Toshiba’s Quantum-Inspired Optimization Framework Enables Constant Optimization in Dynamically Changing Environments

Kawasaki, Japan – Toshiba Corporation has developed a quantum-inspired optimization framework that delivers quick, stable solutions to combinatorial optimization problems in the dynamically changing conditions of real-world environments.
Systems deployed in real-world conditions must contend with dynamically changing environments. Network conditions affect wireless communications; in-vehicle systems must deal with ever-moving objects around them; and sensor information available to robots changes over time. These environmental changes also cause the scale and characteristics of optimization problems to change over time. Real-world systems must constantly respond and adapt to their environment by quickly deriving appropriate solutions and constantly optimizing operation.
Toshiba’s new framework combines its proprietary quantum-inspired optimization computer, the Simulated Bifurcation Machine (SBM), with an optimization control AI that assesses problem conditions, and compression technology for Ising models^*1. The AI automatically selects appropriate SBM parameters and execution platform (machine) based on the characteristics of an input problem. In addition, Ising model compression reduces both data transfer time and computation time, achieving low latency in the overall system. Together, these technologies realize high speed and stable operation in diverse real-world conditions and deliver constant optimization. In an evaluation assuming communication timing scheduling for wireless terminals, results confirmed speed advantages over conventional methods.
This result was published online in Nature Communications on June 16, 2026^*2.

Logistics, communications, mobility, and robotics are among the many fields that must deal with complex combinatorial optimization problems by allocating limited resources or selecting optimized combinations from multiple candidates. For instance, wireless communications require efficient allocation of time slots and frequencies, while in-vehicle systems and robotics both require rapid object recognition and tracking, route selection, and task allocation based on analysis of surrounding moving objects and obstacles. All of these systems must contend with environments that can change dynamically but nonetheless deliver stable performance.
Toshiba has long pioneered research in SBM. Results to date include technologies for high-speed solutions to large-scale combinatorial optimization problems^*3*4*5 and for real-time optimization on embedded devices^*6. Expanding application into real-world environments has required reducing the workload of tuning parameters to reflect a problem’s scale and constraints, and realizing a mechanism to automatically select appropriate execution conditions that reflect the changing environment.

Toshiba has developed a quantum-inspired optimization framework that combines optimization control AI for condition assessment and Ising model compression with SBM. The combination enables constant optimization in a wide range of changing conditions. The main features are described below.

Maximizing the performance of optimization computing technologies like SBM requires adjustment of individual parameters to reflect problem scale, constraints, required solution quality, and allowable computation time. However, systems such as wireless communications, in-vehicle systems, and robotics are all susceptible to environmental changes that make it difficult to maintain stable performance with fixed parameter settings.
In Toshiba’s new framework, the optimization control AI analyzes the characteristics of the target problem and automatically estimates SBM parameters. Based on the scale of the problem and estimation results, it then assigns the machine: either an SBM implemented on an FPGA or in a CPU. The FPGA-based SBM is suitable for larger problems while the CPU-based SBM operates quickly on smaller problems. Selecting the appropriate machine for the scale of the problem reduces the workload of parameter tuning, and delivers fast, stable optimization under diverse conditions. It also expands the range of systems where SBM can be applied.

When applying SBM in deployed systems, the target problem is represented as an Ising model that is transferred to the SBM for solution. If the Ising model data is large, the time required for data transfer and computation can become factors that increase overall system latency. Reducing latency is especially important in applications that require frequent updates.
This framework looks for repeated identical values in the Ising model, groups them, and represents them with assigned numbers (indices). This facilitates model compression, reducing the data transferred to the SBM and the time the transfer takes. Computation methods are also optimized, allowing computation to be performed directly on the compressed representation without fully expanding it back to its original form.　　　　Processing identical values collectively improves computational efficiency and cuts computation time on the SBM. In combination, these factors secure low-latency optimization, even in systems with dynamically changing conditions.

With the framework, Toshiba has constructed a solver for maximum independent set (MIS) problems^*7. Toshiba evaluated the solver on benchmark problems^*8 and on time division multiple access (TDMA) scheduling^*9 in wireless multi-hop networks^*10.
When aggregating data from multiple sensors to a base station, TDMA scheduling determines combinations of nodes that can transmit data simultaneously within a timeslot without wireless interference. This process can be treated as a problem of repeatedly finding an MIS consisting of a group of transmitting node candidates. In the scheduling process, the characteristics of generated MIS problems, such as their scale and structure (e.g., number of nodes and edge density), vary significantly each time a computation is performed. Evaluation results showed that, when solving problems with such varying characteristics, processing was faster than with conventional software-based MIS solvers, because the framework can continuously solve the problems while the optimization control AI selects the SBM parameters and machine. This demonstrates the applicability to systems that must constantly solve optimization problems with varying characteristics.

Video: Demonstration of TDMA scheduling

Toshiba will use the framework to extend the range of systems where SBM can be applied. The company also plans to provide application examples in on-premises development environment with FPGA-based SBM. Through these efforts, Toshiba will promote the implementation of quantum-inspired optimization technologies in society.

• A mathematical model that originates from physics and represents combinatorial optimization problems as energy minimization problems.
• Y.Hamakawa, T.Kashimata, M.Yamasaki, K.Tatsumura, Machine Learning-assisted High-speed Combinatorial Optimization with Ising Machines for Dynamically Changing Problems, Nature Communications 17, 4877 (2026). http://doi.org/10.1038/s41467-026-73725-6
• https://www.global.toshiba/ww/technology/corporate/rdc/rd/topics/19/1904-01.html
• https://www.global.toshiba/ww/technology/corporate/rdc/rd/topics/21/2102-01.html
• https://www.global.toshiba/ww/technology/corporate/rdc/rd/topics/26/2604-01.html
• https://www.global.toshiba/ww/technology/corporate/rdc/rd/topics/26/2602-02.html
• A combinatorial optimization problem that selects as many mutually non-conflicting elements as possible. The larger the scale, the more complex it becomes to reach the most appropriate solution.
• Test problems created to evaluate algorithm and system performance such as speed and solution quality. To evaluate this framework, problems were randomly generated using different node quantities and edge density.
• A process to allocate a timeslot to each device based on time division multiple access (TDMA), a method that allows multiple devices to communicate sequentially in different timing to avoid collisions.
• A wireless network in which multiple data communication devices (nodes) relay data, often used for distant or wide-area connections.