28 results about "Function approximation" patented technology

The invention discloses a mobile robot autonomous navigation planning method based on brain-like pulse deep reinforcement learning. Firstly, a brain-like pulse neural network (SNN) is constructed as a Q function approximator, and the energy efficiency of path planning is improved by using the time sequence processing capability and biological rationality of the SNN. Secondly, designing a double-priority experience playback mechanism, dividing experience samples into a common experience pool and an elite experience pool according to a reward threshold value, adjusting the sampling proportion of the two experience pools by adopting a dynamic self-adaptive sampling strategy, and intelligently balancing the sample utilization rate in the training process according to a loss function and a training stage; in addition, the reward function design is optimized through Manhattan distance reward and a repeated access punishment mechanism, and the robot is guided to efficiently explore the environment.

The invention discloses a security constraint unit commitment method based on tensor completion approximate dynamic programming, and the method comprises the following steps: firstly, building a security constraint unit commitment model based on a Markov decision process; secondly, decoupling the multi-period security constraint unit commitment model into a single-period sub; then, obtaining a value function and a decision function by adopting an approximate dynamic programming algorithm based on tensor completion; according to the approximate dynamic programming algorithm based on tensor completion, the approximate dynamic programming algorithm is improved from the angle of tensor completion, and all value functions and decision functions are approximately obtained by sampling a small number of state variables and decision variables. Decision is made from the angle of tensor complementation, in the face of a discrete state space, the value function of the whole state space can be approximated only by sampling a small number of state points, the calculation burden of an approximate dynamic programming algorithm in the aspect of value function approximation is effectively relieved, and the calculation efficiency of the algorithm can be greatly improved.

The application provides a vibration isolation method, system and device of a control moment gyroscope system, which can be applied to the technical field of control. The method comprises the following steps: determining current kinematic parameters based on disturbance forces received by multiple components during operation; constructing a differential equation representing a dynamic relationship based on the current kinematic parameters and coupled dynamic characteristics; inputting a system state vector into a disturbance estimation network to output a disturbance estimation value; inputting the system state vector into the differential equation to obtain a disturbance theoretical value; inputting the disturbance estimation value and the disturbance theoretical value into a Hamilton function to obtain an initial control input quantity corresponding to a Hamilton function value satisfying a preset constraint condition; and updating the initial control input quantity in a gradient descent direction of a performance index function approximation value corresponding to the initial control input quantity until a difference between adjacent two performance index function approximation values is less than a preset threshold value, so as to obtain a target control input quantity.

The application discloses a kind of based on attention mechanism and reinforcement learning's ultra-short-term DNI prediction method, comprising: collecting solar thermal power station historical measured direct normal irradiance data and meteorological characteristic data;Periodic time characteristics and historical lag term are introduced to enhance time series dependence;Utilize bidirectional long short-term memory network to extract DNI time series features, construct attention mechanism based on dynamic change of cloud amount, strengthen the response capability of model to weather mutation;Adopt KAN to replace traditional full connection to high-dimensional nonlinear characteristics Adaptive function approximation is carried out;Introduce PPO reinforcement learning algorithm, dynamically optimize model feature representation and parameter, realize error self-feedback correction;Output DNI prediction result, for the dispatching and operation control of concentrating solar thermal power generation system.The method can effectively improve the DNI prediction accuracy and robustness under complex weather conditions, provide high-precision prediction support for photo-thermal power generation and renewable energy grid connection.

The invention belongs to the technical field of robot intelligent control, and particularly relates to a mechanical arm tracking control method based on transferable depth incremental reinforcement learning, which effectively solves the nonlinear optimal tracking control problem of a mechanical arm in a model-free mode by constructing a transferable incremental reinforcement learning framework. The method comprises the following steps: firstly, constructing a 1-degree-of-freedom depth increment model by utilizing one-step forward data offline learning, and providing universal dynamic representation for a mechanical arm which is difficult to accurately model; furthermore, an asynchronous depth value network is designed for the one-degree-of-freedom mechanical arm, stable and rapid convergence value function approximation is achieved through a separation base layer and an adaptive layer, and a cross-mechanical-arm migration mechanism is established, so that a pre-training model and a network base layer on the one-degree-of-freedom mechanical arm can be directly migrated to a high-degree-of-freedom mechanical arm subsystem; and the system difference is compensated only by updating the adaptive layer online, so that the repeated training overhead is remarkably reduced, and meanwhile, rapid, robust and adaptive tracking control on the mechanical arms with different degrees of freedom is realized.

