Pith. sign in

REVIEW 3 major objections 5 minor 63 references

Real-Time Black-Box Optimization for Dynamic Discrete Environments Using Embedded Ising Machines

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that a sliding-window, incentive-driven Ising-machine loop can solve dynamic discrete bandit problems with exponentially many actions, holding roughly 80% of optimal reward during environmental drift.

desk verdict A solid engineering extension of FMQA to dynamic MAB with a real embedded Ising machine; the evaluation needs surrogate diagnostics and stronger baselines, but the core claim holds as a heuristic proof-of-concept. read the letter →

arxiv 2506.16924 v1 pith:ROJSTQ34 submitted 2025-06-20 cs.AI cs.ET

Tomoya Kashimata , Yohei Hamakawa , Masaya Yamasaki , Kosuke Tatsumura This is my paper
classification cs.AIcs.ET
keywords real-timeblack-boxoptimizationmulti-armedbanditdiscretecombinatorialIsingmachinefactorizationsimulatedbifurcationdynamicenvironmentswirelessbeamformingcoordination
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that a multi-armed bandit problem over an enormous combinatorial action space, one where no conventional bandit algorithm can enumerate or linearly model the arms, can be solved in real time when an embedded Ising machine is in the loop. The proposed algorithm, RT-BBO, keeps the factorization-machine surrogate used by the static FMQA method but retrains it on a sliding window of recent observations, applies pre-training weight decay, and adds an exploration incentive that rewards spins which have not changed for many cycles. In a synthetic changing Ising environment, the multi-reward version holds about 80% of the white-box-optimal reward during environmental drift and about 90% in static periods. In a wireless beamforming simulator with 19 base stations and moving users, it beats a greedy method that is allowed to look inside the black box, even though RT-BBO sees only throughput values. The reason to care is that many real-time systems have exactly this structure: discrete joint actions, changing conditions, and rewards that depend on variable interactions.

What carries the argument

The load-bearing mechanism is the acquisition function $\hat{r}(s) + H_{\mathrm{exploration}}(s) + H_{\mathrm{encoding}}(s)$, maximized by a simulated-bifurcation Ising machine on an FPGA. Here $\hat{r}(s)$ is a second-order factorization machine, a low-rank quadratic surrogate $\sum_{i<j}\langle v_i,v_j\rangle s_i s_j + \sum_i w_i s_i + w_0$, whose $K$-dimensional factor vectors (here $K=6$) let a small sliding window of 50 samples train a model over 171 spins. The exploration term $H_{\mathrm{exploration}} = -c_{\mathrm{exploration}}\sum_i I_i s_i$ uses elapsed-time counters, squared and sign-flipped by the last spin value, to push rarely changed variables to flip, and $H_{\mathrm{encoding}}$ enforces one-hot constraints for the nine beam patterns per base station. Pre-training weight decay, scaling all parameters by $c_{\mathrm{decay}}=0.999$ before each training pass, is the device that lets the next sample's gradient overwrite the previous environment. The machinery converts the full action space of $9^{19} \approx 1.35\times 10^{18}$ beamforming patterns into a 171-spin Ising problem that the embedded machine can approximately solve in about 17 ms per sampling cycle.

What would settle it

Run the beamforming demonstration while logging, each sampling cycle, the log-cosh error of the surrogate on the newest held-out samples and the gap between RT-BBO's achieved throughput and the white-box optimum of the simulator's SINR model; if surrogate error rises sharply as the users converge at $t=2000$ while achieved throughput still stays high, or if throughput matches the greedy method when the surrogate is trained on time-shuffled rewards, the dynamic-adaptation explanation would be contradicted.

Watch

Extended reading notes

Core claim

The central discovery is that the bottleneck for dynamic discrete black-box optimization is not the Ising machine but the adaptability of the surrogate model, and that a small set of heuristics is enough to make an Ising-machine bandit track a drifting environment. RT-BBO extends FMQA by discarding old data through a sliding window, shrinking model parameters once per sampling cycle before training so recent data dominate, and adding an exploration incentive into the acquisition function, with counters for how long each spin has held its value and an adjustable coefficient that keeps the average dwell time in a target range. When the total reward is a sum of observable sub-rewards, RT-BBO for MR trains one low-rank quadratic sub-surrogate per sub-reward and integrates them; this decomposition improves approximation enough to reach roughly 90% of the white-box Ising optimum in static conditions and about 80% during environmental change, and in the wireless demonstration it keeps network throughput above the greedy beamforming baseline despite having less information. The paper's claim is that these RT-BBO variants are a practical heuristic solution to the dynamic combinatorial multi-armed bandit problem for embedded, low-latency Ising machines.

Load-bearing premise

The load-bearing premise is that a low-rank quadratic factorization machine with $K=6$, retrained on only 50 recent samples, can approximate a highly non-quadratic, time-varying reward function such as SINR-based wireless throughput well enough for the Ising machine to keep selecting near-optimal actions; the paper does not report a direct check of surrogate accuracy.

