Step Size Research Articles

Q‐Learning Based Adaptive Kalman Filtering With Adaptive Window Length

ABSTRACTIn this article, we propose an adaptive Kalman filtering with adaptive window length based on Q‐learning for dynamic systems with unknown model information. The iteration step length of the Q‐function is quantitatively adjusted through the influence function. The adaptive Kalman filtering algorithm is used to set an appropriate weight matrix for the Q‐function to estimate unknown model parameters. One numerical example and a practice‐oriented case are given to illustrate the effectiveness of the proposed method. It is shown that this filtering can provide state estimates of best accuracy among all the compared methods when the model mismatch and noise statistical characteristics change.

PSO-FDM (Particle Swarm Optimization-Finite Difference Method)-Based Simulation Model of Temperature and Velocity of Full-Scale Continuous Annealing Furnace

Improving the accuracy of the temperature field prediction model for continuous annealing line strips and enhancing the model’s adaptability to full-size strips are key technical challenges in continuous annealing lines. This paper developed a continuous annealing temperature prediction model based on a variable step-size strategy for the heating section, even-heat section, slow-cooling section, and fast-cooling section of the continuous annealing line. To improve the prediction accuracy for different strip sizes, the PSO optimization algorithm was employed to determine the optimal heat transfer coefficient for each strip size. Additionally, due to the limited production of certain strip gauges, providing insufficient data for optimization, this study introduces a combined file approach to address gauge vacancies. The experimental results indicate that the optimized model with variable step size can control the absolute prediction error to less than 4 °C, improving prediction accuracy by 61.9% and prediction speed by 26.8% compared to the traditional equal-step prediction model. This study verified that the merger method is effective for addressing side gauge vacancies, while the proposed method is suitable for resolving middle gauge vacancies. The main technical contribution of this study is the establishment of a high-precision prediction model for continuous annealing temperature of variable step length strips, ensuring high temperature control accuracy for full-gauge strips when passing through the continuous annealing production line.

Lake shape and the characteristics of migration behavior modify Atlantic salmon smolt migration success through lakes.

Migration is a high-risk behavior. For the Atlantic salmon, Salmo salar, migrating from its river nursery area to marine feeding grounds, the magnitude of risk varies with habitat type. Passage through lakes, in particular, is associated with low rates of migration success. Downstream migrating salmon smolts are rheotactic when migrating in rivers, but lakes typically provide poorer directional currents for migrating salmon. In this study we tested if, in the absence of clear navigational cues in lakes, Atlantic salmon smolts switch to a random search strategy to find the outflowing river. We constructed random search simulations to test if lake basin shape has an effect on migration success. We also compared simulated migration characteristics with migrations of salmon smolts through five real lakes for which there are telemetry data for migrating salmon. Correlated random walk simulations showed that a random search strategy could be successful for all lake shapes tested but was more successful in curved (round and elliptical) than rectangular basin shapes. Rectangular basins with the migration start and stoppoints at the ends of the lake had a higher success than those where these points were perpendicular to the axis of the lake. In general, a random walk model predicted the migration success rate of fish tracked through real lakes. However, for two lakes the simulated migration success exceeded that of actual success, suggesting that fish passing through these lakes were not adopting a random search strategy. We speculate that this is the result of either conflicting navigational cues which inhibit smolts from finding the lake outlet or that they temporarily suspend migration (e.g., to feed). Modelling predicted that for small lakes, directional swimming in short steps (ca. 100 m) followed by turns with very low variation from the direction of travel resulted in the highest migration success. For larger lakes, longer step lengths but also with low turn variation (simulated turning angle drawn from distributions of standard deviation 2° and 5° around a mean of 0°) resulted in the highest migration success. We conclude that navigation in downstream migrating salmon smolts switches from rheotaxis in rivers to a random search tactic in lakes except where residual flow cues in some lakes prevent this, at times resulting in suboptimal navigation outcomes.

