CN114492165A - Parameter optimization method and system based on genetic breeding method - Google Patents

Parameter optimization method and system based on genetic breeding method Download PDF

Info

Publication number
CN114492165A
CN114492165A CN202111604407.1A CN202111604407A CN114492165A CN 114492165 A CN114492165 A CN 114492165A CN 202111604407 A CN202111604407 A CN 202111604407A CN 114492165 A CN114492165 A CN 114492165A
Authority
CN
China
Prior art keywords
coordinate system
data
orthogonal coordinate
invariance
typical sequence
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202111604407.1A
Other languages
Chinese (zh)
Inventor
常松
盛建东
杜保军
孙家鹏
武红旗
陈冰
石书兵
康定明
马英杰
张红忠
程军回
张凯
石秀楠
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xinjiang Beaver Agriculture And Animal Husbandry Software Co ltd
Original Assignee
Xinjiang Beaver Agriculture And Animal Husbandry Software Co ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Xinjiang Beaver Agriculture And Animal Husbandry Software Co ltd filed Critical Xinjiang Beaver Agriculture And Animal Husbandry Software Co ltd
Priority to CN202111604407.1A priority Critical patent/CN114492165A/en
Publication of CN114492165A publication Critical patent/CN114492165A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00Computer-aided design [CAD]
    • G06F30/20Design optimisation, verification or simulation
    • G06F30/27Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10Complex mathematical operations
    • G06F17/18Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/25Fusion techniques

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Mathematical Optimization (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Pure & Applied Mathematics (AREA)
  • Evolutionary Biology (AREA)
  • Computational Mathematics (AREA)
  • Mathematical Analysis (AREA)
  • Geometry (AREA)
  • Operations Research (AREA)
  • Probability & Statistics with Applications (AREA)
  • Computer Hardware Design (AREA)
  • Algebra (AREA)
  • Medical Informatics (AREA)
  • Databases & Information Systems (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a method and a system for optimizing parameters based on a genetic breeding method, which relate to the technical field of artificial intelligence systems and comprise the following steps: s1: acquiring measurement data, preprocessing the measurement data to obtain typical sequence data meeting detailed balance conditions, and executing step S2; s2: establishing an orthogonal coordinate system based on the typical sequence data meeting the detailed balance condition, judging whether the orthogonal coordinate system has invariance, if not, executing step S3, and if so, executing step S5; s3: calculating the projection of the measurement data to the orthogonal coordinate system, judging whether the orthogonal coordinate system has invariance, if not, executing step S4, and if so, executing step S5; s4: calculating the correlation of the measurement data on the orthogonal coordinate system, judging whether the correlation has invariance, if not, executing step S2, and if so, executing step S5; s5: and outputting typical sequence data, independently optimizing and modeling, and outputting a fitting result meeting an optimization condition.

