Tokyo Institute of Science Builds Quantum Trainable Inequality Constraint Solver

Researchers from the Institute of Science Tokyo and RIKEN Center for Advanced Intelligence Project have developed a new approach to solving complex combinatorial optimization problems without increasing the demands on quantum computing hardware. Their work centers on the deep-unfolded unbalanced penalization Ohzeki method, or DU-UPOM, which learns the most effective way to find solutions from example problems, a departure from earlier methods requiring manual adjustments. Numerical experiments using random knapsack problems show that DU-UPOM improves over original UP and reaches optimal solutions in fewer iterations than fixed-step UPOM and other baseline methods. This framework, the authors report, reduces the tuning and embedding burdens while making the Ohzeki method trainable for problems with inequality constraints, potentially expanding the scope of solvable problems for quantum annealers.

Combinatorial Optimization & Quantum Annealing Introduction

Achieving optimal solutions to complex problems is now faster with a refined approach to quantum annealing. Researchers from the Institute of Science Tokyo and RIKEN Center for Advanced Intelligence Project are focused on tackling combinatorial optimization problems (COPs), challenges where the goal is to find the best solution from a vast number of possibilities, such as logistics, financial modeling, or machine learning. Traditional methods often struggle with these problems, prompting exploration of quantum-based solutions like quantum annealing. However, effectively translating these problems into a format suitable for quantum annealers presents significant hurdles, particularly when dealing with inequality constraints. A common technique for handling these constraints involves adding extra parameters to the equation. While effective, this approach increases the complexity of the problem, demanding more qubits, the quantum equivalent of bits, and escalating the process of mapping the problem onto the quantum hardware.

To circumvent this, researchers have developed the unbalanced penalization (UP) method, which avoids slack variables. However, original UP requires careful manual adjustment of two penalty coefficients, a process that can be both time-consuming and problem-dependent. The study notes that “the two hyperparameters must be tuned jointly, which is difficult because suitable values depend on the characteristics of the target problem and the sampler.” Building on this foundation, the authors proposed the unbalanced penalization Ohzeki method (UPOM), a refinement that replaces these static penalty coefficients with a dynamic auxiliary-variable update. This update is governed by an and the latest iteration, and the “deep-unfolded unbalanced penalization Ohzeki method” (DU-UPOM) takes this a step further. DU-UPOM learns this optimal schedule directly from training instances, eliminating the need for manual tuning. This learning capability, achieved through a technique called deep unfolding, allows the system to adapt and improve its performance over time.

Unbalanced Penalization and the Ohzeki Method

Following advances in quantum annealing as a method for solving complex combinatorial optimization problems, researchers from the Institute of Science Tokyo and RIKEN Center for Advanced Intelligence Project are increasingly focused on minimizing the resources required to implement these solutions on current hardware. A key challenge lies in representing constraints within these problems; traditional approaches often rely on introducing slack variables, which increase the number of qubits needed and complicate the embedding process on quantum annealers. This qubit overhead can severely limit the scale of problems that can be tackled effectively. To address this, a technique known as unbalanced penalization (UP) emerged, offering a way to handle inequality constraints without adding these extra variables. The optimal values for these coefficients are highly problem-dependent, creating a significant tuning burden for users.

The inclusion of a squared residual term within the Hamiltonian, the core component defining the energy landscape explored by the quantum annealer, introduced additional quadratic interactions, potentially exacerbating the embedding challenges. This new approach integrates UP with the Ohzeki method, a technique designed to streamline constraint handling. UPOM replaces the static penalty coefficients of the original UP with a dynamic auxiliary-variable update, governed by a schedule that can be adjusted during the optimization process. Critically, the UPOM Hamiltonian eliminates the squared residual term, reducing the number of quadratic couplings and simplifying the embedding process. This work leverages deep unfolding, a machine learning framework, to learn the optimal update schedule for solving these problems directly from training instances. Unlike earlier methods requiring manual parameter tuning, DU-UPOM automates this process, adapting to the specific characteristics of each problem.

The Institute of Science Tokyo proposes a solution to a critical bottleneck in quantum optimization: the efficient encoding of complex problems for quantum annealers. Researchers from the Institute of Science Tokyo, including Ryo Hagiwara, are focused on minimizing the number of physical qubits required to represent a logical problem, particularly when dealing with inequality constraints. The authors propose a shift away from slack variables with the “unbalanced penalization Ohzeki method (UPOM).” This technique integrates unbalanced penalization with the Ohzeki method, a strategy designed to reduce the number of interactions needing embedding on quantum annealer hardware. Unlike earlier methods, UPOM replaces static penalty coefficients with an auxiliary-variable update, dynamically adjusting to the problem’s characteristics. Building on this work, the authors developed “DU-UPOM”, the “deep-unfolded unbalanced penalization Ohzeki method”, which introduces a machine learning component.

Deep Unfolding for Trainable Iterative Algorithms

The demand for efficient solutions to complex optimization challenges is driving innovation in quantum and classical computing alike, with recent work from researchers at the Institute of Science Tokyo offering a novel approach to tackling combinatorial optimization problems. Researchers Ryo Hagiwara, Shunta Arai, and Satoshi Takabe are moving beyond manually tuned algorithms by integrating machine learning directly into the iterative solution process. A key challenge in implementing combinatorial optimization problems on quantum annealers lies in efficiently encoding constraints. Traditional methods often introduce slack variables to handle inequalities, but these significantly increase the complexity of the problem and the number of qubits required. This work builds on the method, circumventing this need for slack variables, offering a potential reduction in qubit overhead. Deep unfolding is a technique that transforms iterative algorithms into trainable neural network-like models.

Each iteration of the algorithm becomes a layer, and the parameters governing the process, in this case, the step sizes for updating auxiliary variables, are learned from training data.

DU-UPOM: Deep-Unfolded Ohzeki Method for Inequalities

While quantum annealing promises solutions to complex optimization challenges, translating real-world problems into a format suitable for these devices often introduces significant hurdles. A common approach, utilizing slack variables to handle inequality constraints, ironically increases the computational burden by demanding more qubits. Researchers from the Institute of Science Tokyo and RIKEN Center for Advanced Intelligence Project built upon the unbalanced penalization (UP) technique, which avoids slack variables altogether. Original UP, however, presented its own challenges, requiring careful tuning of two penalty coefficients. The approach replaces these static coefficients with an auxiliary-variable update schedule, learned automatically from training data. Each iteration becomes a layer, with the step sizes optimized based on a dataset of problem instances. This isn’t simply automating a process; it’s fundamentally changing how the optimization unfolds. The researchers demonstrated the framework’s effectiveness through numerical experiments on random knapsack problems, a standard benchmark for optimization algorithms.

By removing the need for manual coefficient tuning and eliminating a squared residual term that contributes to quadratic interactions, the work potentially reduces the number of physical qubits required on a quantum annealer, bringing practical quantum solutions closer to reality.

Researchers from the Institute of Science Tokyo and RIKEN Center for Advanced Intelligence Project detailed their findings in recent work focused on improving combinatorial optimization problem (COP) solvers for quantum annealers. The core challenge addressed is minimizing qubit usage, particularly when dealing with inequality constraints, a common feature in problems like the knapsack problem. This update is then further refined through a deep unfolding (DU) technique, resulting in DU-UPOM. Numerical experiments using random knapsack problems show that DU-UPOM improves over original UP and reaches optimal solutions in fewer iterations than fixed-step UPOM and other baseline approaches.