42results about How to "Speed up exploration" patented technology

The invention discloses a self-encoding model training method, and the method comprises the following steps: obtaining a detection result in a report, and performing pre-encoding of the detection result; adding noise to the detection result subjected to precoding; inputting the detection result added with the noise into a preset self-encoding model, wherein the self-encoding model comprises an encoder and a decoder; optimizing the self-encoding model by using an optimizer; when the loss function of the optimized self-encoding model is stable, removing the decoder in the self-encoding model tooutput the self-encoding model with the decoder removed as a trained feature expression model. The invention further discloses a self-encoding model training device and equipment and a computer storage medium. By adopting the embodiment of the invention, the trained feature expression model can improve the exploration of the report detection result space, and solves the problems of incomplete coverage, low efficiency and the like of manually constructed medical examination report detection result feature variables.

The invention discloses a microprocessor micro system structure parameter optimization method based on a Petri network. The method comprises the following steps that a template of a flow line model is built on the basis of the colored Petri network, an instruction sequence of a target application program is obtained, relevant information between the instructions and the function unit types is obtained, a colored Petri network model, running under the current parameter configuration, of the target program is generated, a Petri network simulation tool is used for simulation, a simulation report is generated, the colored Petri network model is used for generating a corresponding directed acyclic graph according to the simulation report, a key path of the directed acyclic graph and nodes passed by the key path are calculated, the release time of each entering edge of each node is calculated, the performance bottle neck or power consumption bottle neck for running the target application program by the microprocessor in the current micro system structure parameter configuration is analyzed, and if the optimization is required, the micro system structure parameters are regulated. The microprocessor micro system structure parameter optimization method has the advantages that the prediction reliability and the precision are high, the searching design space relating range is wide, the optimization algorithm complexity is lower, and the optimization is fast and efficient.

The invention discloses a network node selection method and system based on a whale optimization algorithm, and a storage medium. The method comprises the following steps: setting a binary coded population matrix, setting the maximum number of iterations and the initial number of iterations, and randomly initializing the population matrix; calculating the target function according to the node selection scheme corresponding to each whale individual in the population to obtain the optimal individual position and the optimal fitness function value in the population; calculating a current dynamicconvergence factor and a current dynamic weight according to the current number of iterations; determining a scheme for calculating the whale individual position according to the current dynamic convergence factor and the generated random number, and updating the current whale individual position and the number of iterations; if the number of iterations reaches the maximum number of iterations, returning to the optimal whale individual position, and determining sensor nodes participating in tracking according to the optimal whale individual position; otherwise return computation. According tothe invention, the tracking precision and the real-time performance in the target tracking process in the wireless sensor network can be improved.

The invention discloses an MU-MISO hybrid precoding design method based on multi-agent deep reinforcement learning, and the method is suitable for a downlink system in communication. According to themethod, a base station constructs a plurality of deep reinforcement learning agents used for calculating an analog precoding matrix, each agent comprises an action prediction network and an experiencepool with priority, and the agents share a centralized reward value prediction network and a centralized evaluation network to cooperatively explore an analog precoding strategy. The method comprisesthe following steps: enabling a base station to acquire channel state information of a plurality of users, inputting the user channel information into a constructed intelligent agent, and outputtinga corresponding analog precoding matrix; and calculating a digital precoding matrix containing the digital precoding vector of each user through zero-forcing precoding and a water injection algorithm.According to the method, the problems of high hybrid precoding design complexity and poor reachable rate performance in a large-scale MIMO system can be effectively solved, and the method has relatively high robustness to a channel environment.

A protein conformation space optimization method based on differential evolution local disturbance is disclosed. Under the framework of a differential evolution algorithm, information exchange betweenindividuals in a population is used to enhance the exploration ability of the algorithm; and simultaneously, the differential evolution algorithm is used to achieve the fine tuning of a loop area andincrease the diversity of a loop area structure so that the exploration of the loop area is further enhanced based on an existing structure and then the exploration efficiency and the prediction precision of an integral body are increased. By using the protein conformation space optimization method based on the differential evolution local disturbance, prediction precision is high.

The invention provides a device for synthesising a metal matrix powder material and a high-flux synthesis method thereof. The device comprises a base plate and a heater, wherein the base plate comprises a concentration gradient generation part and a temperature gradient generation part; the heater is arranged at the bottom of the temperature gradient generation part; the cross section of the temperature gradient generation part is in the shape of a right trapezoid; x liquid inlets are arranged in the outer end part of the concentration gradient generation part, and divergently extend inwards in a dendritic form; y liquid outlets are arranged in the inner end part of the concentration gradient generation part, and extend towards the temperature gradient generation part to form y liquid flow channels; and a plurality of reactors are arranged in each liquid flow channel. Compared with the prior art, the device and the high-flux synthesis method, which are provided by the invention, have the following beneficial effect: 1. series concentration gradients can be automatically generated in one experiment by virtue of the repeated splitting and merging of micro-fluids, thus a cumbersome process of manually preparing reaction solutions with different concentrations is greatly omitted.