Editorial extensions

If this is right

  • Dynamic combinatorial bandit problems with exponentially many actions become tractable in closed-loop settings, provided the environment changes slowly relative to the sampling cycle and the surrogate can track it.
  • Because only instant rewards are needed, not internal state observations, RT-BBO applies to black-box systems where the controller has no access to the environment's model; the beamforming demonstration outperforms a greedy method that does have that access.
  • Multi-reward decomposition is a free accuracy boost when the total reward is additive or a linear combination of observable sub-rewards, so system designers should expose sub-rewards wherever possible.
  • The low latency of the simulated-bifurcation Ising machine, roughly 58 sampling cycles per second, is what makes the method real-time; on slower solvers the same heuristic loop would lose its tracking advantage.
  • The same acquisition-function architecture should transfer to other dynamic discrete problems such as scheduling, pricing, and recommendation, since those problems share the combinatorial-action, reward-only-observation structure.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A diagnostic the paper does not report, surrogate prediction error on held-out recent samples, would separate two explanations of the method's success: genuine tracking by the quadratic model versus exploration noise keeping the agent near good regions.
  • The window size, rank $K$, and decay coefficient $c_{\mathrm{decay}}$ are coupled, so a natural extension is to make the window or rank adaptive to a running estimate of environmental non-stationarity rather than fixing them.
  • If the incentive's dwell-time range is the real source of exploration, then a simpler agent that randomly flips a small fraction of spins after long dwell times may capture part of the gain at lower cost; testing that would isolate the value of the Ising machine itself.
  • The performance claim depends on gradual drift, so applying RT-BBO to an abruptly switching or adversarial environment should degrade sharply, which would delimit the class of usable problems.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. This manuscript proposes RT-BBO, a heuristic multi-armed bandit method for dynamic discrete environments. It extends the FMQA black-box optimization framework by (i) training the quadratic factorization-machine surrogate on a sliding window of recent samples, (ii) applying a pre-training weight decay before each round of training, (iii) adding an exploration incentive to the acquisition function maximized by an FPGA-based simulated-bifurcation Ising machine, and (iv) optionally decomposing the total reward into separately trained sub-surrogate models. The authors evaluate RT-BBO on a synthetic environment whose black-box functions are sums of random Ising models and on a 19-cell wireless beamforming simulator with moving users, reporting roughly 80–90% of the white-box Ising energy in the synthetic setting and higher throughput than greedy and random baselines in the wireless setting, with per-cycle latencies around 17 ms.

Significance. If the central claim holds, the paper is a valuable demonstration that an embedded Ising machine can serve as the action-selection engine inside an online combinatorial bandit loop, and the beamforming experiment is a concrete realistic use case in which the method outperforms a greedy policy that uses privileged environmental information. The algorithmic design is clearly decomposed into named components, the training loop is given in pseudocode, and the FPGA/GPU implementation details and measured latencies are reported, which supports reproducibility. The main weakness is validation: the surrogate model that carries the mechanism is never checked for predictive accuracy, the synthetic benchmark is confined to quadratic Ising-model rewards, and the comparison set lacks standard bandit baselines. These gaps are addressable with additional experiments, so the contribution is promising but not yet fully established.

major comments (3)
  1. [§2.2, Fig. 4, and Methods ('Model training unit', 'Implementation')] No diagnostic of the surrogate model's predictive accuracy is reported anywhere, despite the fact that the acquisition function in Eq. (3) is only as good as the surrogate. In the wireless experiment N=171, K=6, and the sliding-window size is 50, giving 1198 free parameters trained on 50 recent reward samples, while the target function in Eqs. (5)–(6) is non-quadratic and time-varying. Please report held-out prediction error or recent-data R², and include ablations over K, window size, c_decay, and the incentive strength. Without this, the paper's central mechanism—that the Ising machine exploits interactions between variables through the surrogate—is not established; the observed throughput advantage could in principle be caused by the exploration incentive or by the one-hot penalty alone.
  2. [§2.1 and Fig. 3] The synthetic black-box functions are constructed by summing ten randomly generated Ising models, i.e., they are quadratic in the spin variables, the same functional family as the factorization-machine surrogate (up to the rank-K truncation). The 80–90% relative performance therefore demonstrates the method on a favorably matched problem class, but it does not test whether the surrogate can track a genuinely non-quadratic dynamic reward. Please add synthetic environments with higher-order interactions or other non-quadratic structure, and ideally a problem where the reward depends on discrete-value combinations in a non-decomposable way.
  3. [§2.2.2 and Figs. 4c–4d] The evaluation would be more persuasive with a standard combinatorial-bandit baseline and with uncertainty estimates. The paper motivates RT-BBO by saying conventional MAB algorithms cannot effectively optimize dynamic discrete environments, but no UCB, Thompson-sampling, or combinatorial-bandit algorithm is included in the comparison. In addition, the reported curves are averages over 50 trials without error bars or significance tests, so the margins between RT-BBO for MR, RT-BBO for SR, and the greedy method in Fig. 4d cannot be assessed statistically. Please add at least one standard bandit baseline and report confidence intervals or trial-level variability.
minor comments (5)
  1. [Methods, Algorithm 1] The expression for \hat{r}^b uses \sum_i \sum_j, whereas Eq. (1) defines the surrogate as \sum_{j>i}. Please reconcile the notation (e.g., restrict the sum or introduce the factor 2) so that the model matches the gradient formulas in Eqs. (14)–(15).
  2. [Eq. (12)] The index s_{dl·i+l} is not defined; please spell out the mapping from control-input indices to spin indices, since the one-hot penalty is central to the encoding.
  3. [Implementation] The reward scaling factors (1000 for MR, 100000/19 for SR) are stated without justification; a sentence explaining how they were chosen would improve reproducibility.
  4. [Additional information] The paper refers to a supplementary video but notes the preprint version does not include it; please state where the video can be obtained during review.
  5. [§2.1.4] Section 2.1.4 controls c_exploration by maintaining the average counter within a target range, but the initial value of c_exploration and the update step size are not given; please specify them for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation; RT-BBO is an empirical online-learning loop benchmarked against external and internal baselines.