Theoretical Analysis on Active Polarization Control of Fiber Laser Based on Root Mean Square Propagation Algorithm

High-power linearly polarized fiber lasers are widely used in coherent beam combination, nonlinear frequency conversion, and gravitational wave detection. With the increase in output power, it is challenging for fiber lasers to maintain a high polarization extinction ratio (PER). Combined with intelligent techniques, active polarization control is a prospective method to obtain the laser output with high PER and high stability. We demonstrate a comprehensive model of an active polarization control system. The root mean square propagation (RMS-Prop) algorithm is used to control the non-polarization-maintaining (non-PM) fiber laser to generate linearly polarized laser. The parameters of the RMS-Prop algorithm are theoretically analyzed, including cost function, perturbation amplitude, and global learning rate. The simulation results show that PER is the optimal cost function. When the perturbation amplitude is 0.06 and the global learning rate is 0.6, the system can achieve the optimal control speed and accuracy. By comparison with the stochastic parallel gradient descent (SPGD) algorithm, the RMS-Prop algorithm has an advantage in obtaining higher PER.

Computational joint action: Dynamical models to understand the development of joint coordination.

Coordinating with others is part of our everyday experience. Previous studies using sensorimotor coordination games suggest that human dyads develop coordination strategies that can be interpreted as Nash equilibria. However, if the players are uncertain about what their partner is doing, they develop coordination strategies which are robust to the actual partner's actions. This has suggested that humans select their actions based on an explicit prediction of what the partner will be doing-a partner model-which is probabilistic by nature. However, the mechanisms underlying the development of a joint coordination over repeated trials remain unknown. Very much like sensorimotor adaptation of individuals to external perturbations (eg force fields or visual rotations), dynamical models may help to understand how joint coordination develops over repeated trials. Here we present a general computational model-based on game theory and Bayesian estimation-designed to understand the mechanisms underlying the development of a joint coordination over repeated trials. Joint tasks are modeled as quadratic games, where each participant's task is expressed as a quadratic cost function. Each participant predicts their partner's next move (partner model) by optimally combining predictions and sensory observations, and selects their actions through a stochastic optimization of its expected cost, given the partner model. The model parameters include perceptual uncertainty (sensory noise), partner representation (retention rate and internale noise), uncertainty in action selection and its rate of decay (which can be interpreted as the action's learning rate). The model can be used in two ways: (i) to simulate interactive behaviors, thus helping to make specific predictions in the context of a given joint action scenario; and (ii) to analyze the action time series in actual experiments, thus providing quantitative metrics that describe individual behaviors during an actual joint action. We demonstrate the model in a variety of joint action scenarios. In a sensorimotor version of the Stag Hunt game, the model predicts that different representations of the partner lead to different Nash equilibria. In a joint two via-point (2-VP) reaching task, in which the actions consist of complex trajectories, the model captures well the observed temporal evolution of performance. For this task we also estimated the model parameters from experimental observations, which provided a comprehensive characterization of individual dyad participants. Computational models of joint action may help identifying the factors preventing or facilitating the development of coordination. They can be used in clinical settings, to interpret the observed behaviors in individuals with impaired interaction capabilities. They may also provide a theoretical basis to devise artificial agents that establish forms of coordination that facilitate neuromotor recovery.

A multi-strategy improved sparrow search algorithm for indoor AGV path planning

To address the problems of weak search ability, easily falling into local optimal solutions and poor path quality of sparrow search algorithm in AGV path planning, a multi-strategy improved sparrow search algorithm (MISSA) is proposed in this paper. MISSA improves the global search ability by improving the discoverer position update operator and introducing the sine cosine algorithm; adopts the adaptive number of vigilantes and adaptive adjustment step size to improve the convergence speed; introduces the Levy flight variation strategy to reduce the probability of falling into any local optimal solution; optimizes the boundary handling mechanism to prevent the loss of population diversity at a later stage; finally, uses the large-scale neighborhood search strategy and path smoothing mechanism for path optimization to further improve the path quality. The superiority-seeking ability of MISSA was verified by 12 standard test functions, and then 30 simulation experiments were conducted in grid maps with two specifications of 20×20 and 30×30. The experimental results showed that, by using MISSA, the path length was reduced by 44.1% and 63.1%, the number of turns was reduced by 68.4% and 78.4%, and the risk degree was reduced by 61.3% and 77.2%, which verifies the superiority of MISSA in path planning. Finally, MISSA was ported to the QBot2e mobile robot for physical verification to prove its feasibility in practical applications.