The application relates to the technical field of computer vision and image processing, in particular to a light, fast and accurate self-supervised depth estimation method and system, which can solve the problem of recovering the depth of a scene from a two-dimensional image and needing a complex depth estimation network model to a certain extent. The method comprises the following steps: acquiring an initial image data set; constructing a weighted coefficient matrix calculation module and a sparse coding module; based on the weighted coefficient matrix calculation module and the sparse coding module, a small depth estimation network model is constructed to speed up the algorithm reasoning speed and reduce the parameter quantity of the model; a large depth estimation network model is constructed and trained by using luminosity loss as a supervision signal; the small depth estimation network model is trained by using luminosity loss and function approximation loss as supervision signals to optimize the depth estimation precision of the small depth estimation network model; finally, input data is input into the small depth estimation network model to obtain an estimation result.

A computer-implemented system and method is provided for determining a risk assessment. The method comprises receiving a plurality of vehicle behaviour data over a defined data collection period. This data is input into a supervised learning prediction model which is trained on historical vehicle behaviour data over a past time period, to generate a predicted value of a frequency of expected claim submissions for the policyholder of the vehicle in a future time period. Then a Shapley estimate is computed for each feature of the behaviour data applied to the model for determining a contribution of each said feature to the predicted value. A spline approximation is applied to the Shapley estimate for each said feature to estimate the contribution of each said feature. Then, a sum of the spline approximation for each said feature is calculated and a corresponding risk score determined based on the sum.

This application belongs to the field of delay differential equation modeling, specifically disclosing a delay differential equation modeling method based on Bayesian optimization and neural networks. The method includes: generating multiple trajectory data of the system; searching for the optimal delay term using a Bayesian optimization algorithm, where the delay term to be solved is the optimization variable, and the error based on neural network simulation is the objective function, obtaining the optimal delay term through iterative optimization using a surrogate model; constructing a state matrix and a delay matrix from the trajectory data based on the optimal delay term, using the concatenated matrix as input to a neural network to construct a neural network for approximating a nonlinear function; integrating a linear multi-step method into the loss function of the neural network, training the neural network using the trajectory data to obtain a nonlinear function approximation model; and separating the numerical discretization error and the neural network approximation error based on this model to construct a total error bound. This application can achieve high-precision, high-efficiency delay differential equation modeling with quantization error guarantee.

Methods and devices of a function approximation unit configured to approximate a nonlinear function within a neural processing unit are described. According to one embodiment, the method includes storing an input value through an input register of the function approximation unit, transmitting the input value to a selected one of a plurality of preprocessing circuits of the unit according to a control signal, generating a preprocessing result corresponding to the input value by the selected one of the preprocessing circuits, transmitting the preprocessing result to a programmable function approximation circuit and a selected one of a plurality of post-processing circuits of the unit, generating an approximated function output based on the preprocessing result in the programmable function approximation circuit, and generating a final output value by post-processing the preprocessing result or the approximated function output in the selected one of the post-processing circuits.

This invention relates to the field of voice coil motor servo control technology, specifically to a fractional-order high-order sliding mode control method for voice coil motors based on a third-order superspiral sliding mode observer (STSMO) and an adaptive reaching law. The method includes the following steps: S1, establishing a mathematical model of the voice coil motor, deriving the dynamic equations of the voice coil motor including parameter uncertainties and external disturbances, defining the position tracking error, and deriving the error dynamic equation; S2, designing a fractional-order high-order integral sliding surface, combining it with an improved Oustaloup filter to achieve rational function approximation of the fractional-order calculus operator, and constructing the first-order derivative equation of the sliding surface; S3, designing an adaptive hyperbolic tangent reaching law that integrates dynamic feedback of the error amplitude, and deriving the initial expression of the sliding mode control law based on the first-order derivative equation of the sliding surface and the adaptive hyperbolic tangent reaching law. This invention achieves real-time accurate observation and feedforward compensation of lumped disturbances through a third-order superspiral sliding mode observer (STSMO).

The invention discloses an urban road traffic state intelligent estimation method based on microphysical representation, and the method comprises the steps: constructing a road section boundary condition based on the arrival and departure accumulated flow information of an intersection and a gate, and forming multi-source traffic perception input through combining with road section observation data; constructing a physical-data hybrid driving model, constructing a deep learning model in a data driving branch, learning a mapping relation between road section boundary cumulative flow and a traffic state in a road section, and converting an indistinguishable Newell traffic flow model into a distinguishable computational graph structure through function approximation and structural conversion in a distinguishable physical branch; and constructing a loss function of the hybrid drive model, so that data drive output and a physical model are kept coordinated, and traffic flow parameters are jointly estimated. Through deep embedding of the microphysical calculation graph, the dependence of a pure physical model on an ideal assumed condition is made up, and meanwhile, the problem that a pure data driving method is insufficient in generalization ability in a sensing blind area is solved.