full rationale

The paper's central claim is that RT-BBO, an FMQA-style loop with a sliding window, pre-training weight decay, and an exploration incentive, can track dynamic discrete environments. The acquisition function in Eq. 3 is fitted to observed rewards and then maximized by the embedded Ising machine; the next action's reward is subsequently observed, so the loop is genuinely predictive rather than self-fulfilling. The surrogate model in Eq. 1 is a low-rank factorization machine trained on a sliding-window dataset, and the paper compares against random selection, a greedy method requiring internal state, and a white-box Ising ceiling. The white-box ceiling is computed by the authors' own SB implementation, which is a legitimate benchmark rather than an input to the algorithm: it provides a normalization target but is not used in training or action selection. The multiple-reward decomposition in Eq. 4 decomposes the approximation problem into smaller sub-models, but improvement is empirical and could fail if the sub-surrogates were inaccurate. Self-citations to prior SB and FMQA work are technical and contextual; no load-bearing step reduces to a self-citation chain or to a parameter fitted to the target outcome. The paper does not invoke a uniqueness theorem, and its contribution is a heuristic combination of existing components rather than a renamed known result. The absence of surrogate-accuracy diagnostics is a correctness or validation concern, not a circularity concern.

Assumptions & free parameters 10 free parameters · 6 assumptions · 0 invented entities

The method rests on heuristic design choices rather than a derivation. All algorithm components (window size, decay, incentive, learning rates, penalty strength) are hand-set with little sensitivity analysis, and the modeling assumptions about the reward structure and the Ising machine are taken from prior work or asserted.

free parameters (10)
  • Factorization rank K = 6
    Chosen by authors for all experiments; no sensitivity analysis.
  • Sliding-window size = 50
    Number of recent samples used to train the surrogate in the wireless demo; hand-set.
  • Pre-training weight decay c_decay = 0.999
    Applied once per sampling cycle; no sweep reported.
  • Incentive counter target range = 100-200
    Target range for average consecutive identical spin values; used to adapt c_exploration.
  • Training iterations Ntrain = 200
    Number of gradient updates per sampling cycle; trade-off noted, no sensitivity analysis.
  • Mini-batch size Lbatch = 20
    Selected experimentally; stability trade-off discussed.
  • Reward scaling factors = 1000 (MR), 100000/19 (SR)
    Applied to rewards before training to adjust dynamic range.
  • Encoding penalty coefficient c_encoding = not specified
    Said to need adjustment but no value or procedure given.
  • Learning rate alpha (Adam) = not specified
    Set depending on parameter kind and experiment; exact values omitted.
  • SB hyperparameters a0, c0, eta, Delta t = not specified
    Used in Eqs. (8)-(10) for ballistic SB; no values in the paper.
assumptions (6)
  • domain assumption Reward depends only on the current action, not on previous actions.
    Sec. 1: 'the reward depends on the action taken in the cycle and does not depend on the previous actions.'
  • domain assumption Environment changes are gradual enough for machine learning to adapt.
    Sec. 1: 'changes in the environment are gradual enough that a machine learning method can work effectively.'
  • domain assumption A low-rank quadratic factorization machine can approximate the reward function sufficiently well.
    Eq. (1) and Sec. 2.2; no approximation error analysis is provided.
  • domain assumption The SB-based Ising machine finds near-optimal solutions to the acquisition function quickly.
    The algorithm relies on ballistic SB performance, cited from Refs [20,21] by the same group.
  • ad hoc to paper One-hot encoding penalty with a suitable coefficient enforces valid actions.
    Eq. (12); coefficient c_encoding is not specified and is said to need adjustment.
  • domain assumption The total reward is the sum or linear combination of sub-rewards for the MR extension.
    Eq. (4) integrates sub-surrogate models; holds in the wireless demo by definition of total throughput.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Real-Time Black-Box Optimization for Dynamic Discrete Environments Using Embedded Ising Machines." pith.science (2026). https://pith.science/paper/ROJSTQ34

