JP2008242927A - Genetic algorithm execution device and genetic algorithm execution method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To improve withdrawability from a local solution by a global search, while maintaining increase in the rate of approaching an optimum solution in a genetic algorithm. <P>SOLUTION: A genetic algorithm execution device comprises a mutation probability determination means 22 for obtaining a congestion degree, showing the convergence state of a parent individual 21 and determining the mutation probability from the congestion degree, and a mutation processing means 23 for performing mutation processing to genetic information of the individual based on the mutation probability. When the individuals of the local solution become predominant in a population 20, the congestion degree increases, thereby the mutation probability and strengthening the broad area search function is increased. By the special mutation processing for dynamically controlling the mutation probability, it is made possible to improve the probability of reaching an optimum solution regardless of a local solution, while maintaining the effect of increasing the rate of approaching the optimum solution. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a genetic algorithm execution apparatus and a genetic algorithm execution method, for example, a genetic algorithm suitable for a parameter adjustment apparatus capable of adjusting a large number of parameters of a physical model of a semiconductor device such as a transistor in a short time. The present invention relates to an execution device and a genetic algorithm execution method.

  When manufacturing an LSI, first, samples of several transistors (MOSFETs) having different shapes such as the channel length L and the channel width W of the gate of the transistor (MOSFET) are manufactured on the manufacturing line. Next, from the measurement result of the electrical characteristics of the prototype, a plurality of parameters of the physical model of the transistor are adjusted so as to match the characteristics of the transistor manufactured in the manufacturing line with high accuracy. Then, using the physical model of the transistor, various LSIs (transistors) manufactured in the manufacturing line are simulated by a known circuit simulator such as SPICE.

  The physical model of a transistor is a representation of the relationship between Vg (gate voltage), Vd (drain voltage), Id (drain current), etc., using a mathematical expression including variables such as channel length L and channel width W of the gate and a plurality of parameters. Many models have been proposed. In the simulation, for example, a typical well-known BSIM (Berkeley Short Channel 1 GFET Model) is used.

The BSIM is composed of a large number of mathematical expressions, and there are 50 or more parameters to be adjusted. Note that the details of the physical model of the transistor and the conventional parameter adjustment method are described in, for example, the following document, and thus detailed description thereof is omitted.
Published by Maruzen on February 28, 2002, supervised by Tatsushi Toriyabe, "MOSFET Modeling and BSIM3 User's Guide".

Conventionally, there has been proposed a parameter adjustment device that automatically performs parameter fitting (adjustment) processing of a physical model including a plurality of parameters using a genetic algorithm from a prototype result or the like. For example, the following document filed earlier by the present applicant proposes a parameter adjustment device that automatically performs parameter adjustment processing of a physical model including a plurality of parameters using a genetic algorithm.
JP 2005-38216 A

  In the conventional parameter adjustment method shown in Patent Document 1 described above, a special crossover method using random numbers is used as a crossover process. Although this crossover method is within a limited range, it has a structure similar to that of a mutation in which a parameter of genetic information changes at random, and therefore, the mutation has not been processed.

  However, the conventional method has a problem in that the wide-area search function, which is a feature of the genetic algorithm, is limited and may be caught by a local solution. An object of the present invention is to solve the above-mentioned problems of the prior art and to simultaneously improve the possibility of leaving from a local solution by a wide area search while maintaining the improvement of the speed of approaching an optimal solution.

  The genetic algorithm execution device according to the present invention is based on the congestion probability calculating means for determining the congestion degree indicating the convergence state of the individual, the mutation probability determining means for determining the mutation probability from the congestion degree, and the mutation probability. The main feature is that it comprises mutation processing means for performing mutation processing on the genetic information of the individual.

  Further, in the genetic algorithm execution device, the congestion degree calculation means, for each parameter that is genetic information, a value obtained by subtracting the minimum value from the maximum value of the individual parameter is calculated from the maximum value that can be taken by the parameter. Is divided by a value obtained by subtracting, and an average value obtained by averaging the values for all parameters is obtained, and a value obtained by subtracting the average value from 1 is used as the congestion degree.