The invention provides a trigonometric function calculation circuit and calculation method based on piecewise quadratic function approximation, and electronic equipment. The calculation circuit comprises an angle mapping unit and a numerical calculation unit, the angle mapping unit is configured to receive an input angle, determine sign bits of a sine value and a cosine value according to a quadrant where the input angle is located, and map the input angle into a mutual complementary angle with the sum of two actual values being 90 degrees in a first quadrant; and the numerical calculation unit is configured to calculate a sine value and a cosine value based on mutual complement angles by adopting a piecewise quadratic function approximation mode, and output a final result in combination with the sign bits. According to the method, efficient calculation of the trigonometric function is achieved through angle mapping and piecewise quadratic function approximation, non-equal-length segmentation and special left end point optimization are adopted in a numerical calculation unit, and storage resources and calculation complexity are reduced. While the precision is guaranteed, hardware resources and calculation delay are remarkably reduced, and the method is suitable for application scenes with high requirements for real-time performance.

The invention discloses an unmanned aerial vehicle communication multi-priority queue modeling and performance analysis method, and the method comprises the steps: building an extensible state model through a WiMarkov chain and a hypercube transfer unit, and achieving the low-delay access of different priorities through a threshold and a window in combination with a weighted COS statistics and fixed backoff mechanism; semi-analytical expressions of the success rate and the time delay are obtained through steady-state distribution and generation function approximation, and parameters are learned through optimization constraints. Simulation shows that the low-load success rate is larger than or equal to 99%, the high-priority delay is still kept at the millisecond magnitude during high load, and the method is suitable for an unmanned aerial vehicle network communication system with the strict real-time requirement.

Methods and systems related to neural networks or other function approximators operate for training a neural network is provided, or another function approximator, for inferring a predetermined time of arrival of a predetermined transmitted signal on the basis of channel-impulse-responses, CIRs, of transmitted signals between a mobile antenna and a fixed antenna, the method having: obtaining a channel impulse response condition characteristic, CIRCC, descriptive of channel impulse responses of transmitted signals associated with mobile antenna positions within a reach of the fixed antenna; generating, by simulation, a training set of simulated CIRs which are associated with different times of arrival in one or more simulated scenes, and which fit to the CIRCC; training the neural network, or other function approximator, using the simulated CIRs and the different associated times of arrivals to obtain a parametrization of the neural network, or other function approximator, associated with the CIRCC.

This application provides a method and system for learning and imitating the three-dimensional trajectory of a robotic arm. During human-computer interaction between a patient's unaffected hand and the robot, a teaching trajectory is collected. The collected three-dimensional trajectory data is preprocessed and then input into a Dynamic Management Model (DMP) to obtain multiple sets of nonlinear terms. A Gaussian Mixture Model (GMM) is used to cluster these nonlinear terms, resulting in multiple Gaussian distribution models. Regression is then performed to determine the target nonlinear term. A Gaussian kernel function is used to approximate the target nonlinear term, and the weights of the Gaussian kernel function are calculated to obtain the learned nonlinear term. The learned nonlinear term is then substituted into the DMP model to solve for the trajectory, obtaining optimized teaching trajectory data. The patient can then train by imitating the optimized trajectory. This demonstration trajectory can preserve the patient's motion characteristics while reducing the impact of non-smooth and unrepresentative trajectories on trajectory imitation, resulting in a smooth imitation trajectory with good convergence at the start and end points.

The present application relates to the technical field of charging facility optimization, and particularly relates to a truck type mobile charging station online collaborative optimization scheduling method, a two-stage optimization scheduling model framework is constructed, the optimization scheduling model framework comprises: an offline training stage, a TMCS multi-period optimization decision model is established, and then a look-ahead rolling value function approximation algorithm (LRH-VFA) is established to iteratively learn from EV charging historical data, so that the influence of current period decision on future profit of the operator is considered; an online scheduling stage, based on the approximate value function obtained through offline training and short-time prediction and real-time information, TMCS online scheduling decision is updated rolling. The present application can fully consider the influence of EV charging demand uncertainty on TMCS scheduling results, effectively utilize the demand change adjustment decision updated dynamically, guarantee the quality of EV charging service, and improve the operator's income by coordinating TMCS to participate in power grid energy arbitrage.