@misc{pith2026250616924,
  author       = {Pith},
  title        = {Pith review of: Real-Time Black-Box Optimization for Dynamic Discrete Environments Using Embedded Ising Machines},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ROJSTQ34}},
  note         = {Machine review of arXiv:2506.16924}
}
read the original abstract

Many real-time systems require the optimization of discrete variables. Black-box optimization (BBO) algorithms and multi-armed bandit (MAB) algorithms perform optimization by repeatedly taking actions and observing the corresponding instant rewards without any prior knowledge. Recently, a BBO method using an Ising machine has been proposed to find the best action that is represented by a combination of discrete values and maximizes the instant reward in static environments. In contrast, dynamic environments, where real-time systems operate, necessitate MAB algorithms that maximize the average reward over multiple trials. However, due to the enormous number of actions resulting from the combinatorial nature of discrete optimization, conventional MAB algorithms cannot effectively optimize dynamic, discrete environments. Here, we show a heuristic MAB method for dynamic, discrete environments by extending the BBO method, in which an Ising machine effectively explores the actions while considering interactions between variables and changes in dynamic environments. We demonstrate the dynamic adaptability of the proposed method in a wireless communication system with moving users.

Figures

Figures reproduced from arXiv: 2506.16924 by the authors.

Figure 1
Figure 1. Dynamic discrete environment and Real-Time Black-Box Optimization Algorithm: The dynamic adap [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Dataflow graph of the RT-BBO algorithm: (a) RT-BBO for SR (single instant reward): Pre-training [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Adaptability of the RT-BBO algorithms (i.e., FMSB with dynamic adaptation mechanism) in a dynam [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Experimental results on the prototype for the wireless control system. (a) Snapshots showing the situation [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Comparison of computation time per sampling cycle. The baseline needs longer time than the others [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: A snapshot from the Supplementary information 2 [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

63 extracted references · 29 canonical work pages

  1. [1]

    Solving flexible job shop scheduling problems in manufactur- ing with Quantum Annealing

    P. Schworm, X. Wu, M. Glatt, and J. C. Aurich. “Solving flexible job shop scheduling problems in manufactur- ing with Quantum Annealing”. In: Production engineering 17.1 (2023), pp. 105–115. doi: 10.1007/s11740- 022-01145-8

    work page doi:10.1007/s11740- 2023

  2. [2]

    Online Combinatorial Optimization under Bandit Feedback

    M. S. Talebi. “Online Combinatorial Optimization under Bandit Feedback”. PhD thesis. KTH Royal Institute of Technology, 2016. url: https://www.diva- portal.org/smash/record.jsf?pid=diva2%3A899137& dswid=5096

    work page 2016

  3. [3]

    Combinatorial Bandits Revisited

    R. Combes, M. S. Talebi Mazraeh Shahi, A. Proutiere, and M. Lelarge. “Combinatorial Bandits Revisited”. In: Advances in Neural Information Processing Systems . Vol. 28. Curran Associates, Inc., 2015. url: https: / / proceedings . neurips . cc / paper _ files / paper / 2015 / file / 0ce2ffd21fc958d9ef0ee9ba5336e357 - Paper.pdf

    work page 2015

  4. [4]

    Combinatorial bandits

    N. Cesa-Bianchi and G. Lugosi. “Combinatorial bandits”. In: Journal of Computer and System Sciences 78.5 (2012), pp. 1404–1422. doi: 10.1016/j.jcss.2012.01.001. 14

    doi:10.1016/j.jcss.2012.01.001 2012

  5. [5]

    Designing metamaterials with quantum annealing and factorization machines

    K. Kitai et al. “Designing metamaterials with quantum annealing and factorization machines”. In: Physical Review Research 2.1 (2020), p. 013319. doi: 10.1103/PhysRevResearch.2.013319

    doi:10.1103/physrevresearch.2.013319 2020

  6. [6]

    Towards optimization of photonic-crystal surface-emitting lasers via quantum annealing

    T. Inoue et al. “Towards optimization of photonic-crystal surface-emitting lasers via quantum annealing”. In: Optics Express 30.24 (2022), pp. 43503–43512. doi: 10.1364/OE.476839

    work page doi:10.1364/oe.476839 2022

  7. [7]

    Exploration of new chemical materials using black- box optimization with the D-wave quantum annealer

    M. Doi, Y. Nakao, T. Tanaka, M. Sako, and M. Ohzeki. “Exploration of new chemical materials using black- box optimization with the D-wave quantum annealer”. In:Frontiers in Computer Science 5 (2023), p. 1286226. doi: 10.3389/fcomp.2023.1286226

    work page arXiv 2023

  8. [8]