Molecular dynamics study on the influence of temperature on the mechanical properties of graphene reinforced magnesium-matrix composites

Magnesium and its alloys have been considered the lightest engineering structural metal materials, but their poor plasticity and the degradation of mechanical properties at high temperatures limit their wide application. Therefore, using molecular dynamics simulation, this work examines how the phase structure and mechanical characteristics of pure magnesium under compressive loading are affected by graphene and varying temperatures. The findings demonstrate that adding graphene can significantly increase magnesium's strength. As well as its Young's modulus, and have a great influence on the secondary strain's strengthening. It is noted that the Gr/Mg-based composites have the same plastic deformation behavior as that of pure magnesium, and both of them undergo a transition from a dense hexagonal structure to other structures during plastic deformation, and the introduction of graphene makes the transition more drastic. The number of embedded graphene layers governs the overall deformation of the Gr/Mg composites; however, it does not significantly affect the deformation behavior of the magnesium matrix in the upper and lower regions of the composites. When there is no graphene embedded, the phase transition of magnesium matrix occurs everywhere, while when graphene is embedded, the phase transition of magnesium matrix in the upper and lower parts of graphene extends along the near-interfacial position of graphene without passing through it, which means graphene has the role of changing the nucleation position of dislocations and blocking the movement of dislocations. In addition, the temperature decreases the phase transition step size of Gr/Mg composites, which is similar to that of pure magnesium, while the temperature affects the phase structure of pure magnesium Gr/Mg composites.

Self-assembled nanostructured membranes with tunable pore size and shape from plant-derived materials.

Nanostructured materials derived from sustainable sources are of interest as viable alternatives to traditional petroleum-derived sources in membrane applications due to environmental concerns. Here, we present the development of pore size-tunable nanostructured polymer membranes based on a plant-derived material. The membranes were fabricated using a tri-functional amine as the templating core species and a cross-linkable ligand synthesized from rose oil-derived citronellol. The self-assembly of a supramolecular complex between the template core and the ligand forms a hexagonally packed columnar (Colh) mesophase, the dimensions of which can be precisely controlled by changing the stoichiometric ratio between these constituents. Within the hexagonal mesophase stoichiometric range, the pore size of the nanostructured membranes can be tuned from 1.0 to 1.3 nm, with a step size of approximately 0.1 nm. The membranes exhibited a clear distinction in molecular size selectivity, as demonstrated by dye adsorption experiments. The membrane fabricated with a ligand-to-core ratio of 3 to 1 demonstrated shape-based selectivity, exhibiting a higher permeability for propeller-shaped penetrants and highlighting its potential for shape-selective transport. We anticipate that this straightforward approach, using plant-derived materials, can contribute to important sustainability aspects while enhancing the performance of current state-of-the-art nanostructured membranes by enabling precise control over pore size.

Computational processes of simultaneous learning of stochasticity and volatility in humans

Making adaptive decisions requires predicting outcomes, and this in turn requires adapting to uncertain environments. This study explores computational challenges in distinguishing two types of noise influencing predictions: volatility and stochasticity. Volatility refers to diffusion noise in latent causes, requiring a higher learning rate, while stochasticity introduces moment-to-moment observation noise and reduces learning rate. Dissociating these effects is challenging as both increase the variance of observations. Previous research examined these factors mostly separately, but it remains unclear whether and how humans dissociate them when they are played off against one another. In two large-scale experiments, through a behavioral prediction task and computational modeling, we report evidence of humans dissociating volatility and stochasticity solely based on their observations. We observed contrasting effects of volatility and stochasticity on learning rates, consistent with statistical principles. These results are consistent with a computational model that estimates volatility and stochasticity by balancing their dueling effects.