The invention provides a timing estimation method and device, electronic equipment, a storage medium and a product. The method comprises the steps of obtaining a raised cosine filtering peak value, a peak value left-side correlation value and a peak value right-side correlation value; based on the peak value, the peak value left-side correlation value and the peak value right-side correlation value, determining a target loss function and a gradient value corresponding to the target loss function; iteratively decreasing and updating the gradient value based on a preset adjustment step length; and when the updated gradient value meets a target threshold value, determining a timing estimation result. Therefore, compared with an existing estimation method based on quadratic function approximate fitting, the method has the advantages that the target loss function is constructed by directly utilizing the peak value and related value information of the raised cosine filter, continuous optimization is carried out in a gradient descent iteration mode, and estimation can be carried out by fitting the actual characteristics of the impulse response function of the raised cosine filter better.

The invention belongs to the field of delay differential equation modeling, and particularly discloses a delay differential equation modeling method based on Bayesian optimization and a neural network, and the method comprises the steps: generating multiple pieces of trajectory data of a system; an optimal delay term is searched by adopting a Bayesian optimization algorithm, the Bayesian optimization takes a delay term to be solved as an optimization variable and an error based on neural network simulation as a target function, and the optimal delay term is obtained through iterative optimization of an agent model; constructing a state matrix and a delay matrix from the trajectory data based on the optimal delay term, taking the spliced matrix as neural network input, and constructing a neural network for approaching a nonlinear function; fusing a linear multi-step method into a loss function of the neural network, and training the neural network by using the trajectory data to obtain a nonlinear function approximation model; based on the model, a numerical discretization error and a neural network approximation error are separated, and a total error bound is constructed. According to the method, high-precision and high-efficiency delay differential equation modeling with quantization error guarantee can be realized.

The application discloses a kind of based on attention mechanism and reinforcement learning's ultra-short-term DNI prediction method, comprising: collecting solar thermal power station historical measured direct normal irradiance data and meteorological characteristic data;Periodic time characteristics and historical lag term are introduced to enhance time series dependence;Using bidirectional long short-term memory network extracts DNI time series features, constructs attention mechanism based on the dynamic change of cloud amount, and strengthens the response capability of model to weather mutation;Adopt KAN to replace traditional fully connected to high-dimensional nonlinear characteristics Adaptive function approximation;Introducing PPO reinforcement learning algorithm, dynamically optimizing model feature representation and parameters, realizing error self-feedback correction;Output DNI prediction result, for the dispatching and operation control of concentrating solar thermal power generation system.The method can effectively improve the DNI prediction accuracy and robustness under complex weather conditions, and provide high-precision prediction support for photo-thermal power generation and renewable energy grid connection.

The application relates to the technical field of image processing, and discloses a compressed sensing image fast reconstruction system based on function approximation, which comprises a data acquisition module for acquiring compressed sensing measurement data; a function approximation module for generating an initial image estimate by using a pre-trained deep neural network; a collaborative optimization reconstruction module for regarding multiple regularization terms as players in a cooperative game, dynamically evaluating the marginal contribution of each regularization term based on a cooperative game model through a Shapley value, and adaptively adjusting a weight coefficient; and an iterative updating module for combining the adaptive weight to perform multi-criteria collaborative optimization iteration until convergence. The application can solve the problem of difficult manual adjustment of the weights of the multiple regularization terms, realize dynamic balance of noise suppression and structure reservation according to image content features, and improve the quality and stability of the reconstructed image.

This invention provides a federated forgetting algorithm based on multi-level knowledge distillation and influence function approximation, belonging to the field of federated forgetting algorithm technology. The algorithm includes the following steps: Step 1: Define input and output; Step 2: Construct a federated learning environment; Step 3: Local forgetting processing; Step 4: Buffer layer aggregation; Step 5: Global model reconstruction; Step 6: Knowledge distillation recovery; Step 7: Forgetting result evaluation. This invention designs a dual-objective joint optimization local forgetting strategy, simultaneously optimizing both forgetting integrity and utility preservation, enabling the model to effectively remove the influence of forgotten data while maximizing predictive performance on retained data.

A device according to one embodiment disclosed herein includes a network configuration unit that uses the generator of the Koopman operator to represent a first neural network defined by the flow of a plurality of differential equations and a function included in a function space defined by a plurality of Fourier functions, and approximates the generator by the Fourier function to configure a second neural network that approximates the first neural network.

