high speed using dedicated hardware or software. Specifically, it includes quantum annealing machines (such as those manufactured by D-Wave Systems), simulated annealing, Digital circuit type annealing machine (manufactured by Fujitsu Limited, etc.) , CMOS annealing machines (such as those manufactured by Hitachi, Ltd.), coherent Ising machines (such as those manufactured by Nippon Telegraph and Telephone Corporation), and Ising machines having functions equivalent to these.
Background Art
[0004] In various sectors of industry, there are combinatorial optimization problems that involve optimizing an objective function under multiple constraints. Examples include optimizing air conditioning, lighting, and power in building and factory energy management; optimizing production scheduling and equipment utilization in manufacturing; optimizing delivery routes and inventory placement in logistics; optimizing signal control and vehicle dispatch in transportation systems; optimizing bandwidth allocation and routing in communication networks; and optimizing portfolio composition in investment.
[0005] Conventional optimization systems primarily relied on pre-set schedules and rule-based control. However, they had the problem of being unable to flexibly respond to dynamic factors such as rapid changes in the external environment, fluctuations in demand, and equipment deterioration over time, making it difficult to maintain an optimal state.
[0006] In recent years, advanced control methods using machine learning and various optimization algorithms have been developed. However, these methods have problems such as requiring long computation times and being unable to perform real-time re-optimization. In particular, when optimizing multiple control targets simultaneously, conventional classical computers sometimes become unable to search for solutions within a realistic timeframe due to combinatorial explosion.
[0007] Ising model optimization techniques have the potential to solve combinatorial optimization problems quickly, and their application to the above-mentioned problems is expected. Ising model optimization includes quantum annealing utilizing quantum effects, classical simulated annealing, and CMOS circuitry. Digital circuit type annealing machine Various implementation forms exist, such as CMOS annealing machines and coherent Ising machines using optical parametric oscillators. However, when applying the optimal solution obtained through Ising model optimization to a real system, methods for dealing with deviations caused by various external factors have not been sufficiently considered.
[0008] Furthermore, conventional technologies had not adequately established a mechanism for systematically analyzing data accumulated through feedback control, learning the unique characteristics and response patterns of the target system, and improving the accuracy of subsequent optimizations. [Prior art documents] [Patent Documents]
[0009] [Patent Document 1] Japanese Patent Publication No. 2020-166670 [Patent Document 2] Japanese Patent Publication No. 2021-140297 [Patent Document 3] Japanese Patent Publication No. 2022-073589 [Patent Document 4] Japanese Patent Publication No. 2023-048752 [Patent Document 5] International Publication No. 2019 / 163655 [Non-patent literature]
[0010] [Non-Patent Document 1] MW Johnson et al., "Quantum annealing with manufactured spins", Nature, Vol. 473, pp. 194-198, 2011 [Non-Patent Document 2] Fujitsu Limited, "Digital Annealer Technical Document", 2018 [Non-Patent Document 3] T. Inagaki et al., "A coherent Ising machine for 20000-node optimization problems", Science, Vol. 354, Issue 6312, pp. 603-606, 2016 [Non-Patent Document 4] M. Yamaoka et al., "A 20k-Spin Ising Chip to Solve Combinatorial Optimization Problems With CMOS Annealing", IEEE Journal of Solid-State Circuits, Vol. 51, No. 1, pp. 303-309, 2016 [Overview of the Initiative] [Problems that the invention aims to solve]
[0011] This invention has been made in view of the problems of the prior art described above, and aims to solve the following problems.
[0012] Firstly, when controlling the target system based on the optimal solution calculated by Ising model optimization, if a deviation occurs from the target state due to factors such as fluctuations in the external environment, unpredictable disturbances, or aging of the target system, a mechanism for rapid and appropriate re-optimization is to be established.
[0013] Secondly, we will introduce a mechanism to dynamically perform deviation analysis when a deviation from the current state is detected, and to adjust the optimization parameters based on that analysis.