Improved active vibration control of a cantilever beam using MFC actuators with Hammerstein model-based hysteresis modeling and VSS-FxLMS control algorithm

This study aims to enhance active vibration control in a cantilever beam using macro fiber composite (MFC) actuators. To address the challenge of accurately modeling the complex hysteresis behavior of MFC actuators, an advanced hysteresis modeling technique is employed, integrating a non-symmetric Bouc–Wen model with an Auto-Regressive model with eXogenous inputs (ARX). The parameter estimation of this model is optimized using an adaptive mutation factor, and multiple mutation strategy differential evolution (AmFMMDE) algorithm. Additionally, a Variable Step Size Filtered-x Least Mean Square (VSS-FxLMS) algorithm with step size adaptation based on normalized error change is utilized for effective vibration control. This approach shows promising potential for practical engineering applications requiring precise and efficient active vibration control.

Optimizing neural network models for predicting nuclear reactor channel temperature: A study on hyperparameter tuning and performance analysis

This study emphasizes how important accurate prediction of channel temperatures in nuclear reactors is for safety and operational efficiency. While traditional methods require long and complex processes such as kernel modeling and mathematical simulations, artificial neural networks (ANN) provide more efficient predictions by accelerating this process. The superior ability of ANNs to process large data sets is intended to demonstrate that this study will provide a valuable alternative compared to conventional methods and increase the accuracy of reactor temperature predictions. In this study, the training performances of Artificial Neural Network (ANN) developed to determine the nuclear reactor channel temperature with different hyperparameter combinations were analysed. It was conducted several experimental studies to assess the influence of hyperparameters on our model for nuclear reactor parameter data prediction. The training and validation results indicates that learning rate, hidden layer sizes and number have critical effects for the more precisive prediction. It was observed that models with a learning rate of 0.05 and 0.5 achieved successful learning with less fluctuation in training and validation errors. When looking at hidden layer sizes, networks with 32 and 64 neurons performed better than networks with 16 neurons. For the test phase our model can successfully predict data with slight error margin. As a result, we demonstrated that neural networks are a powerful tool in nuclear reactor channel temperature prediction through our proposed model.

A Novel Nonlinear Stochastic-Resonance-Enhanced Acceleration Data Processing Approach

Abstract. Smartphones are indispensable tools in modern social life, and they can be used for online shopping, electronic payment, gaming, and navigation. In particular, low-cost inertial measurement unit (IMU) sensors are widely integrated into smartphones, so pedestrian dead reckoning (PDR) positioning techniques based on smartphone IMU sensors have been applied and developed. PDR positioning techniques require acceleration data for step detection, step length estimation, and step heading estimation. However, due to the cost limitations of the built-in IMU sensor in smartphones, acceleration data contains measurement noise and interference, resulting in poor consistency in acceleration peak detection and the generation of false peaks, which is not conducive to step detection and accurate step length estimation. Therefore, this paper proposes a stochastic resonance (SR) enhancement method for smartphone IMU acceleration data. The SR-enhanced acceleration data has better peak consistency and is conducive to step detection. Finally, the algorithm is evaluated using actual measurement data collected from a smartphone. The results show that the SR-enhanced acceleration data has excellent peak consistency and higher step detection accuracy.