The invention discloses a time series data prediction method and device, equipment and a medium, and the method comprises the steps: obtaining a system state of a to-be-predicted target object, and the system state comprises behavior data of the target object at a preset number of historical time points; and determining a prediction result corresponding to the target object through a time sequence prediction model based on the system state. According to the invention, the time sequence enhancement module comprising the Cauchy activation function is introduced into the time sequence prediction model, the rational nonlinear Cauchy function approximation layer and the high-dimensional Cauchy function approximation layer are deployed in the Cauchy activation function, and then the Cauchy activation function is used to carry out approximation on the closure function used for determining the prediction result of the target object. The method facilitates the fitting of high curvature and long tail changes, effectively improves the prediction stability and dynamic consistency under the complex conditions of noise interference, data missing, long-time extrapolation and the like, and remarkably improves the prediction accuracy of time series data.

A function approximation unit includes a programmable function approximation circuit configured to directly compute a piecewise function according to a pre-stored coefficient; one or more dedicated function circuits configured to process at least one specific function among a reciprocal, a reciprocal square root, and a negative exponential function through a fixed hard-wired operation pipeline; and control logic configured to selectively activate either the programmable function approximation circuit or the dedicated function circuit according to the type of the nonlinear function to be processed.

The invention discloses a transform reasoning accelerator based on random calculation. The transform reasoning accelerator comprises a random bit stream generation unit, a random calculation unit and a random calculation unit, wherein the random bit stream generation unit is used for generating a random bit stream which is in probability correspondence with an input fixed-point numerical value according to the input fixed-point numerical value; the random multiplication and addition operation unit is used for executing bitwise multiplication operation on the random bit stream from the random bit stream generation unit and accumulating bitwise multiplication output through a parallel counter in a preset sampling period to generate a multiplication and addition result of probability estimation; the nonlinear function approximation unit is used for executing polynomial approximation on the probability accumulation result from the bit-level multiply-add operation unit so as to generate nonlinear activation function output required in Transform; and the random bit stream generation unit is connected with the random multiplication and addition operation unit and the nonlinear function approximation unit to form a complete reasoning accelerator data path. The invention provides a random calculation Transform reasoning accelerator based on a sobol sequence. The random calculation Transform reasoning accelerator has the advantages of being low in power consumption, small in area and high in robustness.

This invention discloses a power system forward scheduling method based on migration-type cost function approximation, comprising the following steps: S1. Constructing a power system forward scheduling model based on a cost function, with the minimization of power system scheduling cost as the objective function, and determining the constraints of each parameter of the state variables and decision variables in the cost function; S2. Based on the cost function approximation theory and considering the influence of random information, transforming the objective function, and obtaining an approximate cost function by introducing approximate parameters to estimate the influence of random information on the cost function; S3. Establishing a mapping relationship between the approximate parameters and random information, using a Long Short-Term Memory (LSTM) neural network for fitting, identifying the parameters of the LSTM, and obtaining the optimal parameters of the approximate parameters; S4. Inputting the optimal parameters of the approximate parameters into the power system forward scheduling model, the power system forward scheduling model outputs a real-time scheduling optimization scheme, and scheduling the power system according to the real-time scheduling optimization scheme. This improves the accuracy of forward scheduling.

A system and method of designing an integrated circuit for calculating an approximation of a target function over a predetermined interval may include employing an approximation algorithm, to calculate a first approximation function, which approximates the target function. Embodiments may construct an objective function based on the first approximation function. Based on the objective function, embodiments may calculate a first set of outcome coefficient values, which define an outcome approximation function, and generate, based on the outcome approximation function, an approximation schematic. The approximation schematic may represent an electrical approximation circuit, adapted to (i) receive an input value within the predetermined interval, and (ii) produce an estimation of the mathematical function at the input value, according to the outcome approximation function.

A time series anomaly detection method based on a function approximation network is applied to a computing device, the computing device is deployed with a monitoring system, the monitoring system is used to monitor a plurality of key performance indicators KPIs in the computing device, and the method comprises: obtaining a historical feature vector of any one KPI data in the plurality of key performance indicators KPIs at a current time t; training a prediction model based on a plurality of historical feature vectors at different times; the prediction model is a function approximation network model; predicting the feature vector at time t through the trained prediction model to determine the predicted feature vector; and determining whether any one KPI data is abnormal based on the predicted feature vector and the vector of the true value of any one KPI data at time t. The method can model the dependency relationship between indicators in a fine-grained manner and improve the accuracy of anomaly detection.