Description

Parameter optimization method and system based on genetic breeding method
Technical Field
The invention relates to the technical field of artificial intelligence systems, in particular to a method and a system for optimizing parameters based on a genetic breeding method.
Background
In the data processing process of medical health, industry, agriculture and animal husbandry, the number of samples is often insufficient, or the samples are not typical and representative, that is, the obtained data or model is not optimal, for example, census data of a certain region may not represent the census condition of the urban area to which the region belongs, so that the measures carried out in the region cannot be applied to the whole urban area. This is an over-fitting problem, or the data is fitted to a model that looks perfect, but cannot be really and completely applied to the actual process, on one hand, because the data does not have the representativeness of other areas, on the other hand, the fitted model parameters are too much to reflect the actually required data distribution result. Therefore, the invention provides a method and a system for optimizing parameters based on a genetic breeding method, which are used for overcoming the problems.
Disclosure of Invention
Aiming at the defects in the prior art, the invention provides a method and a system for optimizing parameters based on a genetic breeding method.
The parameter optimization method based on the genetic breeding method comprises the following steps:
s1: acquiring measurement data, preprocessing the measurement data to obtain typical sequence data meeting detailed balance conditions, and executing step S2;
s2: establishing an orthogonal coordinate system based on the typical sequence data meeting the detailed balance condition, judging whether the orthogonal coordinate system has invariance, if not, executing step S3, and if so, executing step S5;
s3: calculating the projection of the measurement data to the orthogonal coordinate system, judging whether the orthogonal coordinate system has invariance, if not, executing step S4, and if so, executing step S5;
s4: calculating the correlation of the measurement data on the orthogonal coordinate system, judging whether the correlation has invariance, if not, executing step S2, and if so, executing step S5;
s5: and outputting typical sequence data, independently optimizing and modeling, and outputting a fitting result meeting an optimization condition.
Preferably, the preprocessing of the measurement data in step S1 includes the following steps:
acquiring measurement data;
carrying out localized segmentation on the measurement data through a data segmentation algorithm to obtain a segmentation matrix;
and combining the segmentation matrix and the matrix of the measurement data, and performing multiplication calculation to obtain typical sequence data meeting the detailed balance condition.
Preferably, the measurement data may be acquired by a bridge instrument.
Preferably, the step of determining the invariance of the orthogonal coordinate system in step S2 includes the following steps:
establishing an orthogonal coordinate system based on the typical sequence data satisfying the detailed balance condition;
selecting typical sequence data in an orthogonal coordinate system to carry out data coefficient decomposition to obtain a decomposition coefficient and a sequencing result of the typical sequence data;
invariance of the orthogonal coordinate system is evaluated based on the degree of overlap of the decomposition coefficient of typical sequence data and the sorting result.
Preferably, in the independent optimization modeling in step S5, the calculation process of the model parameters is as follows:
s51: obtaining male parent samples and female parent samples of typical sequence data;
s52: calculating initial noise amplitudes of the male parent sample and the female parent sample, and updating the male parent sample and the female parent sample based on the initial noise amplitudes;
s53: fusing the updated male parent sample and the female parent sample to generate a filial generation sample;
s54: calculating the noise amplitude of the filial generation sample, judging whether the noise amplitude of the filial generation sample meets the set stability, and if not, returning to the step S52; if yes, go to step S55;
s55: and taking the parameters of the filial generation samples as modeling parameters and outputting the parameters.
The parameter optimization system based on the genetic breeding method comprises the following steps:
an acquisition unit configured to acquire measurement data;
the preprocessing unit is used for preprocessing the measurement data to obtain typical sequence data meeting the detailed balance condition;
the evaluation unit is used for screening out typical sequence data meeting invariance conditions;
and the independent optimization unit is used for performing optimization fitting by using the typical sequence data meeting the invariance condition and the local model and outputting a fitting result of the optimization condition.
Preferably, the evaluation unit includes:
the device comprises a first evaluation unit, a second evaluation unit and a third evaluation unit, wherein the first evaluation unit is used for selecting an orthogonal coordinate system from typical sequence data and judging whether the orthogonal coordinate system meets the invariance of the coordinate system;
the second evaluation unit is used for calculating the projection of the measurement data to the orthogonal coordinate system and judging whether the orthogonal coordinate system has invariance or not;
a third evaluation unit for calculating the correlation of the measured data on the orthogonal coordinate system and judging whether the correlation has invariance
The invention has the beneficial effects that:
according to the method, through the optimization of the local area set and the local area model parameters, the problem of insufficient data representativeness is solved, the model parameters are simplified and the maximum influence is maximized through the fixed points or invariance brought by the symmetry of the model, and the problem of excessive model redundant parameters is avoided.
Drawings
In order to more clearly illustrate the detailed description of the invention or the technical solutions in the prior art, the drawings that are needed in the detailed description of the invention or the prior art will be briefly described below. Throughout the drawings, like elements or portions are generally identified by like reference numerals. In the drawings, elements or portions are not necessarily drawn to scale.
FIG. 1 is a schematic diagram of a bridge instrument measurement provided by the present invention;
FIG. 2 is a schematic diagram of a measurement graph result provided by the present invention;
FIG. 3 is a schematic diagram of the fitting results provided by the present invention;
fig. 4 is a schematic diagram of a capacitive sensor fitting process provided by the present invention.
Detailed Description
Embodiments of the present invention will be described in detail below with reference to the accompanying drawings. The following examples are only for illustrating the technical solutions of the present invention more clearly, and therefore are only examples, and the protection scope of the present invention is not limited thereby.
It is to be noted that, unless otherwise specified, technical or scientific terms used herein shall have the ordinary meaning as understood by those skilled in the art to which the invention pertains.
The main reasons for the occurrence of overfitting mentioned in the present invention may be the following four points:
(1) the data is noisy;
(2) limited training data;
(3) the model is excessively trained, so that the model is very complex;