Evaluating the influence of step size in macro-raman mapping experiments

Evaluating the influence of step size in macro-raman mapping experiments

Multi‐Objective Optimization and Experimental Research of Ship Form Based on Improved Bare‐Bones Multi‐Objective Particle Swarm Optimization Algorithm

ABSTRACTShip form optimization poses a complex and high‐dimensional engineering challenge. Therefore, when conducting multi‐objective optimization research of ship forms, traditional intelligent optimization algorithms are prone to falling into local optima solution and difficult to converge. In order to effectively improve the diversity and convergence performance of the algorithm, this paper improves the bare‐bones multi‐objective particle swarm optimization (BBMOPSO) algorithm by dynamically adjusting the local and global search step sizes, and verifies the algorithm's reliability through standard function testing. Then, a multi‐objective optimization design framework with high efficiency and high integration is constructed. Taking DTMB 5512 as the research case, Free Form Deformation (FFD) method is used for hull deformation, and the proposed algorithm is used for multi‐objective optimization of resistance performance and motion response. Ship model tests were conducted on the DTMB 5512's original hull. And the numerical simulations were compared with the ship model tests. Finally, under the constructed multi‐objective optimization design framework, satisfactory solutions were obtained through the improved algorithm, which confirms the effectiveness and practicality of the improved algorithm. The results show that the algorithm improved in this paper can provide some theoretical basis and technical support for green ship design and low‐carbon shipping.

Mixing enhancement with minimal thrust loss for the Mach 1.76 jet issuing from different beveled collar nozzles

The main objective of the present investigation is to enhance the jet mixing by introducing a nozzle collar exit inclination having bevel angles of 30°, 45°, and 60°. In addition to this, the present study also addresses the problem of investigating the jet which attains maximum mixing from beveled collar nozzle with the imposition of a minimal overall thrust loss penalty. The nozzles of the present study have been designed to deliver a uniform Mach 1.76 jet at the nozzle collar exit. The results thus obtained through numerical computations have been compared with that of the reference nozzle which has zero exit inclination. The investigations have been performed under both design and off-design conditions of the jet by varying the nozzle pressure ratio (NPR) from 4.5 to 6.5 with unit step size. It is seen that the rate of jet mixing is augmented with the increase in the bevel angle. Thus, the 60° beveled collar nozzle demonstrated the highest jet mixing in comparison to the reference nozzle, 30° and 45° beveled collar nozzle at each nozzle pressure ratio. However, 60° beveled collar nozzle has been substantially penalized with the highest thrust loss of about 6.5% at nozzle pressure ratio 4.5 and about 3.5% at nozzle pressure ratio 5.5 with respect to the reference nozzle and other nozzles of the present study. The 45° and 30° beveled collar nozzles reported the thrust penalty below 2.3% and 1%, respectively. The overall performance, which includes mixing enhancement along with the minimal thrust loss, has been found in the 45° beveled collar nozzle, which clearly justifies its preference over the other investigated nozzle configurations.

Abrupt and spontaneous strategy switches emerge in simple regularised neural networks.

Humans sometimes have an insight that leads to a sudden and drastic performance improvement on the task they are working on. Sudden strategy adaptations are often linked to insights, considered to be a unique aspect of human cognition tied to complex processes such as creativity or meta-cognitive reasoning. Here, we take a learning perspective and ask whether insight-like behaviour can occur in simple artificial neural networks, even when the models only learn to form input-output associations through gradual gradient descent. We compared learning dynamics in humans and regularised neural networks in a perceptual decision task that included a hidden regularity to solve the task more efficiently. Our results show that only some humans discover this regularity, whose behaviour was marked by a sudden and abrupt strategy switch that reflects an aha-moment. Notably, we find that simple neural networks with a gradual learning rule and a constant learning rate closely mimicked behavioural characteristics of human insight-like switches, exhibiting delay of insight, suddenness and selective occurrence in only some networks. Analyses of network architectures and learning dynamics revealed that insight-like behaviour crucially depended on a regularised gating mechanism and noise added to gradient updates, which allowed the networks to accumulate "silent knowledge" that is initially suppressed by regularised gating. This suggests that insight-like behaviour can arise from gradual learning in simple neural networks, where it reflects the combined influences of noise, gating and regularisation. These results have potential implications for more complex systems, such as the brain, and guide the way for future insight research.