[0014] Thirdly, systematically accumulate the history of feedback control and build a knowledge base to improve the accuracy of subsequent optimizations.
[0015] Fourthly, we will introduce a mechanism that automatically extracts characteristics specific to the target system from accumulated feedback data and learns and optimizes the initial weighting parameters for subsequent optimization.
[0016] The technologies disclosed in the above-mentioned Patent Documents 1 to 5 and Non-Patent Documents 1 to 4 relate to the basic principles and various implementation forms of Ising model optimization, but do not disclose integration with feedback control, dynamic weighting adjustment based on deviation detection, or learning of target system characteristics by machine learning. [MEANS FOR SOLVING THE PROBLEMS]
[0017] To solve the above problems, the present invention adopts the following configurations.
[0018] A. Feedback-type optimization function
[0019] (1) State monitoring unit: Collect data indicating the state of the target system via a group of sensors installed in the target system, aggregate the data in real time, and continuously monitor the deviation from the target state.
[0020] (2) Deviation determination unit: After control execution based on the optimal solution, at the point when a certain period of time has elapsed, determine whether the deviation between the current state and the target state exceeds a threshold value.
[0021] (3) Weight adjustment unit: When a deviation is detected, dynamically adjust the weighting parameter in the objective function of the optimization problem based on the deviation analysis of the actual value and the target value of the deviation and the deviation direction. Specifically, a weight is preferentially assigned to a state element with a large deviation, and the optimization problem is configured to preferentially improve that state element.
[0022] (4) Ising model optimization unit: Using the adjusted weighting parameter, search for a new optimal solution in a short time through Ising model optimization calculation. With the high-speed search capability of Ising model optimization, this re-optimization process is efficiently executed until the deviation from the target state is eliminated. The Ising model optimization unit in the present invention is implemented by any of a quantum annealing machine, a simulated annealing execution environment, Digital circuit type annealing machine , a CMOS annealing machine, a coherent Ising machine, or an Ising machine having functions equivalent to these.
[0023] (5) Database recording unit: In each optimization cycle, record the following information in a history database: state data, parameter data used in the relevant optimization cycle, weighting values used for optimization, target state values, solutions used for optimization, and control results.
[0024] (6) Control execution unit: Based on the new optimal solution, it executes control of the target system and repeats the processes (1) to (5) above until the deviation from the target state is eliminated.
[0025] B. Target characteristic learning function using machine learning
[0026] (7) Target Characteristics Learning Unit: This unit analyzes the database records accumulated by the above-mentioned feedback-type optimization function using a machine learning model to extract characteristics specific to the target system. Specifically, it extracts the following information: Extraction of response patterns, learning under what conditions and to what extent deviations occur. Evaluation of weighting effects, analyzing how effective the parameters were. Construction of a target system model, identifying the response characteristics, baseline, and architecture specific to the target system. Analysis of the impact of external environmental changes on the target system, as part of external factor impact analysis.
[0027] (8) Weighted parameter optimization unit: Based on the target system characteristics extracted by the machine learning model, the initial weighted parameters used in subsequent optimization cycles are optimized. This improves the convergence speed of the feedback control, enabling optimization that approaches the target state more quickly.
[0028] (9) Prediction model update unit: The state prediction model is periodically updated using the learning results to improve the optimization accuracy. [Effects of the Invention]
[0029] The present invention provides the following effects.
[0030] (1) Rapid Adaptation: Leveraging the Ising model optimization's ability to search for optimal solutions in seconds, deviations can be detected and the optimal solution can be found in an extremely short time. This enables rapid response to sudden changes in the external environment and unpredictable disturbances.
[0031] (2) Continuous improvement of control accuracy: By dynamically adjusting the weightings that are constantly being improved, it is possible to continuously search for an optimization that prioritizes improving state elements with large deviations. This is expected to lead to continuous improvement of control accuracy without manual intervention.