  Further, in the above-described genetic algorithm execution apparatus, a parent individual selection unit that selects a predetermined number of parent individuals from a population and sets the plurality of parent individuals, and generates a plurality of child individuals from the plurality of parent individuals A child individual generating means, an evaluation value calculating means for calculating an evaluation value for a plurality of child individuals generated by the child individual generating means, and selecting a predetermined number of individuals having a good evaluation value from the plurality of child individuals Child individual selection means, parent individual extraction means for randomly extracting a predetermined number of parent individuals from the parent individual population, child individuals selected by the child individual selection means, and parent individuals extracted by the parent individual extraction means There is also a feature in that a predetermined number of individuals having a good evaluation value are selected from the above, and an upper individual selecting means for adding to the population is provided.

  In the above genetic algorithm execution device, the child individual generation means obtains a centroid in a vector space represented by the gene information of the parent individual, and the super-element on the vector space determined from the centroid and the value of the parent gene information. Another feature is that a generation range determining means for determining the generation range of the child population is provided inside the polyhedron.

  Further, in the genetic algorithm execution device described above, the parent individual selection means includes weight selection means for selecting a predetermined number of individuals so that an individual having a higher evaluation value has a higher probability of being selected, and individuals of the population. There is also a feature in that two-dimensional selection means for two-dimensionally arranging at random and selecting a predetermined number of individuals selected by the weighting selection means and individuals adjacent to the individual is provided.

  Further, in the above-described genetic algorithm execution device, an individual having each of a plurality of parameters of a physical model of a semiconductor device as a gene is defined, and a genetic algorithm is used based on the characteristic measurement data of the prototyped semiconductor device. There is also a feature in that a parameter adjusting means for optimizing the parameters is provided.

  The genetic algorithm execution method of the present invention includes a step of obtaining a degree of congestion indicating a convergence state of a plurality of parent individuals, a step of determining a mutation probability from the degree of congestion, and generating a child individual from the plurality of parent individuals. And a step of performing a mutation process on the genetic information of the offspring based on the mutation probability.

  In the genetic algorithm execution device of the present invention, the degree of congestion increases when the local solution becomes dominant in the population. Then, the probability of mutation increases and the wide area search function is further strengthened. The effect of improving the probability of reaching the optimal solution without being caught by the local solution while maintaining the effect of improving the speed of approaching the optimal solution by special mutation processing that dynamically controls the mutation probability. Can be made compatible.

  In the following embodiments, an example in which the present invention is applied to a parameter adjustment device is disclosed, but the present invention is applicable to any genetic algorithm execution device and execution method. Example 1 will be described below.

  FIG. 2 is a block diagram illustrating a hardware configuration example of a computer system for carrying out the present invention. For example, a keyboard 11 and a mouse 12 are connected as input devices to a known computer 10, and a known display device 15 is connected as an output device. Further, an external storage device 13 and a LAN 16 are connected for data input or output. The genetic algorithm execution apparatus of the present invention is realized by creating a program for executing the processing shown in the flowcharts described later and installing this program in a computer system.

  FIG. 1 is a flowchart showing an overall procedure when a simulation is performed using a parameter adjusting apparatus to which the present invention is applied. As described above, when an LSI is manufactured, first, in S10, samples of several transistors (MOSFETs) having different shapes such as the channel length L and the channel width W of the gate are manufactured on the LSI manufacturing line.

  FIG. 10 is an explanatory diagram showing a method for selecting the shape of a transistor to be prototyped. The shape selection method is, for example, such that the maximum value and the minimum value of the length L and the width W are equally spaced or more finely divided closer to the minimum value to form a cross shape on the LW plane. Select the category (shape) of the transistor to be prototyped.

  In S11, the electrical characteristics of the prototyped transistor are measured. Specifically, measurements for obtaining a plurality of sample values (discrete data) for the IdVd characteristics (Vb fixed), IdVd characteristics (Vg fixed), and IdVg characteristics (Vd fixed) are performed a plurality of times while changing the fixed values.

  In S12, the parameter adjustment process of the physical model of the transistor is performed by a method described later using the parameter adjustment apparatus to which the present invention is applied so as to match the characteristics of the transistor manufactured in the manufacturing line with high accuracy.

  In S13, using a physical model with adjusted parameters, a known circuit simulation program such as SPICE is used to perform an operation simulation of a transistor circuit having an arbitrary channel length and channel width manufactured in the manufacturing line. . By using the parameter adjustment apparatus to which the present invention is applied, a highly accurate parameter of the physical model can be obtained in a short time, and a highly accurate simulation can be performed.

  FIG. 3 is an explanatory diagram showing an outline of the crossover and mutation process of the present invention. This process is performed in the parameter adjustment process using the genetic algorithm of S12. FIG. 3 shows the contents of generation change processing of one generation of individuals (population) in the genetic algorithm. In actual processing, for example, this generation change processing is repeated until an individual with a good evaluation value is found.