    Machine Learning Supported Annealing for Prediction of Grand Canonical Crystal Struc- tures

    Y. Couzini´ e et al. “Machine Learning Supported Annealing for Prediction of Grand Canonical Crystal Struc- tures”. In: Journal of the Physical Society of Japan 94.4 (2025), p. 044802. doi: 10.7566/JPSJ.94.044802

    doi:10.7566/jpsj.94.044802 2025

Show all 63 references
  1. [9]

    Application of factorization machine with quantum annealing to hyperparameter optimization and metamodel-based optimization in granular flow simulations

    J. Xiao et al. “Application of factorization machine with quantum annealing to hyperparameter optimization and metamodel-based optimization in granular flow simulations”. In: International Journal for Numerical and Analytical Methods in Geomechanics 48.13 (2024), pp. 3432–3451...

    2024 doi

  2. [10]

    Lossy compression of matrices by black box optimisation of mixed integer nonlinear programming

    T. Kadowaki and M. Ambai. “Lossy compression of matrices by black box optimisation of mixed integer nonlinear programming”. In: Scientific Reports 12.1 (2022), p. 15482. doi: 10.1038/s41598-022-19763-8

    2022 doi

  3. [11]

    Application of QUBO solver using black-box optimization to structural design for resonance avoidance

    T. Matsumori, M. Taki, and T. Kadowaki. “Application of QUBO solver using black-box optimization to structural design for resonance avoidance”. In:Scientific Reports 12.1 (2022), p. 12143. doi: 10.1038/s41598- 022-16149-8

    2022 doi

  4. [12]

    Topology optimization of analog circuit design via global optimization using factorization machines with digital annealer

    M. Hida, H. Ikeda, A. Maruo, M. Sato, and T. Yamazaki. “Topology optimization of analog circuit design via global optimization using factorization machines with digital annealer”. In: Journal of Advanced Mechanical Design, Systems, and Manufacturing 18.6 (2024), JAMDSM0076–JAM...

    2024

  5. [13]

    Adaptive operator selection with dynamic multi-armed bandits

    L. DaCosta, A. Fialho, M. Schoenauer, and M. Sebag. “Adaptive operator selection with dynamic multi-armed bandits”. In: Proceedings of the 10th Annual Conference on Genetic and Evolutionary Computation . GECCO ’08. Atlanta, GA, USA: Association for Computing Machinery, 2008, p...

    2008 doi

  6. [14]

    Thompson sampling for dynamic multi-armed bandits

    N. Gupta, O.-C. Granmo, and A. Agrawala. “Thompson sampling for dynamic multi-armed bandits”. In: 2011 10th International Conference on Machine Learning and Applications and Workshops . Vol. 1. IEEE. 2011, pp. 484–489. doi: 10.1109/ICMLA.2011.144

    2011 doi

  7. [15]

    DBA: dynamic multi-armed bandit algorithm

    S. Nobari. “DBA: dynamic multi-armed bandit algorithm”. In: Proceedings of the AAAI Conference on Arti- ficial Intelligence . Vol. 33. 01. AAAI Press, 2019, pp. 9869–9870. doi: 10.1609/AAAI.V33I01.33019869

    2019 doi

  8. [16]

    Gambling in a rigged casino: The adversarial multi- armed bandit problem

    P. Auer, N. Cesa-Bianchi, Y. Freund, and R. Schapire. “Gambling in a rigged casino: The adversarial multi- armed bandit problem”. In: Proceedings of IEEE 36th Annual Foundations of Computer Science . 1995, pp. 322–331. doi: 10.1109/SFCS.1995.492488

    1995

  9. [17]

    Online Non-convex Learning in Dynamic Environments

    Z. Xu and L. Zhang. “Online Non-convex Learning in Dynamic Environments”. In: The Thirty-eighth Annual Conference on Neural Information Processing Systems . 2024. url: https://openreview.net/forum?id= DrQXDKbGgy

    2024

  10. [18]

    Online Non-Convex Optimization with Imper- fect Feedback

    A. H´ eliou, M. Martin, P. Mertikopoulos, and T. Rahier. “Online Non-Convex Optimization with Imper- fect Feedback”. In: Advances in Neural Information Processing Systems . Vol. 33. Curran Associates, Inc., 2020, pp. 17224–17235. url: https : / / proceedings . neurips . cc / p...