[0032] (3) Automatic learning of target characteristics: Machine learning models can automatically extract the characteristics, response patterns, and effective control strategies specific to the target system from the feedback data. This eliminates the need for individual tuning for each target system, significantly reducing implementation and operating costs.
[0033] (4) Continuous improvement of prediction accuracy: As the learned target system model is reflected in subsequent optimization, the accuracy of the initial weighting improves and the convergence speed of the feedback control increases. The longer the system is operated, the more a positive feedback loop is formed that enables highly accurate solution search and convergence.
[0034] (5) Improved Robustness: By combining continuous feedback and learning of target characteristics through machine learning, the system's robustness is significantly improved because it can address external factors that cannot be captured by predictive models and data based on characteristics specific to the target system.
[0035] (6) Achieving objectives and optimizing costs simultaneously: By searching for the solution that minimizes costs at each point in time while reliably achieving the target state, it becomes possible to achieve both objectives and optimize operating costs.
[0036] (7) Flexibility of implementation: The Ising model optimization unit can be a quantum annealing machine, simulated annealing, Digital circuit type annealing machine Because it can be configured in any implementation form, such as a CMOS annealing machine or a coherent Ising machine, it allows for flexible system construction according to available hardware resources and cost requirements. [Brief explanation of the drawing]
[0037] [Figure 1]This is an overall diagram of a feedback-type Ising model optimization system according to an embodiment of the present invention. [Figure 2] This is a flowchart of a feedback-type optimization process. [Figure 3] This is a flowchart of the process of learning target characteristics using machine learning. [Figure 4] This is a conceptual diagram of the weighting adjustment algorithm. [Figure 5] This is a schematic diagram of the database record structure. [Figure 6] This is a diagram illustrating its application to an energy management system. [Figure 7] This is a diagram illustrating the application to a manufacturing process optimization system. [Figure 8] This is a diagram illustrating the application to a logistics optimization system. [Figure 9] This is a diagram illustrating the application to a traffic control system. [Figure 10] This is a conceptual diagram showing various implementation configurations of Ising machines. [Modes for carrying out the invention]
[0038] Embodiments of the present invention will be described in detail below.
[0039] 1. System Configuration
[0040] The system of this embodiment consists of the following main components. The sensor group consists of various sensors that measure the state of the target system. The edge computing layer includes a state monitoring unit, a deviation determination unit, and a control execution unit. The Ising model optimization calculation unit consists of a quantum annealing machine (manufactured by D-Wave Systems, etc.), a simulated annealing execution environment, Digital circuit type annealing machine (manufactured by Fujitsu Limited, etc.) This includes a CMOS annealing machine (manufactured by Hitachi, Ltd., etc.), a coherent Ising machine (manufactured by Nippon Telegraph and Telephone Corporation, etc.), or a cloud service for these. In addition, it includes a machine learning model engine, a history database, a control interface, and an integrated management server.
[0041] 2. Implementation form of the Ising model optimization unit
[0042] The Ising model optimization unit in the present invention can be configured in any of the following implementation forms.
[0043] (a) Quantum annealing machine: This machine uses quantum effects (quantum tunneling effect) to search for the ground state of the Ising model. A representative example is the quantum annealing machine manufactured by D-Wave Systems. It is implemented using superconducting qubits and operates in cryogenic environments.
[0044] (b) Simulated annealing: This method simulates thermal fluctuations on a classical computer to search for the optimal solution in the Ising model. It can run on a general-purpose CPU or GPU and does not require dedicated hardware.
[0045] (c) Digital circuit-based annealing machines: These are Ising machines based on digital circuits, and one example is the one developed by Fujitsu Limited. It uses CMOS technology to rapidly search for solutions to the Ising model. It operates at room temperature and has lower deployment and operating costs compared to quantum annealing machines.
[0046] (d) CMOS Annealing Machine: This is a CMOS circuit-based Ising machine developed by Hitachi, Ltd. and others. It uses semiconductor technology to search for solutions to the Ising model.