  First, a predetermined number of individuals are selected (copied) from a population 20 including a predetermined number of individuals having a plurality of parameter values to be adjusted as genes, and a parent individual set 21 is obtained. As a selection method, in addition to a method of randomly selecting (Example 1), a method of combining selection based on an evaluation value and random selection (Example 2) is also disclosed.

  Next, the mutation probability is calculated from the parent individual set 21. Although details will be described later, first, the congestion degree S is calculated from the parent individual set 21, and the mutation probability M is calculated from the congestion degree S. In the crossover / mutation process 23, a special crossover process to be described later is executed, and a mutation process is also executed based on the mutation probability M to generate a set 24 of child individuals.

  In the evaluation value calculation / upper selection process 25, an evaluation value is calculated for each individual of the child individual set 24, and a predetermined number of individuals 26 with higher evaluation values are selected. Separately, a predetermined number of solids 27 are randomly extracted from the population 20 (not left in the population). In the upper selection process 28, a predetermined number of individuals 29 with higher evaluation values are selected from a set obtained by adding the individuals 26 and 27 and added to the population. This completes the generation change of one generation. If the number of solids 27 is equal to the number of solids 29, the number of solids in the population will not change.

  Hereinafter, the content of the parameter adjustment process using the genetic algorithm of S12 will be described. FIG. 4 is a schematic flowchart showing parameter adjustment (fitting) processing using a genetic algorithm. In S20, one or more discrete data groups (measurement results) obtained from the prototype transistor are prepared. In S21, N individuals having gene parameters as arbitrary parameters to be adjusted in the physical model function of the transistor are generated and set as an individual population.

  As described above, the BSIM has 50 or more parameters, but depending on the content to be simulated, there are parameters that may be fixed to representative values without adjustment or may be ignored. Accordingly, the number n of parameters to be adjusted differs depending on the purpose of simulation, and may be 50 or more, but may be 10, for example. Therefore, parameters such as the number N of individuals and the number of children generated (b) in the genetic algorithm are changed depending on the number pn of parameters to be adjusted. By this, if pn is small, a process will also become quick. In the embodiment, for example, the number of individuals N = pn × 15.

  In addition, in BSIM, a range of recommended parameter initial values for each parameter is determined, and therefore, for each parameter, an initial value is determined at random within the recommended parameter initial value range to be a gene value.

  In S21, the evaluation value of each individual is also calculated. An evaluation value is a value indicating how close a gene in an individual is to an ideal value as a model parameter. The evaluation value is calculated from the prepared discrete data group and the estimated data group calculated by the transistor electrical characteristic model function using the gene in the individual as a model parameter. As the evaluation function, a square error or an error rate is used.

  Note that when the scaling is different between discrete data groups, if a square error is taken, a data group having a small scale has a small influence on the evaluation value. Therefore, in the present invention, the discrete data in each data group is normalized, and the adjustment accuracy is improved by unifying the scale.

  In S22, a parent individuals are selected from the individual population generated in S21. For example, when the number of parameters to be adjusted is pn, a may be a = pn + 1. In S23, b child individuals are generated from the selected parent individual by crossover and mutation processing described later. In S24, the evaluation values of all the child individuals generated in S23 are calculated.

  In S25, c (for example, 2) are selected from the child individuals generated in S23 in the order of evaluation. In S26, c individuals are randomly selected from the population. In S27, c are selected from the 2c individuals selected or taken out in S25 and S26 in the order of good evaluation, and added to the population. One generation is updated by the above processing.

  In S28, it is determined whether or not to end the generation update process. If the determination result is negative, the process proceeds to S22, but if the determination is positive, the process ends. The termination condition is whether the evaluation value has reached a predetermined value or more, whether the increase rate of the evaluation value has fallen below a predetermined value, or whether the number of generation changes has reached a predetermined number.

  FIG. 5 is a flowchart showing the contents of the crossover and mutation process in S23. The crossover method used here is a crossover method for real values that generates a gene of a child individual from a polyhedron calculated from genes of a plurality of parent individuals. In S30, the center of gravity G of a parent individuals is obtained. That is, an average value is obtained for each individual parameter. In S31, the congestion degree S is calculated, and the mutation probability M is calculated based on the congestion degree S.