(4) the influence of all factors of the training model is uneven, so that the model cannot reflect the average condition.
Generally, the conventional solutions to overfitting are as follows:
more representative data were acquired: acquiring more data from a data source;
using the appropriate model: the number of layers of the network, the number of neurons and the like can be reduced to limit the fitting capacity of the network;
dropout;
regularization, wherein the limit weight is increased during training;
limiting the training time; passing an evaluation test;
randomization: input and weight randomization (gaussian initialization);
data cleaning: correcting or deleting the error label;
various models are combined.
However, these approaches have the disadvantage of requiring an ever-increasing knowledge of the models and data, which is often not readily available, or that the local data is not fully available until the learning results are known, but is progressively available, so that progressive acquisition of the local data is often a difficult task for a large-scale learning system. Another problem is that the output result is not a smooth random distribution, and the mean value or the value obtained by simple calculation and conversion cannot be converged to a predicted expected result, which increases the difficulty of smooth convergence of data. Finally, the non-linearity increases the computational complexity, and some factors cannot reach the maximum influence (i.e., the maximum yield or growth rate of each factor in the computation), and are easily ignored in the computation process, resulting in a new overfitting phenomenon.
The over-fit phenomenon is that the dimensions that the predetermined model can express exceed the dimensions of the currently available data. A simple method for solving the overfitting is to perform dimension reduction on the model, optimize the model parameters and generate a proper low-dimensional model, but if a good global processing model is constructed, the adaptability of the model is finally lost due to the excessive dimension reduction.
In order to solve the problem that the data set is too small or not representative, the invention provides a genetic breeding method, namely the Liu's algorithm mentioned below.
The Liu's algorithm is proposed by Liu academy, and firstly, the individuals (brother, sister, parents and grandparents) with blood affinity are mated and bred by using a genetic breeding method to obtain an optimized variety, so that the range of available data can be increased. Compared with genetic breeding, the traditional breeding method is to continuously expand the sample range, increase representative samples, realize hybrid breeding and obtain hybrid advantages (the hybrid advantages are the hybridization among samples with relatively far relation, thereby obtaining filial generations with relatively good characters).
Compared with the Liu's algorithm, the classical dimensionality reduction method comprises the following steps: LLE (local linear embedding) and ISOMAP (equal metric mapping) are results of simulating high dimensions by fitting in a large range by using all local dimension reduction data, and can be regarded as a global optimal fitting process. The problem is that on one hand, data may not be representative, and the established model can easily ignore data in different areas; on the other hand, the amount of calculation is large, and the calculation error increases exponentially as the number of local regions increases.
In contrast, "Liu's algorithm" includes local area data algorithms and local area model algorithms. Local area data algorithms use either existing neighboring sample (sibling parents) data or local area sample data generated with a local data model for processing (local dimensionality reduction) to obtain representative data. Specifically, the measured local data sequence is continuously fused with itself (or sample data with the most recent relationship) to generate a local reduced-dimension or simplified sequence, and then the local reduced-dimension or simplified sequence is used for a fitting process (including obtaining a reasonable estimation without deviation on distribution through an adaptive process). The local area model algorithm uses stepwise regression, and independent Model Parameter Optimization (MPO) is performed by using a corresponding local area data set, or an optimal data processing model is obtained by using a local area data evaluation model system obtained by other regression methods.
The Liu's algorithm aims at outputting fixed, stable and reliable results, and the required results are calculated by a similar parameter optimization fitting method. This can be seen in the machine learning models of neural networks, which can always perform well with thousands to tens of thousands of parameters, despite thousands of billions of image data. In addition, many quantum problems all exhibit exponential growth as systems increase, but can often be parameterized efficiently within polynomial complexity. The reason behind this is that the hamiltonian of a physical system has locality and symmetry. Symmetry can also be considered as an intrinsic factor. Vignette (e.p. wigner) in 1927 demonstrated invariance in atom splitting, one of the consequences of symmetry. Therefore, the locality of the physical system can be regarded as an extension of symmetry, that is, the symmetry makes the local parameter solution global.
According to the information theory, the local data sequence is complex enough and long enough, when the length N is reached, according to AEP (gradual equal partitionability principle), the entropy of the local data sequence is H (S), the maximum information entropy of the whole data sequence is reflected, and the result represents the global result. This can also be seen as a sufficiently large matrix, which must not be full rank, plus the fitting result, to be full rank, but to give a unique result.
The "Liu's algorithm" solution relies on the addition of local data sets, the most efficient approach being to segment and re-segment the data in a random process, with a continuous recursion of the typical sequence. A typical sequence is a sequence that satisfies randomness, averaging, and convergence, generated from producing actual local data, whose local entropy is equal to the global entropy. The "Liu's algorithm" may be used to generate data and parameter adjustments. Since the segmentation and fitting process may bring multiple results, this may be done by stepwise regression with "Liu's algorithm", parametric optimization or other multivariate fitting methods to select a suitable model from the model system. The stepwise regression parameter optimization fitting method is to change the method of generating the typical sequence by fusion in a recursive method by means of several optimized results (possibly from different fusions of several local data and the typical sequence), and then compare the calculated results with respect to the invariance fitting. This method can be applied to the estimation of mean and variance.
The relationship referred to above is a phylogenetically displayed relationship of the biological group, which is traditionally based on morphological or anatomical similarities and differences of animals, and can be represented by two sequences { x }1,x2,x3,.. } and y1,y2,y3,.. } the Euclidean distance S between:
Figure BDA0003433200930000071
mathematically this may consider the relationship to be a geographical coordinate expression, with siblings regarding the relationship between any geographical location on the ground to its respective local point, such that at each local location there is a geodetic coordinate system that satisfies in south, east and north. The building or measurement according to the geodetic coordinate system can meet the coordinate of the spherical surface of the earth and the horizontal and vertical geodetic coordinate.