The invention relates to an ant colony algorithm optimization method based on reverse learning, which is used for solving a traveling salesman problem. The improvement of the algorithm mainly comprises the following points: 1, after an initial path is solved, reversing the serial number of each city in the initial path, and constructing a reverse path; 2, sorting the initial paths and the reversepaths from small to large according to the lengths, and taking part of the paths to form a group of new paths; and 3, setting an iteration threshold value, and if the current iteration frequency doesnot reach the iteration threshold value, carrying out pheromone updating on a new group of paths; otherwise, performing pheromone updating on the initial path. The pheromone updating aspect of the basic ant colony algorithm is improved, in the early stage of iteration, reverse learning is introduced to construct a reverse path, and participates in pheromone updating, so that the search range of ants is expanded, the ants are prevented from falling into local extremum, and exploration and development of the understanding space are balanced.

The invention belongs to the technical field of natural language data processing, and discloses an unregistered word identification method and system using five-stroke character root deep learning. The method comprises the steps of converting a Chinese character into four English letters according to a five-stroke character root table; then inputting an embedded vector serving as an embedded vector of the model into an embedded vector corresponding to the words in a corpus to train a neural network model; and finally, enabling the model to output a most similar vocabulary vector in a previous corpus, and using the vocabulary vector as an important basis for identifying the unlogged vocabularies to better identify the unlogged vocabularies. According to the present invention, the Chinese character words with close radicals mostly have the same part-of-speech, and the five-stroke codes of the Chinese character words are similar, so that the neural network entity identification method based on the five-stroke roots is provided and can improve the performance of identifying the unlogged words through the neural network model. According to the present invention, the word vectors are used for representing the words based on deep learning, so that the sparse problem of the high-latitude vector space is solved, and the method is simpler and more effective.

The invention discloses a two-arm robot precise assembly method based on a six-dimensional force sensor. The method comprises the following steps that a robot controller reads all joint angles of a left mechanical arm and all joint angles of a right mechanical arm; Cartesian space poses of the left mechanical arm and the right mechanical arm are obtained; according to the Cartesian space poses of the left and right mechanical arms, the Cartesian space poses of pin hole workpieces at the tail ends of the left and right arms relative to the base coordinates of the robot are obtained through coordinate conversion; a left arm tail end pin hole planning path and a right arm tail end pin hole planning path are obtained through a fast potential energy exploration-based random tree method; wherein the planned path is short in movement distance, short in assembling operation time, low in mechanical arm energy consumption and capable of avoiding interference; and according to the left arm tail end pin hole and right arm tail end pin hole planning path, a control method based on a learning variable impedance control system is used for controlling the left mechanical arm and the right mechanical arm to be flexibly assembled. According to the method, the high-quality path can be planned, and the compliant assembly capacity is achieved.

The invention discloses a semantic-enhanced large-scale multi-element graph simplification visualization method, which comprises the following steps of: establishing a large-scale multi-element graph,and extracting a hierarchical structure of the large-scale multi-element graph; constructing a multi-scale community set according to the hierarchical structure of the large-scale multi-element graphby utilizing the attributes of the large-scale multi-element graph, the attributes of the large-scale multi-element graph including modularity and multi-dimensional attribute information entropy; constructing a multi-level force guiding layout for the multi-scale community set according to the hierarchical structure of the large-scale multivariate graph, and displaying the semantic expression ofthe communities through mapping; And using the community after mapping display to obtain a hierarchical view and an attribute mulberry-based view, and performing visual analysis on the large-scale multi-element view by using a multi-level force guide layout, the hierarchical view and the attribute mulberry-based view. According to the method, the visual expression of the large-scale multi-elementgraph can be effectively simplified, the association structure and semantic composition of the large-scale multi-element graph in different application fields can be rapidly analyzed, and the practicability is high.

The invention discloses an industrial robot-oriented reinforcement learning reward value calculation method. The method comprises the following steps of S1, carrying out initialization calculation on state parameters of an industrial robot; s2, calculating a pose reward value of an end execution mechanism of the industrial robot; s3, calculating a collision reward value of the industrial robot; s4, calculating an exploration reward value of the industrial robot; s5, calculating a target reward value; according to the method, the target nearby area and the non-target nearby area are divided, so that the tail end of the industrial robot can be quickly close to the target position in the early stage and can be adjusted to be in a proper posture while being close to the target position in the later stage, and the exploration process of the industrial robot is accelerated; various state information (positions, postures, collision and the like) of the industrial robot are comprehensively considered, and the problem that postures of a motion track finally planned by the industrial robot cannot meet actual production requirements is solved.