  FIG. 8 is an explanatory diagram showing the content of the degree of congestion S. In FIG. 8, for ease of illustration, the number of parameters is 2, the number of parent individuals is 3, and the minimum value that can be taken by the parameter is 0. For example, Ri is a value obtained by subtracting the minimum value from the maximum value that can be taken by the parameter Pi, and Qi is a value obtained by subtracting the minimum value from the maximum value of the parameter of the selected parent individual.

  When the parameter Pi does not converge to a specific value and is widely distributed, such as in the initial stage of search, Qi is large and close to Ri, but the parameter Pi has converged in the vicinity of the specific value. In this case, Qi decreases and approaches 0.

  Therefore, the degree of congestion S is defined as in Equation 1 below. That is, for each parameter, the value Qi obtained by subtracting the minimum value from the maximum value of the parameter of the parent individual is divided by the value Ri obtained by subtracting the minimum value from the maximum value that the parameter can take, and this is an average value for all parameters. A value obtained by subtracting 1 from 1 is defined as a congestion degree S. n is the number of parameters. The congestion degree S takes a value between 0 and 1.

  Further, the mutation probability M is obtained from the congestion degree S by the following formula 2. FIG. 9 is an explanatory diagram showing a linear function for obtaining the mutation probability M from the degree of congestion S. The constants e and f are determined by experiment so that the mutation probability M falls within the range of 0 to 1 and the search speed and the search range are balanced. For example, e is about 0.3 to 0.5. , F may be about 0 to 0.1.

In S32, the value of the variable i is set to 1, and in S33, a crossover process is performed to generate one child individual C from Equation 3 below using the center of gravity G and the uniform distribution random number. Here, p is the number of selected parent individuals, C is a vector indicating an individual child individual to be generated, and P _k is a vector indicating an individual of the selected parent individual. In this embodiment, it is assumed that the number of selected parent individuals a is pn + 1. U (0,1) is a uniformly distributed random number in the interval [0,1].

In Equation 3, first, x _k is obtained from the vector P _k of each parent individual, C _k is sequentially obtained up to C _a , and finally x _a and C _a are added to obtain a child individual C.

  FIG. 7 is an explanatory diagram showing a search range (child individual generation range) in the crossover process when the parameters to be adjusted are two, α and β, and the number of parent individuals randomly selected from the individual group is three. It is. The vector from the center of gravity G to each parent individual P1 to P3 is multiplied by ε to determine the generation range of the child individual (inside the outer triangle in FIG. 7), and the child individual is generated from the range using a uniform random number. . The recommended value of ε is √ (p + 1) when the number of parent individuals is p. When the number of parameters is 3 or more, the generation range of the child individual is an internal space of a super polyhedron surrounded by a plurality of hyper planes.

  By using the crossover method as described above, it is possible to handle the parameter explicitly for the problem that the parameter to be adjusted is a real value, and to perform effective adjustment. In addition, such a crossover method is robust to the dependency between variables, and does not depend on how to set the scale. The electrical characteristic model function of the transistor has a large dependency between parameters and a large number of parameters with different scaling. Suitable for parameter adjustment.

  In S34, a mutation process is performed. FIG. 6 is a flowchart showing the contents of the mutation process of S34. In S40, the variable i is set to 1. In S41, a uniformly distributed random number R is generated. In S42, it is determined whether or not the random number R is greater than the mutation probability M. If the determination result is negative, the process proceeds to S43, but if the determination is affirmative, the process proceeds to S44.

  In S43, a uniform random number is further generated to mutate the i-th parameter of the individual. That is, a new random parameter value that may be uniformly distributed within a range that the parameter can take is generated using a uniform random number, and the i-th parameter of the individual is overwritten.

  In S44, 1 is added to the variable i. In S45, it is determined whether or not the variable i is greater than the constant pn. If the determination result is negative, the process returns to S41, but if the determination is affirmative, the process ends. .

  In S35, 1 is added to the variable i. In S36, it is determined whether or not the variable i is greater than the constant b. If the determination result is negative, the process returns to S33, but if the determination is positive, the crossover process is terminated. To do. This process generates b child individuals. b is preferably about 10 × pn.

  By the parameter adjustment process as described above, highly accurate parameter adjustment can be performed in a short time. Then, by adopting the parameter in the physical model, a highly accurate circuit simulation can be performed without making a prototype, thereby improving the manufacturing efficiency of the semiconductor element.