According to the analysis: a typical sequence of sufficient length results in a constant orthogonal coordinate system. From the local data set, an output covering the global is obtained according to the invariant orthogonal system. It is also possible to introduce a constant projective transformation, for example, rotating an image does not change the distance of a point from a curve. From the local data set, an output covering the global can also be obtained, back-calculated according to the invariant projective transformation process. The representative sequence may also bring about invariant correlations, such as the transformation of any rigid object. From the local data set, an output covering the global can also be obtained by back-computing according to the invariant correlation. Invariance of the correlation can be expressed in terms of conservation of momentum and kinetic energy corresponding to movement of the rigid object. The local data set is the corresponding assumption of locality, which is the principle of locality. In physics, the Principle of localization (also called localization Principle or regionality Principle) means that a specific object can only be influenced by the surrounding forces. The physical theory, including the principle of localization, is called a localization theory. According to classical physics's view of field theory, the action of one point, affecting another, in the intervening space, e.g., a field, becomes an intermediary in motion. To affect another point, a wave or particle must first pass through the space between the two points before it can be affected. According to the localization principle, the current state of a certain local area can be calculated by using a model of the local area from the past state of the local area and an event occurring in the local area. That is, due to localization, local data and local models and global data satisfy the invariance of global correlation relationships as global models, which in this example of physics is the law of conservation of kinetic energy and momentum. This invariance can bring a sufficient number of correlation equations, that is, the system input and output satisfy a one-to-one relationship, i.e., an ergodic relationship, under the invariance relationship. Thus, the self-mapping of a → a, or a set, brings a sufficiently long self-mapped data sequence, satisfying the principle of progressive equal partitionability, and the distribution of outputs that generate independent equal probabilities is a global distribution relationship for each input X and sufficiently long (the number of self-mappings is sufficiently large) in the local area. These relationships can be checked and confirmed from data changes in the system.
And, in the analysis of the algorithmic model: the model used for fitting starts with a set of learned data sequences that do not meet the predicted conditions due to insufficient data representativeness or too complex a model, i.e., some factors that do not reflect the maximum influence (e.g., previous overfitting plots reflect that the boundary conditions are not consistent, or are asymmetric, i.e., that
Figure BDA0003433200930000081
Where G ═ F + λ H, the lagrange term), such that the maximum influence of x is satisfied, the maximum influence of y, i.e., y, cannot be satisfied
Figure BDA0003433200930000082
). The condition of the Liu's algorithm above is that the data is not sufficiently simplified and does not reach a sufficiently low dimensionality. The solution is to generate a stable, average, convergent canonical sequence of contiguous data (e.g., in two-dimensional euclidean distance) that is continuously reduced by being continuously accepted or rejected with the minimum row-column position distance. For example, the sequence to be fitted is 0.12, 0.27, 0.56, 1.23.. so that it is combined with random noise, such as 0.1, 0.7, 0.3, 0.6, 0.3, 0.6.. the random sequence obtained by continuous combination is a simplest non-redundant sequence, such as 0.12, 0.22, 0.43, 0.89.. when N is sufficiently long, the number of occurrences of the sequence is 2NH(X)Probability of 2-NH(X)Average self-information content of the sequence
Figure BDA0003433200930000083
Is equal to
Figure BDA0003433200930000084
Or h (x). In addition, according to the first Shannon theorem, in XiWhen each character is independently and equally distributed, the simplest length L of the sequence is equal to log (p (X)i))=log(2-NH(X)) Nh (x), then
Figure BDA0003433200930000085
Approaches the source entropy H (x), or the limit entropy HTherefore, as long as the simplified sequence is long enough, the occurrence of the sequence character satisfies randomness, and the sequence satisfies the conditions of a typical sequence, namely nh (x) ═ h (x), in accordance with the definition of the amount of self-information. The sum of the probabilities of all representative sequences equals 2NH(X)(number) times 2-NH(X)(probability of each step) is equal to 1. The result can be said to be a balanced system, which is equivalent to the combination of an encryption process and a simplification process, and directly reflects the result. Because of the independence of the results to be achieved, the fusion problem can be viewed as a decryption process, which is aimed at being separated from other data. Since the sum of the probabilities is 1, it can also beSeen as a simplified problem, so that the most probable situation can be achieved (see later proof). Both of these constitute a traversal segmentation process, reflecting the limited attributes.
Assuming an independent equal probability distribution coin-throwing sequence of 0110101010, where L is 10 and N is 10, the simplest sequence is:
Figure BDA0003433200930000091
Figure BDA0003433200930000092
the simplest sequence is a typical sequence. Since the entropy of the simplest sequence is equal to the overall entropy of the system, it can be extended, compressed, concatenated, or pruned in any way without loss of system information. Meanwhile, the continuous use of the method is also equal to a continuous sub-optimal sequencing process, and finally a good sequencing result can be always obtained to generate a stable and convergent fitting result.
To speed up this process, the sequence needs to be optimized in addition to the simplified sequence. The method of optimizing the sequence is to perform the reverse data exchange on the sequence continuously, i.e. the reverse process of the random fusion of the previous data-the process of data generation is reversed, and a new sequence which looks all the same but has completely different subsequence immediately is obtained. When a sufficient number is exchanged, the average self-information amount I (X) with characterj) Must be reduced because the mutual information I (Y, X) of each step is reduced. The formula is calculated as I (Y, X) ═ H (X) -H (X | Y), or H (X) ═ I (Y, X) + H (X | Y), where H (X | Y) is zero because its conditional probability is zero; while according to the definition of I (Y, X), it represents the amount of information that can be considered as a random variable Y contained in the random variable X, or the uncertainty of the reduction of the random variable X due to the known random variable Y, as a process of probability superposition, thisThe uncertainty will decrease. According to the shannon's first theorem and the progressive equal-partitionability theorem, the condition h (X) ═ I (X) is reachedi)=-log(p(Xi))=-log(2-NH(X))=NH(X)The typical sequence is obtained.
Reducing the probability of occurrence with increasing process length (forward and reverse operation) (2)-NH(S)) The method (1) monotonically increases the self-information amount (the self-information amount is defined as the logarithm of the negative probability, namely:
-log(P(Xi))=-log(2-NH(S))=NH(S)