The invention relates to a reinforcement learning method and device based on a short-time access mechanism, and a storage medium. The method comprises the steps: configuring a state cache list which is used for storing state increment information obtained through the change of a current environment state of an intelligent agent under the condition that the intelligent agent meets a preset short-time access mechanism; inputting all actions of the intelligent agent at the next moment into the environment state transition probability model, and outputting a plurality of environment states of allactions corresponding to the next moment; comparing the plurality of environment states at the next moment with the state increment information in the state cache list, and determining an action corresponding to the environment state with the maximum difference in the plurality of environment states as a first alternative action executed by the intelligent agent at the next moment; and executing exploration operation for reinforcement learning according to the first alternative action. According to the invention, through the state cache list, repeated exploration of the explored environment state is avoided; through the environment state transition probability model, the exploration of the intelligent agent to the unknown state is strengthened and guided, and the learning efficiency is effectively improved.

The invention relates to a robot path exploration method based on double-agent competitive reinforcement learning, and the method comprises the following steps: S1, constructing a Markov decision model, and initializing agents and an experience pool; s2, recording a current state st of an agent Agent1, exploring k steps, and recording a current track sequence to an experience pool Buffer 1; s3, the intelligent agent Agent2 is placed at the state st, the intelligent agent Agent2 explores k steps, and a current track sequence is recorded to an experience pool Buffer 2; s4, taking the similarity between the exploration trajectories as an additional reward of the agent Agent1, and taking an opposite number as an additional reward of the agent Agent2; s5, updating strategies of the agents Agent1 and Agent2 when the number of data in the experience pool meets the requirement; s6, repeatedly executing the steps S2-S5 until the intelligent agent Agent1 reaches the target state or exceeds the set time tlimit; and S7, repeatedly executing the steps S1-S6 until the set training episode number is completed. Compared with the prior art, the method has the advantages that the intelligent agent can explore more effectively, the training speed is increased, the utilization efficiency of samples is improved, random noise can be effectively eliminated, and the robustness is higher.

The invention discloses a construction method of an oil storage device capable of achieving freezing through freezing pipes. The construction method comprises the specific steps that an appropriate geographic position is selected, a main body structure is constructed through reinforcing steel bars, and a fiber concrete space enclosing structure is formed from high-pressure jet grouting fiber cement; a working well is dug; freezing pipes are inserted to the periphery of the fiber concrete space enclosing structure; a refrigerant is introduced into the freezing pipes for freezing soil, and frozen oil with the high strength and good sealing performance is formed; deep-stratum soil is dug away, and an oil storage space is formed; the inner side of the fiber concrete space enclosing structure is paved with waterproof roll, and the upper soil body layer of the oil storage space reinforced through high pressure spouting; and hydraulic oil is injected through high pressure, and sealing is carried out. The invention further discloses the oil storage device capable of achieving freezing through the freezing pipes. The problems of excessive spending and explosion caused by storage in deep-layer soil in the prior art are solve, implementation is convenient, adaptability is good, safety and reliability are high, and no pollution is produced.

The invention discloses an ionized layer high-dimensional data feature selection method based on an improved BBA algorithm. The method comprises the following steps: acquiring ionized layer data; taking a dimension classification loss function as a target function; adopting an improved BBA algorithm to solve the target function, wherein the improved BBA algorithm comprises the following steps: after the individual speed is updated in a single dimension, mapping from a continuous space to a discrete space is conducted on the updated individual speed according to a time-varying V-shaped conversion function; and determining a target dimension after solving, and performing dimension reduction processing on the ionized layer data according to the target dimension to obtain ionized layer features corresponding to the target dimension. A random black hole model is introduced, a time-varying V-shaped conversion function is provided to improve a BBA algorithm, after ionized layer high-dimensional data is subjected to dimension reduction based on an improved discrete binary bat algorithm, a minimized feature subset is generated, the data error rate is reduced, the dimension classification precision is improved, and accurate ionized layer data features are selected.

The invention discloses an autonomous exploration type semantic map construction method and system, and the method comprises the steps: carrying out the autonomous exploration of an unknown environment based on a robot autonomous exploration algorithm of an improved rapid extension random tree, and carrying out the global front point detection and local front point detection through employing a global random tree and a local random tree in the exploration process; clustering the front edge points after the front edge points are obtained, selecting the front edge point with the maximum income as an optimal target point, and controlling the robot to arrive at the optimal target point along the optimal path; an image sequence of a current scene is collected in the moving process of the robot, the position of the robot is continuously updated, and when the optimal target point changes, the path is re-planned, and the robot is controlled to reach a new optimal target point; according to the method, the RGB image collected in the autonomous exploration process of the robot is subjected to semantic segmentation, and the RGB image and the corresponding depth map are combined to serve as input of the semantic map construction system, so that the autonomous semantic map construction task is completed, and the autonomy and intelligence of the robot are improved.