  In the first embodiment, an example in which a individual is randomly selected from the population and used as a parent individual has been disclosed, but there is a problem in that the speed of evolution is slow. In addition, it is conceivable to select an individual having a high evaluation value with a higher probability. In this case, however, there is a problem that the individual of the local solution becomes dominant and the wide area search function is deteriorated. Therefore, in the second embodiment, in order to solve the above problem, a parent individual is selected by combining selection based on the evaluation value and random selection using a two-dimensional arrangement.

The difference between the first embodiment and the second embodiment is that the random selection process of S22 in FIG. 4 of the first embodiment is changed to the process of FIG. 11 in the second embodiment, and other processes are the same. It is. Therefore, only the differences will be described below.
FIG. 11 is a flowchart illustrating the contents of the parent individual selection process according to the second embodiment. In S50, the selection probability is calculated from the evaluation value for each individual of the population, and the winning range is further determined.

  FIG. 12 is an explanatory diagram illustrating a processing example of S50. In FIG. 12, four individuals having evaluation values as shown are shown. Therefore, first, the selection probability is calculated by dividing the evaluation value of each individual by the sum of the evaluation values of all the individuals (10 + 15 + 5 + 20 = 50 in FIG. 12). For example, the selection probability of the individual A is 10/50 = 0.2.

  Next, for each individual, a winning range is calculated with a value obtained by adding the selection probabilities of the individuals prior to the individual as a lower limit, and a value obtained by adding the selection probability of the individual to the lower limit as an upper limit. In this winning range, the range between 0 and 1 is divided without overlap by a length proportional to the evaluation value of each individual.

  In S51, the variable i is set to 1. In S52, random numbers uniformly distributed in the interval (0, 1) are generated. In S53, an individual whose random value is included in the winning range is determined as a winning individual. For example, in FIG. 12, if the random value is 0.4, the individual B is within the winning range of 0.2 to 0.5, and therefore the individual B is selected as the winning individual.

  In S54, it is determined whether or not the selected individual already overlaps. If the determination result is negative, the process proceeds to S55, but if the determination is affirmative, the process returns to S52. In S55, the winning individual is stored in the list. In S56, 1 is added to the variable i. In S57, it is determined whether or not the variable i is larger than a predetermined constant h. If the determination result is negative, the process proceeds to S52. Goes to S58. For example, h is 1 / n of the number of parent individuals a to be finally selected (n is an integer of 2 or more).

  In S58, all individuals in the population are randomly arranged in two dimensions. In S59, the variable i is set to 1, and in S60, an individual existing inside a rectangle having a predetermined size including the i-th selected individual is selected as a parent individual.

  FIG. 13 is an explanatory diagram showing the contents of the process of S60. Individuals are randomly arranged so as to fill a two-dimensional large rectangle (in FIG. 13, 10 horizontal × 6 vertical), and a rectangle 52 of a predetermined size including the winning individual 50 (in FIG. All the individuals 50 and 51 included in the upper left corner and a square of 2 × 2 = 4 individuals are selected as parent individuals. When the winning individual 50 is at the edge and the rectangle 52 protrudes from the two-dimensional arrangement, the individual is selected by folding the rectangle to the opposite side of the two-dimensional arrangement as shown in FIG.

  In S61, 1 is added to i. In S62, it is determined whether or not the variable i is larger than a predetermined constant h. If the determination result is negative, the process proceeds to S60. The process proceeds to S63. The vertical and horizontal lengths of the rectangle 52 are arbitrary as long as they are integer multiples of the individual interval, and the number of parent individuals to be finally selected is the above-described constant h × (number of individuals in the rectangle). Accordingly, the constant h and the vertical and horizontal lengths of the rectangle 52 are determined so that the number of parent individuals is close to the desired number. In S63, duplicate individuals are eliminated.

  Through the processing described above, the probability that an individual with a high evaluation value will be selected can be increased, and an individual with a low evaluation value can be selected with a certain probability. It is possible to simultaneously improve the possibility of leaving from the solution.

  Although the embodiments have been described above, the following modifications may be considered for the genetic algorithm execution apparatus of the present invention. Although an example in which the degree of congestion is calculated by using the average of the degrees of congestion of multiple parameters is disclosed, the degree of congestion is obtained for each parameter, and based on this degree of congestion, a different mutation probability is obtained for each parameter. Mutation processing may be performed.

  In the embodiment, an example has been disclosed in which the parameters e and f relating to mutation are constants. However, these parameters may be dynamically changed according to the initial stage, middle stage, and final stage of the search. For example, in the initial stage, the values of the parameters e and f may be increased in order to enhance the wide area search function, and the values of the parameters e and f may be decreased in order to increase the search speed after the middle stage.