decreasing with increasing probability), according to the theory that the monotone increasing function of the real variable function brings fixed points, the system has nash equalization, that is, the random sequence has a dynamic process of maintaining equalization to achieve the maximization of self information quantity, or a process of becoming a typical sequence.
It can also be demonstrated by the hasani conversion in the incomplete information game, which is to generate the average self-information quantity of N length or single character by using the length and probability hypothesis of the typical sequence
Figure BDA0003433200930000101
The new sample set approaching h (x) brings an imperfect game, which is an independent equal probability random sampling process. Because a random relation of random equiprobability exists everywhere in the new sample sequence set, the immobile point P (X) is obtainedk)=Xk,XkThat is, the probability of each character of a sequence of random equiprobable everywhere of length N, P (X)k) That is, the conditional probability of the subsequent character is an independent equal probability distribution, and then Nash equilibrium can be obtained.
When the output code is a random sequence, the expectation value of the sum of the self-information amount of the input sequence according to the theorem of large numbers
Figure BDA0003433200930000102
When N is increased, the entropy approaches to the information entropy H (S), namely
Figure BDA0003433200930000103
Wherein SiIs the input sequence satisfying the random distribution, and S is the probability distribution of the random sequence defined by this sequence. The entropy of each information source is the sum of the probabilities of the self-information volumes, or its expectation, and the resulting result is the output since the sum of their probabilities and the probability of the entropy of the information sources are equivalent for sufficiently long sequences. The system calculates the probability of the entropy, and when the entropy converges, the probability is converged together, and the obtained result is the final optimized result.
Obtained by the first law of Shannon, a discrete memoryless stable information source with entropy of H (X) is coded, if a sequence consisting of N information source symbols is coded, a code word consists of L binary codes and is used for any epsilon>0,δ<0, as long as there is
Figure BDA0003433200930000104
Then when N is large enough, distortion-free encoding can be achieved. When the codes are randomly distributed, the equal sign is established. Therefore, when the output result is randomly distributed, there is a code composed of the output symbol code sequence as the typical sequence, that is, NH (X) is equal to L, L is the code length of the code, and thus the condition I (X) of the typical sequence is satisfiedi) (ii) a probability of 2 for either NH (X), or p-NH(x)And then:
I(Xi)=-log(p)=-log(2-NH(x))=L
this simplified optimized random sequence must be a typical sequence. When the length of the sequence is N and continuously increases, the result calculated by the method necessarily reflects the results of all typical sequences. According to the definition of a typical sequence, the combination is the simplest coding combination, as long as N is continuously increased, the total probability is 1, the result of data processing is the probability 1, and a uniform result is obtained.
As mentioned above, this is achieved by first generating a sequence (possibly over-fit) by machine learning, starting a new sequence when the average information content of the sequence reaches or exceeds the entropy of the candidate data set, and then calculating its entropy and coding length, specifically by using a methodThe continuous simplified process forms a coding sequence, which is equivalent to coding the data sequence to form a chain with continuous length and short length until reaching a stable state. These sequences are then exchanged continuously to produce an optimized coding result. Rearranging these overfitting sequences by codes, according to the first law of Shannon, the optimal randomly distributed sequences are obtained, the average code length of which satisfies that L/N is more than or equal to H(s), and L/N<H(s) +1/N, that is to say it is a typical sequence which can be used to fit the desired result (supervised learning) or as a proposed distribution sequence to continue to produce the optimal sequence until the result stabilizes, which satisfies pij=F(pij) If the fixed point appears, the required result can be found by searching or re-fitting.
The data segmentation according to the invention can be implemented by the following algorithm: fourier transform method, Tamadder algorithm, Leeberg algorithm, machine learning method, NlogN algorithm, MC2Algorithm, image region segmentation algorithm, quantum normalization algorithm, multiplication-by-multiplication algorithm and recursion method.
Two-dimensional symmetric or localized segmentation matrices are performed using the above-described segmentation method (localization is known to be a necessary consequence of symmetry):
and continuously multiplying the segmented matrix obtained by the method with the matrix of the original data to obtain a typical sequence meeting the randomness, the stationarity and the convergence, and establishing an orthogonal coordinate system.
Or by using a quantum normalization method,
Figure BDA0003433200930000111
the segmentation data is converted to a value between (-1, 1) so that a threshold value can be used for segmentation.
Selecting typical sequences representing an orthogonal coordinate system, performing a data coefficient decomposition process, and evaluating invariance of the current coordinate system by using similarity of decomposition coefficients of the data to be detected on the typical sequences and sequencing results of the data to be detected, such as Nash coefficients or overlapping degrees (IoU).
And checking the invariance of orthogonal coordinate values formed by the typical sequences, and selecting an optimal coordinate system from the decomposition coefficients of the data to be detected on each typical sequence and the similarity of the sequencing result of the data to be detected.
Continuously adjusting segmentation, continuously adjusting a coordinate system, continuously iterating, simulating an actual genetic breeding process, and selecting a most representative typical sequence.
The present invention lists the following four algorithms for solving the output model parameters.
Basic fusion algorithm:
firstly, obtaining the projection of original data in an orthogonal coordinate system;
generating a male parent sample and a female parent sample pair with similar relation through projection on an orthogonal coordinate system, and selecting an optimal projection by using the similarity of the projection of the male parent and the projection of the female parent;
checking the invariance of the projections of the male parent and the female parent, selecting the male parent and the female parent to be fused from the classification result of the projections of the male parent and the female parent as classification seeds and the projections of other data to be detected;
checking projections of the male parent and the female parent as classification seeds, and selecting the best projection of the male parent and the female parent from the similarity by using invariance of classification results of projections from a K-means algorithm to other data to be detected;
fifthly, fusing the male parent sample and the female parent sample with similar relatives to generate filial generations of the relatives;
step six, the step is enough to withdraw, or withdraw after the output is stable, otherwise carry out the second step repeatedly;
and seventhly, outputting.
The self-adaptive fusion algorithm:
1. simplifying data to obtain a typical sequence meeting randomness, stationarity and convergence;