[0047] (e) Coherent Ising machine: This is an Ising machine based on an optical parametric oscillator, developed by Nippon Telegraph and Telephone Corporation and others. It represents the Ising model using the phase state of optical pulses and searches for a solution through optical interaction.
[0048] The present invention is applicable to any of the above implementation forms, and the specific implementation form of the Ising model optimization unit can be selected according to available hardware resources, required computing speed, cost requirements, etc.
[0049] 3. Feedback-type optimization processing flow
[0050] The feedback-type optimization process is performed in the following steps.
[0051] In step B1, the optimization cycle is initiated.
[0052] In step B2, the initial optimal solution is calculated using the Ising model optimization based on the current state data and the target state.
[0053] Based on the optimal solution calculated in step B3, control of the target system is executed.
[0054] In step B4, wait for the evaluation of the results to elapse.
[0055] In step B5, the current state is measured using a group of sensors.
[0056] In step B6, it is determined whether the deviation between the current state and the target state is within a threshold.
[0057] If the deviation in step B7 is within the threshold, proceed to step B14; otherwise, proceed to step B8.
[0058] In step B8, the optimization results, state data, weighting values, and target state at the time of deviation are recorded in the database.
[0059] In step B9, the weighting of the optimization problem is dynamically adjusted based on the divergence analysis and the direction of the divergence.
[0060] In step B10, a new optimal solution is searched using the adjusted weighting.
[0061] In step B11, the parameters used for re-optimization, the target values, and the new optimal solution are recorded in the database.
[0062] In step B12, check whether the divergence is converging.
[0063] If convergence is achieved in step B13, proceed to step B15; otherwise, return to step B3.
[0064] In step B14, the target state is achieved, and the cycle completes successfully.
[0065] If the discrepancy does not converge in step B15, an alert will be sent to the administrator.
[0066] 4. Machine Learning Process Flow for Learning Target Characteristics
[0067] The process of learning target characteristics using machine learning is performed in the following steps.
[0068] In Step B16, the learning cycle is initiated with a regular learning schedule.
[0069] In step B17, data for the learning period is extracted from the historical database.
[0070] In Step B18, we will analyze under what conditions and to what extent the deviation occurs.
[0071] In step B19, evaluate how effective each weighting parameter was.
[0072] In step B20, identify the system-specific response characteristics, baseline, and architecture.
[0073] In step B21, we analyze the impact of external environmental changes on the target system.
[0074] In step B22, the initial weight parameters are optimized based on the learning results.
[0075] In step B23, the state prediction model is periodically updated using the learning results to improve optimization accuracy. [Explanation of symbols]
[0076] 100 Feedback-type Ising model optimization system 110 Sensor Group 111 Temperature sensor 112 Humidity Sensor 113 Power Sensor 114 Position Sensor 120 Edge Computing Layers 121 Status Monitoring Unit 122 Discrepancy Determination Unit 123 Weight adjustment section 124 Control Execution Unit 130 Ising Model Optimization Computation Unit 131 Quantum Annealing 132 Simulated Annealing 133 Digital circuit type annealing machine 134 CMOS annealing 135 Coherent Ising Machine 140 Machine Learning Engines 141 Target Characteristics Learning Unit 142 Weighting parameter optimization unit 143 Predictive Model Update Section 150 History Database 160 control interfaces 170 Integrated Management Server 171 Management screen 180 Target Systems
Claims
1. A state monitoring unit collects sensor data indicating the status of the target system and monitors the deviation from the target state, A weighting adjustment unit dynamically adjusts the weighting parameters of the optimization problem based on the direction and magnitude of the deviation when the deviation exceeds a predetermined threshold. An Ising model optimization unit calculates a new optimal solution by performing an Ising model optimization operation using the adjusted weighting parameters, A control execution unit that executes control of the target system based on the new optimal solution, A database recording unit records state data, weighting parameters, optimal solution, and control results for each optimization cycle in a history database. A feedback-type Ising model optimization system equipped with the following features.