    2020

  11. [19]

    Decoupled weight decay regularization

    I. Loshchilov. “Decoupled weight decay regularization”. 2017. doi: 10.48550/arXiv.1711.05101

    2017 doi arXiv

  12. [20]

    Combinatorial optimization by simulating adiabatic bifurcations in nonlinear Hamiltonian systems

    H. Goto, K. Tatsumura, and A. R. Dixon. “Combinatorial optimization by simulating adiabatic bifurcations in nonlinear Hamiltonian systems”. In: Science advances 5.4 (2019), eaav2372. doi: 10.1126/sciadv.aav2372

    2019 doi

  13. [21]

    High-performance combinatorial optimization based on classical mechanics

    H. Goto et al. “High-performance combinatorial optimization based on classical mechanics”. In: Science Advances 7.6 (2021), eabe7953. doi: 10.1126/sciadv.abe7953

    2021 doi

  14. [22]

    Bifurcation-based adiabatic quantum computation with a nonlinear oscillator network

    H. Goto. “Bifurcation-based adiabatic quantum computation with a nonlinear oscillator network”. In: Scien- tific reports 6.1 (2016), p. 21686. doi: 10.1038/srep21686

    2016 doi

  15. [23]

    FPGA-based simulated bifurcation machine

    K. Tatsumura, A. R. Dixon, and H. Goto. “FPGA-based simulated bifurcation machine”. In: 2019 29th International Conference on Field Programmable Logic and Applications (FPL) . IEEE. 2019, pp. 59–66. doi: 10.1109/FPL.2019.00019. 15

    2019

  16. [24]

    Scaling out Ising machines using a multi-chip architecture for simulated bifurcation

    K. Tatsumura, M. Yamasaki, and H. Goto. “Scaling out Ising machines using a multi-chip architecture for simulated bifurcation”. In: Nature Electronics 4.3 (2021), pp. 208–217. doi: 10.1038/s41928-021-00546-4

    2021 doi

  17. [25]

    Efficient and Scalable Architecture for Multiple- Chip Implementation of Simulated Bifurcation Machines

    T. Kashimata, M. Yamasaki, R. Hidaka, and K. Tatsumura. “Efficient and Scalable Architecture for Multiple- Chip Implementation of Simulated Bifurcation Machines”. In: IEEE Access 12 (2024), pp. 36606–36621. doi: 10.1109/ACCESS.2024.3374089

    2024

  18. [26]

    A Currency Arbitrage Machine Based on the Simulated Bifurcation Algorithm for Ultrafast Detection of Optimal Opportunity

    K. Tatsumura, R. Hidaka, M. Yamasaki, Y. Sakai, and H. Goto. “A Currency Arbitrage Machine Based on the Simulated Bifurcation Algorithm for Ultrafast Detection of Optimal Opportunity”. In: 2020 IEEE International Symposium on Circuits and Systems (ISCAS) . 2020, pp. 1–5. doi: ...

    2020

  19. [27]

    Pairs-Trading System Using Quantum-Inspired Combinatorial Optimization Accelerator for Optimal Path Search in Market Graphs

    K. Tatsumura, R. Hidaka, J. Nakayama, T. Kashimata, and M. Yamasaki. “Pairs-Trading System Using Quantum-Inspired Combinatorial Optimization Accelerator for Optimal Path Search in Market Graphs”. In: IEEE Access 11 (2023), pp. 104406–104416. doi: 10.1109/ACCESS.2023.3316727

    2023

  20. [28]

    Real-Time Trading System Based on Selections of Potentially Profitable, Uncorrelated, and Balanced Stocks by NP-Hard Combinatorial Opti- mization

    K. Tatsumura, R. Hidaka, J. Nakayama, T. Kashimata, and M. Yamasaki. “Real-Time Trading System Based on Selections of Potentially Profitable, Uncorrelated, and Balanced Stocks by NP-Hard Combinatorial Opti- mization”. In: IEEE Access 11 (2023), pp. 120023–120033. doi: 10.1109/...

    2023

  21. [29]

    Enhancing In-vehicle Multiple Object Tracking Systems with Embeddable Ising Machines

    K. Tatsumura, Y. Hamakawa, M. Yamasaki, K. Oya, and H. Fujimoto. “Enhancing In-vehicle Multiple Object Tracking Systems with Embeddable Ising Machines”. 2024. doi: 10 . 48550 / arXiv . 2410 . 14093. arXiv: 2410.14093 [cs.CV]

    2024 doi

  22. [30]

    Distance-based clustering using QUBO formu- lations

    N. Matsumoto, Y. Hamakawa, K. Tatsumura, and K. Kudo. “Distance-based clustering using QUBO formu- lations”. In: Scientific reports 12.1 (2022), p. 2669. doi: 10.1038/s41598-022-06559-z

    2022 doi

  23. [31]

    Deep Reinforcement Learning for 5G Networks: Joint Beam- forming, Power Control, and Interference Coordination

    F. B. Mismar, B. L. Evans, and A. Alkhateeb. “Deep Reinforcement Learning for 5G Networks: Joint Beam- forming, Power Control, and Interference Coordination”. In: IEEE Transactions on Communications 68.3 (2020), pp. 1581–1592. doi: 10.1109/TCOMM.2019.2961332

    2020

  24. [32]

    Deep reinforcement learning for distributed dynamic MISO downlink-beamforming coordination

    J. Ge, Y.-C. Liang, J. Joung, and S. Sun. “Deep reinforcement learning for distributed dynamic MISO downlink-beamforming coordination”. In: IEEE Transactions on Communications 68.10 (2020), pp. 6070–

    2020

  25. [33]

    Deep Reinforcement Learning-Based On-Off Analog Beamform- ing Coordination for Downlink MISO Networks

    H. Zhou, X. Wang, M. Umehira, and Y. Ji. “Deep Reinforcement Learning-Based On-Off Analog Beamform- ing Coordination for Downlink MISO Networks”. In: 2023 IEEE Intl Conf on Dependable, Autonomic and Secure Computing, Intl Conf on Pervasive Intelligence and Computing, Intl Conf...