Dynamics of finite width Kernel and prediction fluctuations in mean field neural networks*

Abstract We analyze the dynamics of finite width effects in wide but finite feature learning neural networks. Starting from a dynamical mean field theory description of infinite width deep neural network kernel and prediction dynamics, we provide a characterization of the O ( 1 / width ) fluctuations of the dynamical mean field theory order parameters over random initializations of the network weights. Our results, while perturbative in width, unlike prior analyses, are non-perturbative in the strength of feature learning. We find that once the mean field/µP parameterization is adopted, the leading finite size effect on the dynamics is to introduce initialization variance in the predictions and feature kernels of the networks. In the lazy limit of network training, all kernels are random but static in time and the prediction variance has a universal form. However, in the rich, feature learning regime, the fluctuations of the kernels and predictions are dynamically coupled with a variance that can be computed self-consistently. In two layer networks, we show how feature learning can dynamically reduce the variance of the final tangent kernel and final network predictions. We also show how initialization variance can slow down online learning in wide but finite networks. In deeper networks, kernel variance can dramatically accumulate through subsequent layers at large feature learning strengths, but feature learning continues to improve the signal-to-noise ratio of the feature kernels. In discrete time, we demonstrate that large learning rate phenomena such as edge of stability effects can be well captured by infinite width dynamics and that initialization variance can decrease dynamically. For convolutional neural networks trained on CIFAR-10, we empirically find significant corrections to both the bias and variance of network dynamics due to finite width.

A Novel FBG Placement Optimization Method for Tunnel Monitoring Based on WOA and Deep Q-Network

By employing the whale optimization algorithm’s (WOA) capability to reduce the probability of being stuck in a locally optimal solution, this study proposed an improved WOA-DQN algorithm based on the Deep Q-Network algorithm (DQN). Firstly, the mathematical model of Fiber Bragg Grating (FBG) sensor placement was established to calculate the reward of DQN. Secondly, the effectiveness and applicability of WOA-DQN were validated through experiments in nine cases. It indicated that the algorithm is far superior to other methods (Noisy DQN, Prioritized DQN, DQN, WOA), especially with the learning rate of 0.001, the initial noise 0.4, the hidden layer 3–512, and the updated frequency of 20. Finally, the FBG sensors were placed at [0°, 27°, 30°, 47°, 51°, 111°, 126°, 219°, 221°, 289°] to detect the accurate deformation of the tunnel with the maximum error 8.66 mm, which is better than the traditional placement. In conclusion, the algorithm provides a theoretical foundation for sensor placement and improves monitoring accuracy. It further shows great promise for deformation monitoring in tunnels.

Stochastic collapse: how gradient noise attracts SGD dynamics towards simpler subnetworks*

Abstract In this work, we reveal a strong implicit bias of stochastic gradient descent (SGD) that drives overly expressive networks to much simpler subnetworks, thereby dramatically reducing the number of independent parameters, and improving generalization. To reveal this bias, we identify invariant sets, or subsets of parameter space that remain unmodified by SGD. We focus on two classes of invariant sets that correspond to simpler (sparse or low-rank) subnetworks and commonly appear in modern architectures. Our analysis uncovers that SGD exhibits a property of stochastic attractivity towards these simpler invariant sets. We establish a sufficient condition for stochastic attractivity based on a competition between the loss landscape’s curvature around the invariant set and the noise introduced by stochastic gradients. Remarkably, we find that an increased level of noise strengthens attractivity, leading to the emergence of attractive invariant sets associated with saddle-points or local maxima of the train loss. We observe empirically the existence of attractive invariant sets in trained deep neural networks, implying that SGD dynamics often collapses to simple subnetworks with either vanishing or redundant neurons. We further demonstrate how this simplifying process of stochastic collapse benefits generalization in a linear teacher-student framework. Finally, through this analysis, we mechanistically explain why early training with large learning rates for extended periods benefits subsequent generalization.

Adaptive learning rate in dynamical binary environments: the signature of adaptive information processing

Adaptive learning rate in dynamical binary environments: the signature of adaptive information processing