2. establishing a male parent sample and a female parent sample of a typical sequence;
3. calculating the amplitude of the noise of the starting sample;
4. updating the amplitude of the noise of the sample;
5. fusing the male parent sample and the female parent sample to generate filial generations;
6. preferentially selecting the male parent sample and the female parent sample as filial generations which are close to each other;
7. calculating the output amplitude and storing the amplitude as the amplitude of new noise;
8. the step number is enough and quit, or quit after the output is stable, otherwise, the third step is repeatedly executed;
9. and (6) outputting.
And (3) coding fusion algorithm:
the method comprises the following steps of firstly, obtaining data, an orthogonal coordinate system and a projection on the orthogonal coordinate system;
secondly, encoding the data, the orthogonal coordinate system and the projection on the orthogonal coordinate system;
enumerating each code by using a diagonal method;
establishing a correlation relationship among various codes through an orthogonal coordinate system, a projection or a functional relationship, also comprising a collection of the correlation relationships represented by the codes, and according to the segmentation principles such as the progressive and the like, the codes are long enough to express the information and the correlation relationship of various objects;
step five, finding the male parent sample and the female parent sample which are closest in code through the enumerated code correlation relationship, and selecting the optimal parent sample by using the code similarity;
fusing the male parent sample and the female parent sample with the most similar codes to generate a genetic filial generation;
step seven, the step number is enough and withdraws, or withdraw after the output is stable, otherwise carry out the fourth step and recode repeatedly;
and step eight, outputting.
The model parameter optimization fitting method comprises the following steps:
a, setting the fitting range and the fitting precision of each parameter value of a male parent and a female parent of a model, wherein the fitting range and the fitting precision generally comprise two parameters of the female parent plus noise coefficient phi (0.0-1.0) and male parent MCMC coefficient omega (0.0-1.0);
b, gradually refining the selection of two parameters phi and omega;
c, enumerating parameter values phi and omega by using a diagonal method;
d, continuously adjusting the generation and fusion processes of the male parent, the female parent and the typical sequence according to the enumerated parameter values phi and omega;
e, calculating the amplitude of the noise of the starting sample;
f, updating the amplitude of the noise of the sample;
g, calculating the output amplitude and storing the amplitude as the amplitude of new noise;
h, randomly selecting a male parent sample group and a female parent sample group;
fusing the male parent sample group and the female parent sample group to generate hybrid offspring;
j, circularly selecting the typical sequence N, continuing segmenting, otherwise updating the fitting adjustment typical sequence N according to the output result, and executing the step B;
k, recording the relation between the parameter value and the result;
l, the step number is enough to exit, or exit after the output is stable, otherwise, the step B is repeatedly executed;
m, taking a small-change constant parameter, and clearing a large-change parameter;
and N, outputting the model parameters.
It is also possible to estimate the noise parameters (such as mean and variance) using AMCMC (adaptive MCMC) samples, and then to adjust this sample sequence by parameters until randomness, smoothness and convergence are reached. The sequence can be used for partitioning a data sequence and establishing local data samples, and can also be used for establishing a data local model. The establishment of the data local area model is realized independently by continuously debugging and optimizing in the continuous adjustment of the self-adaptive noise parameters.
Local features can be identified by a manual assistance or convolution method, because the features meet extreme conditions and have approximate invariance, the result is fed back to the input and is equivalent to the ergodic Markov chain sampling of stable distribution, namely the detailed balance condition is met, the symmetry is brought, and the input reflects the prior probability of the features. The local model and the global model can be independently optimized, the expected value and the probability distribution variance of typical output are estimated according to the orthogonal coordinate, the orthogonal coordinate projection and the correlation relationship, and the local probability distribution result and the global probability distribution result are obtained through fitting.
The local sequences may also be listed by enumeration or recursive enumeration. And adding each local feature obtained in the previous step as a background, wherein each local feature has a fixed point, and as long as the sequence generated by each local feature is long enough, a dimension-reduced sequence is generated through the random background generated by each feature sequence, and can be used for independently fitting each local or global result.
The values of the minute capacitance range are measured by a bridge instrument as shown in fig. 1. The height of the cylinder after tilting reflects the slight capacitance change. The value reflected by the precision bridge is a range and oscillations within that range.
As shown in fig. 2 and 3, the measurement value is continuously oscillating, and the average value does not reflect the change in capacitance. A typical sequence is obtained by using a measurement result, and the number and the range of optimized parameters are obtained by using a randomly generated male parent and female parent to simply simulate the charge motion rule of an analog circuit, or integrating an information source and a measurement process.
Finally, a well-fitted equation is obtained which fits the height of the cylinder with the values of the capacitance parameters calculated from the measurements to obtain a representative sequence fusion estimate.
As shown in fig. 4, the capacitive sensor fitting algorithm implements the process:
obtaining instrument data;
obtaining segmented typical sequence data by using quantum normalization;
judging whether the careful balance condition is met, if so, continuing, otherwise, returning to the previous step;
selecting an orthogonal coordinate system from the typical sequence, and judging the invariance of the coordinate system;
calculating the projection of the measurement data to an orthogonal coordinate system, and judging the invariance of the projection;
calculating the correlation among the data, and judging the invariance of the correlation;
if not, repeatedly returning to the second step;
fitting data or establishing a machine learning model;
independently optimizing each model, judging whether the randomness, the stationarity and the convergence of input and output are met, and repeating the previous step if the randomness, the stationarity and the convergence are not met;
and outputting the result.
Finally, it should be noted that: the above embodiments are only used to illustrate the technical solution of the present invention, and not to limit the same; while the invention has been described in detail and with reference to the foregoing embodiments, it will be understood by those skilled in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some or all of the technical features may be equivalently replaced; such modifications and substitutions do not depart from the spirit and scope of the present invention, and they should be construed as being included in the following claims and description.