  At the end of the search, the genetic algorithm may be switched to, for example, the known Powell method or other known local search methods. By switching the search method at the end, the parameter adjustment time is further shortened.

  The present invention can be applied to any apparatus that uses a genetic algorithm, including a parameter adjustment apparatus that uses a physical model of a transistor to simulate various LSIs manufactured in an LSI manufacturing line.

It is a flowchart which shows the whole procedure in the case of performing a simulation using the parameter adjustment apparatus to which this invention is applied. It is a block diagram which shows the hardware structural example of the computer system for implementing this invention. It is explanatory drawing which shows the outline of the crossover of this invention, and a mutation process. It is a schematic flowchart which shows the parameter adjustment (fitting) process using a genetic algorithm. It is a flowchart which shows the content of the crossover of S23, and a mutation process. It is a flowchart which shows the content of the mutation process of S34. It is explanatory drawing which shows the production | generation range of the child individual in a crossover process. It is explanatory drawing which shows the content of the congestion degree. It is explanatory drawing which shows the linear function which calculates | requires the mutation probability M from the congestion degree S. FIG. It is explanatory drawing which shows the selection method of the shape of the transistor to test. It is a flowchart which shows the content of the parent individual selection process in Example 2. FIG. It is explanatory drawing which shows the process example of S50. It is explanatory drawing which shows the content of the process of S60.

Explanation of symbols

20 ... Population 21 ... Parent individual 22 ... Displacement probability calculation process 23 ... Crossover and mutation process 24 ... Child individual 25 ... Evaluation value calculation, upper selection process 26 ... Upper child individual 28 ... Upper selection process 29 ... Upper individual

Claims (7)

A congestion degree calculating means for obtaining a congestion degree indicating a convergence state of a plurality of individuals;
A mutation probability determining means for determining a mutation probability from the degree of congestion;
A genetic algorithm execution device comprising: mutation processing means for performing mutation processing on gene information of the offspring based on the mutation probability.   The degree-of-congestion calculation means divides, for each parameter that is genetic information, a value obtained by subtracting the minimum value from the maximum value of the parent individual parameter by a value obtained by subtracting the minimum value from the maximum value that can be taken by the parameter. The genetic algorithm execution device according to claim 1, wherein an average value obtained by averaging all parameters is obtained, and a value obtained by subtracting the average value from 1 is set as the congestion degree. Furthermore,
A parent individual selection means for selecting a predetermined number of parent individuals from a population and making the plurality of parent individuals,
A child individual generating means for generating a plurality of child individuals from the plurality of parent individuals;
Evaluation value calculating means for calculating an evaluation value for a plurality of child individuals generated by the child individual generating means;
A child individual selecting means for selecting a predetermined number of individuals having a good evaluation value from the plurality of child individuals;
A parent individual extraction means for extracting a predetermined number of parent individuals at random from the parent individual population;
Selecting a predetermined number of individuals having a good evaluation value from the child individuals selected by the child individual selecting means and the parent individuals extracted by the parent individual extracting means, and adding to the population an upper individual selecting means. The genetic algorithm execution device according to claim 1, further comprising:   The child individual generation means obtains the center of gravity in the vector space represented by the gene information of the parent individual, and sets the generation range of the child individual group within the super polyhedron on the vector space determined from the value of the center of gravity and the parent gene information. 4. The genetic algorithm execution device according to claim 3, further comprising a generation range determination means for determining. The parent individual selection means is a weight selection means for selecting a predetermined number of individuals so that an individual with a better evaluation value has a higher probability of being selected;
4. The apparatus according to claim 3, further comprising: a two-dimensional selection unit that randomly two-dimensionally arranges individuals of a population and selects a predetermined number of individuals selected by the weight selection unit and individuals adjacent to the individual. The genetic algorithm execution device described.   A parameter adjustment means is provided that defines an individual having each of a plurality of parameters of a physical model of a semiconductor device as a gene and optimizes the parameter using a genetic algorithm based on the characteristic measurement data of the prototyped semiconductor device. The genetic algorithm execution device according to claim 1, wherein: Obtaining a degree of congestion indicating a convergence state of a plurality of parent individuals;
Determining a mutation probability from the degree of congestion;
Generating a child individual from the plurality of parent individuals;
Applying a mutation process to the gene information of the offspring individual based on the mutation probability.

Images