    2023

  26. [34]

    Factorization Machines

    S. Rendle. “Factorization Machines”. In: 2010 IEEE International Conference on Data Mining . 2010, pp. 995–

    2010

  27. [35]

    Statistical Properties of the log-cosh Loss Function Used in Machine Learning

    R. A. Saleh and A. K. M. E. Saleh. “Statistical Properties of the log-cosh Loss Function Used in Machine Learning”. 2024. doi: 10.48550/arXiv.2208.04564. arXiv: 2208.04564 [stat.ML]

    2024 doi arXiv

  28. [36]

    Customer acquisition via display advertising using multi- armed bandit experiments

    E. M. Schwartz, E. T. Bradlow, and P. S. Fader. “Customer acquisition via display advertising using multi- armed bandit experiments”. In: Marketing Science 36.4 (2017), pp. 500–522. doi: 10.1287/mksc.2016.1023

    2017

  29. [37]

    Dynamic pricing and learning: historical origins, current research, and new directions

    A. V. Den Boer. “Dynamic pricing and learning: historical origins, current research, and new directions”. In: Surveys in operations research and management science 20.1 (2015), pp. 1–18. doi: 10.1016/j.sorms.2015. 03.001

    2015 doi

  30. [38]

    A contextual-bandit approach to personalized news article recommendation

    L. Li, W. Chu, J. Langford, and R. E. Schapire. “A contextual-bandit approach to personalized news article recommendation”. In: Proceedings of the 19th international conference on World wide web. 2010, pp. 661–670. doi: 10.1145/1772690.1772758

    2010

  31. [39]

    Hedging using reinforcement learning: Contextual k-armed bandit versus Q-learning

    L. Cannelli, G. Nuti, M. Sala, and O. Szehr. “Hedging using reinforcement learning: Contextual k-armed bandit versus Q-learning”. In: The Journal of Finance and Data Science 9 (2023), p. 100101. doi: 10.1016/ j.jfds.2023.100101

    2023

  32. [40]

    Ising formulations of many NP problems

    A. Lucas. “Ising formulations of many NP problems”. In: Frontiers in Physics 2 (2014). doi: 10.3389/fphy. 2014.00005

    2014

  33. [41]

    On the computational complexity of Ising spin glass models

    F. Barahona. “On the computational complexity of Ising spin glass models”. In: Journal of Physics A: Math- ematical and General 15.10 (1982), p. 3241. doi: 10.1088/0305-4470/15/10/028. 16

    1982 doi

  34. [42]

    Quantum annealing with manufactured spins

    M. W. Johnson et al. “Quantum annealing with manufactured spins”. In: Nature 473.7346 (2011), pp. 194–

    2011

  35. [43]

    Ising machines as hardware solvers of combinatorial optimization problems

    N. Mohseni, P. L. McMahon, and T. Byrnes. “Ising machines as hardware solvers of combinatorial optimization problems”. In: Nature Reviews Physics 4.6 (2022), pp. 363–379. doi: 10.1038/s42254-022-00440-8

    2022 doi

  36. [44]

    Digital Annealer for High-Speed Solving of Combinatorial optimization Problems and Its Applications

    S. Matsubara et al. “Digital Annealer for High-Speed Solving of Combinatorial optimization Problems and Its Applications”. In: 2020 25th Asia and South Pacific Design Automation Conference (ASP-DAC) . 2020, pp. 667–672. doi: 10.1109/ASP-DAC47756.2020.9045100

    2020

  37. [45]

    STATICA: A 512-Spin 0.25M-Weight Annealing Processor With an All-Spin-Updates-at- Once Architecture for Combinatorial Optimization With Complete Spin–Spin Interactions

    K. Yamamoto et al. “STATICA: A 512-Spin 0.25M-Weight Annealing Processor With an All-Spin-Updates-at- Once Architecture for Combinatorial Optimization With Complete Spin–Spin Interactions”. In:IEEE Journal of Solid-State Circuits 56.1 (2021), pp. 165–178. doi: 10.1109/JSSC.202...