Claims (7)

1. The parameter optimization method based on the genetic breeding method is characterized by comprising the following steps:
s1: acquiring measurement data, preprocessing the measurement data to obtain typical sequence data meeting detailed balance conditions, and executing step S2;
s2: establishing an orthogonal coordinate system based on the typical sequence data meeting the detailed balance condition, judging whether the orthogonal coordinate system has invariance, if not, executing step S3, and if so, executing step S5;
s3: calculating the projection of the measurement data to the orthogonal coordinate system, judging whether the orthogonal coordinate system has invariance, if not, executing step S4, and if so, executing step S5;
s4: calculating the correlation of the measurement data on the orthogonal coordinate system, judging whether the correlation has invariance, if not, executing step S2, and if so, executing step S5;
s5: and outputting typical sequence data, independently optimizing and modeling, and outputting a fitting result meeting an optimization condition.
2. The genetic breeding method-based parameter optimization method of claim 1, wherein the step of preprocessing the measurement data in step S1 comprises the steps of:
acquiring measurement data;
carrying out localized segmentation on the measurement data through a data segmentation algorithm to obtain a segmentation matrix;
and combining the segmentation matrix and the matrix of the measurement data, and performing multiplication calculation to obtain typical sequence data meeting the detailed balance condition.
3. The genetic breeding method-based parameter optimization method according to any one of claims 1-2, wherein the measurement data can be obtained by a bridge instrument.
4. The method for optimizing parameters based on genetic breeding method as claimed in claim 1, wherein the step of determining the invariance of the orthogonal coordinate system in step S2 comprises the steps of:
establishing an orthogonal coordinate system based on the typical sequence data satisfying the detailed balance condition;
selecting typical sequence data in an orthogonal coordinate system to carry out data coefficient decomposition to obtain a decomposition coefficient and a sequencing result of the typical sequence data;
invariance of the orthogonal coordinate system is evaluated based on the degree of overlap of the decomposition coefficient of typical sequence data and the sorting result.
5. The genetic breeding method-based parameter optimization method of claim 1, wherein the independent optimization modeling in step S5 is performed by the following model parameter calculation process:
s51: obtaining male parent samples and female parent samples of typical sequence data;
s52: calculating initial noise amplitudes of the male parent sample and the female parent sample, and updating the male parent sample and the female parent sample based on the initial noise amplitudes;
s53: fusing the updated male parent sample and the female parent sample to generate a filial generation sample;
s54: calculating the noise amplitude of the filial generation sample, judging whether the noise amplitude of the filial generation sample meets the set stability, and if not, returning to the step S52; if yes, go to step S55;
s55: and taking the parameters of the filial generation samples as modeling parameters and outputting the parameters.
6. The parameter optimization system based on the genetic breeding method is characterized by comprising the following steps:
an acquisition unit configured to acquire measurement data;
the preprocessing unit is used for preprocessing the measurement data to obtain typical sequence data meeting the detailed balance condition;
the evaluation unit is used for screening out typical sequence data meeting invariance conditions;
and the independent optimization unit is used for performing optimization fitting by using the typical sequence data meeting the invariance condition and the local model and outputting a fitting result of the optimization condition.
7. The genetic breeding method-based parameter optimization system according to claim 6, wherein the evaluation unit comprises:
the first evaluation unit is used for selecting an orthogonal coordinate system from the typical sequence data and judging whether the orthogonal coordinate system meets the invariance of the coordinate system;
the second evaluation unit is used for calculating the projection of the measurement data to the orthogonal coordinate system and judging whether the orthogonal coordinate system has invariance or not;
and the third evaluation unit is used for calculating the correlation of the measurement data on the orthogonal coordinate system and judging whether the correlation has invariance.
CN202111604407.1A 2021-12-24 2021-12-24 Parameter optimization method and system based on genetic breeding method Pending CN114492165A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111604407.1A CN114492165A (en) 2021-12-24 2021-12-24 Parameter optimization method and system based on genetic breeding method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111604407.1A CN114492165A (en) 2021-12-24 2021-12-24 Parameter optimization method and system based on genetic breeding method