2. The feedback-type Ising model optimization system according to claim 1, wherein the Ising model optimization unit is composed of at least one selected from the group consisting of a quantum annealing machine, a simulated annealing execution environment, a digital circuit type annealing machine, a CMOS annealing machine, and a coherent Ising machine.
3. The feedback-type Ising model optimization system according to claim 1 or 2, further comprising a target characteristic learning unit that analyzes data accumulated in the historical database using a machine learning model and extracts response characteristics, response patterns, and effective weighting parameters specific to the target system.
4. The feedback-type Ising model optimization system according to claim 3, further comprising a weighting parameter optimization unit that optimizes the initial weighting parameters in subsequent optimization cycles based on the target system characteristics extracted by the target characteristic learning unit.
5. The feedback-type Ising model optimization system according to claim 4, wherein the target characteristic learning unit analyzes the degree of influence that changes in the external environment have on the target system, and the weighting parameter optimization unit adjusts the weighting parameters taking that degree of influence into consideration.
6. The Ising model optimization unit performs the process from deviation detection to re-optimization completion within a few seconds, as described in claim 1 or 2, in the feedback-type Ising model optimization system.
7. The feedback-type Ising model optimization system according to claim 1 or 2, wherein the target system is a building energy management system, and the optimization target includes at least one of air conditioning, lighting, and power consumption.
8. The feedback-type Ising model optimization system according to claim 1 or 2, wherein the target system is a manufacturing process, and the optimization target includes at least one of production scheduling, equipment utilization rate, and energy consumption.
9. The feedback-type Ising model optimization system according to claim 1 or 2, wherein the target system is a logistics system, and the optimization target includes at least one of delivery routes, vehicle dispatch, and inventory placement.
10. The feedback-type Ising model optimization system according to claim 1 or 2, wherein the target system is a traffic control system, and the optimization target includes at least one of signal control timing and path guidance.
11. The steps include: collecting sensor data indicating the state of the target system and monitoring the deviation from the target state; If the aforementioned deviation exceeds a predetermined threshold, the weighting parameters of the optimization problem are dynamically adjusted based on the direction and magnitude of the deviation. The steps include: calculating a new optimal solution by performing an Ising model optimization operation using the adjusted weighting parameters; The steps include: executing control of the target system based on the new optimal solution; The steps include recording state data, weighting parameters, optimal solution, and control results for each optimization cycle in a historical database, A feedback-type Ising model optimization method that includes this.
12. The feedback-type Ising model optimization method according to claim 11, wherein the Ising model optimization operation is performed by at least one selected from the group consisting of a quantum annealing machine, a simulated annealing execution environment, a digital circuit type annealing machine, a CMOS annealing machine, and a coherent Ising machine.
13. The feedback-type Ising model optimization method according to claim 11 or 12, further comprising the step of analyzing the data accumulated in the historical database using a machine learning model to extract response characteristics, response patterns, and effective weighting parameters specific to the target system.
14. The feedback-type Ising model optimization method according to claim 13, further comprising the step of optimizing the initial weighting parameters in subsequent optimization cycles based on the extracted target system characteristics.
15. A program for causing a computer to execute the feedback-type Ising model optimization method according to claim 11 or 12.
16. A program for causing a computer to execute the feedback-type Ising model optimization method described in Claim 13.
17. A program for causing a computer to execute the feedback-type Ising model optimization method described in Claim 14.
Citation Information
Patent Citations
Intelligent fault diagnosis and active control system for ship multi-source heterogeneous data fusion
CN120779923A
Property-specific balance management device, property-specific balance management method, and property-specific balance management program
JP2020166670A
Abnormality diagnosis device
JP2021140297A
Mechanical seal device
JP2022073589A
Toilet device
JP2023048752A