    2021

  38. [46]

    Amorphica: 4-Replica 512 Fully Connected Spin 336MHz Metamorphic Annealer with Programmable Optimization Strategy and Compressed-Spin-Transfer Multi-Chip Extension

    K. Kawamura et al. “Amorphica: 4-Replica 512 Fully Connected Spin 336MHz Metamorphic Annealer with Programmable Optimization Strategy and Compressed-Spin-Transfer Multi-Chip Extension”. In: 2023 IEEE International Solid-State Circuits Conference (ISSCC) . 2023, pp. 42–44. doi:...

    2023

  39. [47]

    100,000-spin coherent Ising machine

    T. Honjo et al. “100,000-spin coherent Ising machine”. In: Science Advances 7.40 (2021), eabh0952. doi: 10.1126/sciadv.abh0952

    2021 doi

  40. [48]

    K. P. Kalinin, A. Amo, J. Bloch, and N. G. Berloff. In: Nanophotonics 9.13 (2020), pp. 4127–4138. doi: 10.1515/nanoph-2020-0162

    2020 doi

  41. [49]

    A poor man’s coherent Ising machine based on opto- electronic feedback systems for solving optimization problems

    F. B¨ ohm, G. Verschaffelt, and G. Van der Sande. “A poor man’s coherent Ising machine based on opto- electronic feedback systems for solving optimization problems”. In: Nature communications 10.1 (2019), p. 3538. doi: 10.1038/s41467-019-11484-3

    2019 doi

  42. [50]

    A 1,968-node coupled ring oscillator circuit for combinatorial optimization problem solving

    W. Moy et al. “A 1,968-node coupled ring oscillator circuit for combinatorial optimization problem solving”. In: Nature Electronics 5.5 (2022), pp. 310–317. doi: 10.1038/s41928-022-00749-3

    2022 doi

  43. [51]

    Ultrafast Ising Machines using spin torque nano-oscillators

    D. I. Albertsson et al. “Ultrafast Ising Machines using spin torque nano-oscillators”. In: Applied Physics Letters 118.11 (2021). doi: 10.1063/5.0041575

    2021 doi

  44. [52]

    Solving combinatorial optimisation problems using oscil- lator based Ising machines

    T. Wang, L. Wu, P. Nobel, and J. Roychowdhury. “Solving combinatorial optimisation problems using oscil- lator based Ising machines”. In: Natural Computing 20.2 (2021), pp. 287–306. doi: 10.1007/s11047- 021- 09845-3

    2021 doi

  45. [53]

    Scaling advantage of chaotic amplitude control for high-performance combinatorial optimiza- tion

    T. Leleu et al. “Scaling advantage of chaotic amplitude control for high-performance combinatorial optimiza- tion”. In: Communications Physics 4.1 (2021), p. 266. doi: 10.1038/s42005-021-00768-0

    2021 doi

  46. [54]

    An integrated coupled oscillator network to solve optimization problems

    M. Graber and K. Hofmann. “An integrated coupled oscillator network to solve optimization problems”. In: Communications Engineering 3.1 (2024), p. 116. doi: 10.1038/s44172-024-00261-w

    2024 doi

  47. [55]

    Black-box optimization for integer-variable problems using Ising machines and factorization machines

    Y. Seki, R. Tamura, and S. Tanaka. “Black-box optimization for integer-variable problems using Ising machines and factorization machines”. 2022. doi: 10.48550/arXiv.2209.01016

    2022 doi arXiv

  48. [56]

    Deep Learning is Robust to Massive Label Noise

    D. Rolnick, A. Veit, S. Belongie, and N. Shavit. “Deep Learning is Robust to Massive Label Noise”. 2018. doi: 10.48550/arXiv.1705.10694

    2018 doi arXiv

  49. [57]

    Adam: A Method for Stochastic Optimization

    D. P. Kingma and J. Ba. “Adam: A Method for Stochastic Optimization”. 2017. doi: 10.48550/arXiv.1412

    2017 doi

  50. [58]

    Correlation-Diversified Portfolio Construction by Finding Maximum Independent Set in Large-Scale Market Graph

    R. Hidaka, Y. Hamakawa, J. Nakayama, and K. Tatsumura. “Correlation-Diversified Portfolio Construction by Finding Maximum Independent Set in Large-Scale Market Graph”. In:IEEE Access 11 (2023), pp. 142979– 142991. doi: 10.1109/ACCESS.2023.3341422

    2023

  51. [59]

    Optimal insertion of pilot symbols for transmissions over time-varying flat fading channels

    M. Dong, L. Tong, and B. Sadler. “Optimal insertion of pilot symbols for transmissions over time-varying flat fading channels”. In: IEEE Transactions on Signal Processing 52.5 (2004), pp. 1403–1418. doi: 10.1109/ TSP.2004.826182. 17 Supplementary Information 1 Supplementary In...

    2004

  52. [198]

    doi: 10.1038/nature10012

    doi

  53. [1000]

    doi: 10.1109/ICDM.2010.127

    2010 doi

  54. [6085]

    doi: 10.1109/TCOMM.2020.3004524

    2020

  55. [6980]

    arXiv: 1412.6980 [cs.LG]

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.