Publications (1)

Publication Number Publication Date
CN114492165A true CN114492165A (en) 2022-05-13

Family

ID=81495538

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111604407.1A Pending CN114492165A (en) 2021-12-24 2021-12-24 Parameter optimization method and system based on genetic breeding method

Country Status (1)

Country Link
CN (1) CN114492165A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117290750A (en) * 2023-07-03 2023-12-26 北京大学 Classification, association and range identification method for traditional village concentrated connection areas

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117290750A (en) * 2023-07-03 2023-12-26 北京大学 Classification, association and range identification method for traditional village concentrated connection areas
CN117290750B (en) * 2023-07-03 2024-03-19 北京大学 Classification, association and range identification method for traditional village concentrated connection areas

Similar Documents

Publication Publication Date Title
US11620487B2 (en) Neural architecture search based on synaptic connectivity graphs
Yan et al. An adaptive surrogate modeling based on deep neural networks for large-scale Bayesian inverse problems
US20230229891A1 (en) Reservoir computing neural networks based on synaptic connectivity graphs
US11593627B2 (en) Artificial neural network architectures based on synaptic connectivity graphs
US11625611B2 (en) Training artificial neural networks based on synaptic connectivity graphs
US11568201B2 (en) Predicting neuron types based on synaptic connectivity graphs
CN112464004A (en) Multi-view depth generation image clustering method
US11631000B2 (en) Training artificial neural networks based on synaptic connectivity graphs
Jiang et al. Active object detection in sonar images
He et al. Robust adaptive graph regularized non-negative matrix factorization
CN109409434B (en) Liver disease data classification rule extraction method based on random forest
CN114492165A (en) Parameter optimization method and system based on genetic breeding method
EP4073713A1 (en) Hyper-opinion classification using a neural network
CN115359297B (en) Classification method, system, electronic equipment and medium based on higher-order brain network
Ramasinghe et al. A learnable radial basis positional embedding for coordinate-mlps
Tiomoko et al. Deciphering lasso-based classification through a large dimensional analysis of the iterative soft-thresholding algorithm
Zeng et al. Contextual bandit guided data farming for deep neural networks in manufacturing industrial internet
CN111882441A (en) User prediction interpretation Treeshap method based on financial product recommendation scene
CN117238453A (en) Storage medium storing medical image feature processing method program
Watanabe et al. Summarizing data structures with gaussian process and robust neighborhood preservation
Rolon et al. A multi-class structured dictionary learning method using discriminant atom selection
Du et al. Probabilistic and Bayesian networks
Liang et al. Supervised and unsupervised learning models
Fischer et al. Preserving local densities in low-dimensional embeddings
Malhotra Transfer learning based entropy optimized semi-supervised decomposed vector-quantized variational autoencoder model for multiclass text classification and generation

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination