WO2003071480A1 - Active learning system, active learning method used in the same, program for the same, and recording medium containing the program - Google Patents

Abstract

A boosting device (1) obtains a hypothesis from initial data input from input means (100) and initial distribution of the data, calculates a prediction value of current data from the hypothesis, calculates an error from the prediction value, and updates the data distribution. Input candidate data selection means (22) uses the final hypothesis obtained by the boosting device (1) and selects input candidate data not contained in the past data, so that this input candidate data is instructed as a new experiment to an experimenter. Function value input means (23) receives a function value obtained by an experiment (30) of the experimenter for the aforementioned input candidate data. Data update means (24) adds the obtained input candidate data and its function value to the past data. Thus, when the I/O data items are continuous values and it is impossible to actually perform all the trials in the experiment, it is possible to estimate a target function with a small number of experiments.

Description

BACKGROUND OF THE INVENTION Active learning system, active learning method used therefor, program thereof, and recording medium storing the program

 The present invention relates to an active learning system, an active learning method used for the active learning system, a program for the active learning method, and an active learning method.

Regarding the recording medium on which the program is recorded, particularly experiments in fields such as molecular biology.

The present invention relates to a method for developing a plan for efficiently executing a plan.

 In general, in the conventional “design of experiment”, for example, a factor (which is called a factor) that has an effect on the purpose of improving the performance of a certain machine (this is called a characteristic) is extracted, and the factor is extracted. The purpose is to detect whether the characteristics can be improved by the possible values (this is called the level). In other words, the “experimental design method” analyzes the correlation between factors and its effect on characteristics by performing experiments with varying levels of factors.

 However, the conventional “experimental design method” does not always determine the level of factors to be used in the next experiment based on the experimental results obtained in the actual experiment. In other words, the traditional “design of experiment” is not a method of deciding the strategy of what experiment procedure to pursue in pursuit of characteristics.

 The problems arising in the above-mentioned conventional “experimental design method” will be described below with reference to “a specific problem of the binding ability of immunopeptides” as an example. The immune system recognizes antigens by binding cells in the body called MHC (Major Histocoatability Croples) cells to small fragments of proteins, called peptides, that the antigen has. It is considered. For this reason, the problem of identifying the binding ability of immunopeptides is one of the most important issues in medicine, especially in the field of immunology.

For example, suppose that a certain protein binds to an amino acid sequence having a length of 10 (hereinafter referred to as a peptide). At this time, it is assumed that the binding ability changes depending on the amino acid sequence. How If we try to find out if a peptide has a high binding ability to this protein by conducting biochemical experiments, we will produce all 10-length peptides and conduct experiments to measure each binding ability It is necessary to go through the procedure.

However, since the number of amino acids is 20, there are 201 ^Q combinations for a peptide of length 10. It is almost impossible to obtain data on binding ability to all these sequences from biochemical experiments.

 In other words, more generally, when we want to know the function f (y = f (χ)) that expresses the relationship between X and y from the variable X and the data of the corresponding label (output) y, The range of the variable X is very wide, and it is often not practically possible to experimentally determine the value of f (X) for all X. In such a case, according to the conventional “experimental design method”, analysis cannot be performed unless a number of experiments that cannot be actually performed is performed, that is, analysis cannot be actually performed. As described above, there has not been known a highly accurate method for estimating the function f while deciding the next input candidate data to be tested from the data (X, y) obtained in the past. Such a method is long-awaited. Summary of the Invention

An object of the present invention is to solve the above-described problem, and to perform an experiment in which it is practically impossible to perform all of the target trials when input / output data is a continuous value. It is an object of the present invention to provide an active learning system capable of estimating the value, an active learning method used therein, a program thereof, and a recording medium storing the program. An active learning system according to the present invention is an active learning system that executes an active learning method for estimating a functional relationship that at least approximately holds between input and output data based on selected input and output data, A function relationship is estimated from past input / output data using a certain expression, and a plurality of learning algorithms of any of a predetermined single type and a plurality of types are used as lower-level algorithms, and the lower-level algorithm is used as the lower-level algorithm. Learning means for learning input / output data up to the point of selecting the next input data as training data; and a plurality of input candidate data using a final hypothesis obtained from a plurality of hypotheses obtained from an algorithm of each of the learning means. Function predictor that predicts function values And input data designating means for selecting input data from the input candidate data based on a function value using the variance of the learned predicted values of the plurality of hypotheses. Another active learning system according to the present invention further improves the accuracy by repeating data sampling in combination with a boosting technique that improves the accuracy of a passive learning algorithm.

 In one configuration example of the active learning system according to the present invention, in a learning unit and a function prediction unit, a lower algorithm performs learning on a plurality of input candidate data selected at random. In the input data designating means, the data in which the variance of the predicted values of a plurality of hypotheses is the largest is selected as the input data.

 Another configuration example of the active learning system according to the present invention includes data updating means for adding input candidate data and function values to past data.

 Still another configuration example of the active learning system according to the present invention includes an experiment means for experimentally finding a function value for input candidate data.

 Another configuration example of the active learning system according to the present invention includes simulation means for obtaining a function value for input candidate data by simulation.

 In still another configuration example of the active learning system according to the present invention, in the learning means, any one of a conclusion according to an arbitrary regularity and a result obtained by a simulation calculation is used as input data of the lower algorithm.

 The active learning method according to the present invention is an active learning method for estimating a functional relationship that at least approximately holds between input and output data based on selected input and output data. Use shape! /, Input / output up to the point when the next input data is selected for the lower-level algorithm by estimating the functional relationship and using a plurality of learning algorithms of one or more of predetermined single types as the lower-level algorithm A learning step of learning data as training data; and a function prediction step of predicting a function value for a plurality of input candidate data using a final hypothesis obtained from a plurality of hypotheses obtained from the respective algorithms in the learning step. And an input data designating step of selecting input data from the input candidate data based on a function value using a variance of predicted values of the plurality of learned hypotheses.

Another active learning method according to the present invention is to repeat data resampling. The method further includes a boosting step of improving the learning accuracy of the learning step by a boosting technique for improving the accuracy of a passive learning algorithm.

 In one configuration example of the active learning method according to the present invention, in a learning stage and a function prediction stage, a lower algorithm performs learning on a plurality of input candidate data selected at random. In the input data designating step, the data having the largest variance of the predicted values of a plurality of hypotheses is selected as input candidate data.

 A recording medium on which a program for an active learning method according to the present invention is recorded is an active learning system for executing an active learning method for estimating a functional relationship that at least approximately holds between input and output data based on selected input and output data. A computer-readable recording medium on which a program for causing a computer to function as a computer is recorded. A learning process of learning input / output data as training data up to a point at which the next input data is selected by the lower algorithm using a plurality of learning algorithms of any of a plurality of types as lower algorithms; and the learning process. Using the final hypothesis obtained from the multiple hypotheses obtained from each algorithm, Function prediction processing for predicting a function value for force candidate data; and input data for selecting input data from the input candidate data based on a function value using a variance of predicted values of the plurality of learned hypotheses. The program for executing the specified process is recorded.

A recording medium on which another active learning method program according to the present invention is recorded is provided with a boosting technique for improving the accuracy of a passive learning algorithm by repeating the resampling of data by a computer. A program for further executing the boosting process for improving the learning accuracy of the learning process is recorded. One configuration example of a recording medium on which the program of the active learning method according to the present invention is recorded is a computer in which a lower algorithm learns a plurality of randomly selected input candidate data in a learning process and a function prediction process. It records a program to execute the process of performing Another example of the configuration of the recording medium is to allow a computer to execute a process of selecting data having a maximum variance of predicted values of a plurality of hypotheses as input candidate data in an input data designation process. Program Have recorded.

 A program for an active learning method according to the present invention causes a computer to function as an active learning system that executes an active learning method for estimating a functional relationship that at least approximately holds between input and output data based on selected input and output data. A program for estimating a functional relationship from a past input / output data using a certain expression form, and learning one or a plurality of learning algorithms of one or more of predetermined types. Is used as a lower algorithm, the lower algorithm learns input / output data up to the point at which the next input data is selected as training data, and a plurality of hypotheses obtained from the respective algorithms of the learning process. A function prediction process for predicting a function value for a plurality of input candidate data using a final hypothesis; More it is made to execute the input data designation processing for selecting input data from the input candidate data distribution based on the value of functions with the predicted value of the hypothesis. Another active learning program according to the present invention is a boosting technique for improving the accuracy of a learning process by boosting a computer to improve the accuracy of a passive learning algorithm by repeating data resampling. The process is being executed further.

 One example of the configuration of the program of the active learning method according to the present invention causes a computer to execute a process in which a lower-level algorithm performs learning on a plurality of randomly selected input candidate data in a learning process and a function prediction process. ing. Another configuration example of the program causes a computer to execute a process of selecting, as input candidate data, data in which a variance of predicted values of a plurality of hypotheses is maximum in an input data designation process.

 That is, the active learning system of the present invention proposes an experiment design method using active learning for continuous input / output data in order to solve the above-mentioned problems. In other words, in the active learning system of the present invention, the “boosting” technology, which is a method of improving the accuracy of the passive learning algorithm by repeating data sampling, is one of the leading methods of active learning. By appropriately combining with a certain group question learning, high-accuracy active learning is realized and applied to the planning of experiment.

As described above, in the active learning system of the present invention, the accuracy between the input and the output is In a design of experiment, which estimates functional relationships that approximately hold based on input / output data obtained from experiments, a predetermined expression that estimates functional relationships from past input / output data using a certain expression A learning step in which the learning algorithm is used as a lower-level algorithm, and the lower-level algorithm learns input / output data up to that point as training data; a boosting step in which the learning accuracy of the learning step is improved by boosting technology; and a boosting step. A function value prediction step of predicting function values for a plurality of input candidate data using a final hypothesis obtained as a weighted average of a plurality of hypotheses, and a function value using a variance of predicted values of a plurality of hypotheses. Input data selection step of selecting input candidate data as input based on the input data. Programming can be obtained.

 In other words, in the active learning system of the present invention, in an experimental design method for estimating a functional relationship established between an input and an output based on input / output data obtained by experiments, a hypothesis is obtained from past data by passive learning. Estimating a hypothesis, estimating the next input candidate data based on the hypothesis, specifying input data based on the hypothesis, experimentally obtaining a function value for the input caught data by experiment, and calculating the input candidate data and the function value. An experimental design method is obtained, which includes a data updating step to be added to past data. In this case, the above steps are executed at each time of selecting the input data of the next experiment, that is, each time the input data of the next experiment is selected until the estimation of the functional relationship is completed.

 Here, it is preferable to use boosting in the hypothesis estimation stage. In the input data designating step, an input candidate data generating step of generating a plurality of input candidate data not present in the past data, and an input candidate data of selecting the next input candidate data from the plurality of input candidate data based on a hypothesis. Data selection stage.

Further, in the active learning system of the present invention, an experimental design program for estimating a functional relationship that is established accurately or approximately between an input and an output based on input / output data obtained by an experiment includes: An arbitrary learning algorithm that estimates a functional relationship from data using a certain expression is used as a lower algorithm, and the lower algorithm learns input / output data up to that point as training data. Boo boosted by boosting technology The computer performs a sting process and an input data designation process that selects input candidate data as input based on the value of a function that uses the variance of the predicted values of multiple hypotheses that have been improved by the boosting process. .

 In other words, in the active learning system of the present invention, a hypothesis is estimated from past data by passive learning in an experiment design program for estimating a functional relationship between input and output based on input / output data obtained by experiments. Hypothesis estimation processing, input candidate data instruction processing to indicate the next input candidate data based on the hypothesis, experimental processing to experimentally determine the function value for the input candidate data, and past input data and function values to past data The data update process to be added to the computer is executed by the computer. In this case, each of the above steps is performed at each time of selecting the input point of the next experiment, that is, each time the input of the next experiment is selected until the estimation of the functional relationship is completed. Here, the hypothesis estimation processing preferably includes a boosting processing. The input candidate data processing includes input candidate data generation processing for generating a plurality of input candidate data that are not included in past data, and input candidate data selection for selecting the next input candidate data from the plurality of input candidate data based on a hypothesis. And data selection processing.

 The “experimental design method” according to the present invention is as follows when the expression of the conventional “experimental design method” is used with respect to the above-mentioned conventional “experimental design method”. That is, the "experimental design method" according to the present invention is such that, when estimating a function f having factors as inputs and characteristics as outputs, the level of the factor of the experiment for which the result is desired is to be specified one by one. This is a method to estimate the function で with a small number of experiments. Furthermore, generalizing the expression, the order of experiments to be performed to efficiently estimate the function (hypothesis) f (y = f (x)) representing the relationship between the input X and its label (output) y Point to.

 In the present invention, the above-described experiment design method is realized by active learning. Active learning refers to learning as described in, for example, Abe and Nakamura's commentary “Active Learning Overview” (“Information Processing”, Vol. 38, No. 55, pp. 58-56, 1997). Is a method by which a person selectively obtains information from the environment and learns.

Many methods have been proposed for active learning. For example, the paper "Baijian esperimentanore design: Alebiyu" by Charoner (Cha 1 oner) and Verdenerelli (Vayesianex) peri me nta 1 design: a Revi ew). ("Statistical Science", Vol. 10, pp. 273-304, published in 1995) An active learning method based on it has been proposed.

 Also, a paper by Baum, “Neuralnetalgorithms thatlearninpo 1 ynomia 1 time fr om ex amp lesandqueries” (“I EEE Transactions on Neural Networks”, Vol. 2, pp. 5-19, published in 991) is based on past learning data. A learning method has been proposed in which a point estimated as is used as new input candidate data. Along with these methods, a learning method called "group query learning" for realizing data efficiency has been proposed, and is considered as one of the leading active learning methods. A typical group question learning method is the query by the SENG et al., “Query pie committee” (“Proceeding sub-sub-fifth annual EM workshop on-committee”). The method is described in the Proceedings of the Shonan Reeling Series (Proceedings of fiftha nn ua 1 ACM wo rkshoponc omp utationa 1 learningthoery), pp. 287-294, 1992. In this method, a plurality of randomized learners make predictions, and learning is performed with the point having the highest degree of mismatch as new input candidate data.

On the other hand, learning without active learning is called passive learning. 'In passive learning, the learner does not selectively obtain information from the environment, but estimates the rules (hypotheses) inherent in the data from a given large amount of data. In passive learning, in recent years, a technique called `` boosting '', which improves learning accuracy by repeating resampling so that data areas for which the hypothesis up to that point is not good at given data is trained, has been attracting attention in recent years. Has been confirmed to have high performance experimentally. Is coming.

 The method of "boosting" is described in a paper by Freund and Shapire, "Advice Decision Theoretic Generalization-Online Learning and Application Application Boostink" (A decision—theoreticgeneralizationofon—linelearninganaanappi icationtoboosting) (Proceedings of the Proceedings of the second europeanconferenceoncomp utational 1 earningthoery, pp. 23, 23, 9-95) It is described in. In addition, a paper by Freund and Shapire, "EX perimentswithan ew boostingalgor ltm" ("Proceedings of the Third International") Proceedings of Conference on Machine Learning (Proceedingsofthethirte enthinternationalconierenceon machinelearning), pp. 148-156, published in 1996), the “boosting” technique in passive learning is a technique for learning algorithms with reasonable learning performance. It demonstrates experimentally the superiority of using it to construct a learning algorithm with arbitrary learning accuracy.

 However, boosting techniques have only been used in passive learning and have not yet been applied to active learning. In other words, in active learning, there has never been a concept unique to “boosting” technology, such as repeating resampling, and no method of using this technology in the framework of active learning has been proposed.

As active learning, a technique disclosed in Japanese Patent Application Laid-Open No. 11-316754 is already known. However, the feature of this active learning method is that the point at which the difference between the sum of weights of the output values with the largest sum of the weights and the sum of the weights of the output values with the next largest sum is the smallest is the input candidate point. And select them as input from them. However, this method targets only certain discrete numerical values with output values. The most significant feature of the present invention is that a variance function is used as a solution to the problem in the case of continuous values, and the technique disclosed in Japanese Patent Application Laid-Open No. different. BRIEF DESCRIPTION OF THE FIGURES

 FIG. 1 is a diagram showing a flow of an active learning algorithm according to an embodiment of the present invention.

 FIG. 2 is a block diagram showing a configuration of the active learning machine according to the embodiment of the present invention. FIG. 3 is a block diagram showing the configuration of the active learning system according to the first embodiment of the present invention.

 FIG. 4 is a block diagram showing a configuration of an active learning system according to a second embodiment of the present invention.

 FIG. 5 is a block diagram showing the configuration of the active learning system according to the third embodiment of the present invention.

 FIG. 6 is a block diagram showing a configuration of an active learning system according to a fourth embodiment of the present invention.

 FIG. 7 is a block diagram showing the configuration of the active learning system according to the fifth embodiment of the present invention.

 FIG. 8 is a block diagram showing a configuration of an active learning system according to a sixth embodiment of the present invention.

 FIG. 9 is a block diagram showing the configuration of the active learning system according to the seventh embodiment of the present invention.

 FIG. 10 is a block diagram showing a configuration of an active learning machine according to an eighth embodiment of the present invention. Detailed description of the embodiment

Next, embodiments of the present invention will be described with reference to the drawings. FIG. 1 is a diagram showing a flow of an active learning algorithm according to an embodiment of the present invention, and FIG. It is a block diagram showing composition of an active learning machine by an example.

 In FIG. 1, the boosting device 1 includes a learning prediction error calculation unit 11, a data distribution updating unit 12, and a judgment unit 13, a function calculation unit 21, and input candidate data selection unit 22. , A function value input means 23, a data update means 24, a determination means 25, an apparatus including an experiment simulation 30, and an input means 100. Here, the experiment simulation 30 need not necessarily be included in the apparatus, and may be replaced with an external simulator / manual experiment / simulation calculation. It should be noted that it is also possible to realize by a program by replacing each of the above means with a step.

 In FIG. 2, the active learning machine 4 includes a preprocessing device 41, a lower learning machine device 42 composed of lower learning machines 4 2 1 to 4 2 N, and a hypothesis' predicted value storage area 4 3 1 to 4 3 N. It is composed of a prediction value storage device 43, a post-processing device 44, and a recording medium 45 for recording a program to be executed by a computer when each of the above devices is realized by a computer. I have.

 The active learning machine 4 includes a computer and software that causes the computer to function as the active learning machine 4. In this case, when the program recorded on the recording medium 45 is read into the computer, this program and the hardware resources of the computer cooperate with each other, and the preprocessing device 41 and the lower learning machine device 42 are connected to the computer. Hypothesis · The predicted value storage device 43 and the post-processing device 44 are realized. The experiment design method according to the embodiment of the present invention will be described with reference to FIGS. In this embodiment, first, initial data and initial distribution of the data are input to the input means 100.

 Next, as shown in FIG. 1, the questions in the active learning are repeated between the learning prediction error calculating means 11 and the judging means 25. The first part of the iteration is performed in the process of the boosting device 1. This processing is performed in the active learning machine 4 by the preprocessing device 41 and the lower learning machine 42.

In the boosting device 1, resampling of data is repeated by the learning prediction error calculating means 11, the data distribution updating means 12, and the judging means 13. First, in the boosting device 1, the learning prediction error calculation means 11 Learning from the current data is performed using a learning algorithm, and a hypothesis is obtained as the learning result. The lower learning algorithm may be any learning algorithm that handles continuous values, such as a regression tree. The result is output to the hypothesis / predicted value storage device 43. Further, the learning prediction error calculation means 11 calculates a prediction value from the hypothesis to the current data, and calculates an error from the prediction value. Next, the data distribution updating means 12 updates the data resampling distribution from the error obtained by the learning prediction error calculating means 11. '

 Finally, if the number of times of resampling has reached the fixed value (T) in the judging means 13, the processing by the boosting device 1 ends, otherwise, the process returns to the learning prediction error calculating means 11, and The above processing is repeated.

 Further, the number of repetitions of the processing of the boosting device 1 may be one. At that time, the data distribution updating means 12 updates the data distribution and the condition judgment statement of the judging means 13 and the loop returning from the judging means 13 to the learning prediction error calculating means 11 becomes unnecessary. In other words, lower learning processing without boosting processing is performed.

 After the processing by the boosting device 1 is completed, the processing by the post-processing device 4 is performed. In this case, the post-processing device 44 includes function calculation means 21, input candidate data selection means 22, function value input means 23, data update means 24, determination means 25, and an experimental simulation. 30. The function calculation means 21 calculates the final hypothesis using the function value using the variance of the predicted values of the plurality of learned hypotheses and the plurality of hypotheses, and outputs the final hypothesis. The input candidate data selection means 22 selects input candidate data based on the value of the function using the variance obtained from the calculation result by the function calculation means 21. Indicate as a new experiment for simulation 30).

The function value input means 23 receives the function value of the input candidate data tested in the experiment (experimental simulation 30) by the experimenter. The data updating means 24 adds the obtained input candidate data and its function value to the past data. Finally, the determining means 25 determines whether or not the number of questions has reached a certain value (N). If not, the process returns to the boosting device 1 to continue active learning. If the number of questions reaches a certain value, the active learning ends. Next, the correspondence between each device shown in FIG. 2 and each main means shown in FIG. 1 will be described. In this embodiment, the lower-order learning machine device 42 and the hypothesis / prediction value storage device 4 3 are combined and operated as the learning prediction error calculation means 11, so that the time until the next input data is selected for the lower-order algorithm is obtained. It functions as learning means for learning input / output data as training data and function prediction means for predicting function values for a plurality of input candidate data.

 Further, the lower learning machine device 42 also operates as the data distribution updating device 12 and the judging device 13 and acts as a boosting device for improving the learning accuracy of the learning device.

 In addition, the post-processing device 44 operates as the function calculating means 21 and the input candidate data selecting means 22, so that the input candidate data based on the variance of the predicted values of the plurality of learned hypotheses is used. Acts as input data designating means for selecting input data from Further, the pre-processing device 41 and the post-processing device 44 cooperate with each other to operate as the function value input means 23, the data updating means 24, and the experiment simulation 30, so that the input candidate data and the function value can be obtained. Acts as a data updating means to add to past data. As a result, it becomes possible to update the experimental result database 6 and input to the active learning machine 4 without manual intervention. In this way, it can be regarded as another learning system, and the user's labor can be greatly reduced. Although the embodiment of the active learning machine 4 according to the present invention has been described above, the present invention is not limited to this, and it is possible to make changes and improvements within the ordinary knowledge of those skilled in the art. it is obvious.

 For example, in the learning prediction error calculation means 11 acting as the learning means and the function prediction means, the lower algorithm may perform learning on a plurality of input candidate data selected at random. By choosing at random, the lower learning algorithms become independent of each other and parallelization becomes easier. Furthermore, since only random numbers are generated, the computation time is shorter than other methods.

Further, the input candidate data selecting means 22 acting as a part of the input data designating means may select, as the input candidate data, data in which the variance of the predicted values of a plurality of hypotheses is maximum. Data with the largest variance has the least amount of information available. Since it is large, the effect of learning can be maximized by selecting it as input candidate data.

 Further, in the learning prediction error calculating means 11 acting as a learning means, any one of a conclusion according to an arbitrary regularity and a result obtained by simulation calculation may be used as input data of the lower algorithm. Les ,. By adopting such a system, it is possible to handle calculated or converted values based on experimental or simulation results, and to actively learn a wider range of data. .

 FIG. 3 is a block diagram showing the configuration of the active learning system according to the first embodiment of the present invention. In FIG. 3, the active learning system 5 includes an active learning machine 4 and a storage device [experiment result database (attribute values and results)] 6 for storing past inputs and outputs.

 Here, the active learning machine 4 performs the same operation as the above-described embodiment of the active learning machine 4 according to the present invention. Although the active learning system 5 has the same input and output as the active learning machine 4 shown in FIG. 2, a system that does not require an external storage device can be constructed.

 FIG. 4 is a block diagram showing a configuration of an active learning system according to a second embodiment of the present invention. In FIG. 4, the active learning system according to the second embodiment of the present invention is an example of a system incorporating an active learning machine 4.

 The next input candidate output by the system of the active learning machine 4 is manually experimented by the experimenter and simulation calculation 7 by the experimenter.The experimenter actually performs the experiment, and the set of the input candidate and the experiment result is stored in the storage device 6. . In addition, the active learning machine 4 performs learning by using it as training data, and outputs the next input candidate.

 By repeating the above operations, the experimenter can efficiently select an infinite number of experimental objects from the innumerable experimental objects according to the instructions of the active learning machine 4. Experiments can be performed efficiently.

FIG. 5 is a block diagram showing a configuration of the Noh learning system according to the third embodiment of the present invention. In FIG. 5, an active learning system 5 is a system in which a storage device 6 is combined with an active learning machine 4. In this embodiment, the storage device 6, which was required in the second embodiment of the present invention, is no longer required outside the active learning system 5, so that the space can be reduced and the data transfer time between the database and the active learning machine can be reduced. Can be reduced.

 FIG. 6 is a block diagram showing a configuration of an active learning system according to a fourth embodiment of the present invention. In FIG. 6, the active learning system according to the fourth embodiment of the present invention is an example of a system incorporating an active learning machine 4. The simulator 8 automatically performs a simulation corresponding to the next input candidate output by the system of the active learning machine 4, and stores a set of the input candidate and the simulation result in the storage device 6. Further, the active learning machine 4 performs learning using the training data as training data, and outputs the next input candidate. By repeating such an operation, the experiment target can be efficiently selected from the myriad of experiment targets, and the user can select only the combination of the attribute of the input candidate data and the result value. If you do, you will be able to automatically carry out experiments that meet your purpose. This eliminates the need for the experimenter to process the input candidate data selected by the active learning engine, thus reducing the burden on the experimenter. However, in this active learning system, it is necessary for the user to prepare a system (for example, a computer or a device for performing a simulation) necessary for the simulator 8 or the like.

 FIG. 7 is a block diagram showing a configuration of an active learning system according to a fifth embodiment of the present invention. In FIG. 7, an active learning system 5 is configured by incorporating a storage device 6 into an active learning machine 4.

 By using this active learning system 5, the storage device 6, which was required in the above-described fourth embodiment of the present invention, is not required outside the system. The data transfer time between the two can be reduced.

FIG. 8 is a block diagram showing a configuration of an active learning system according to a sixth embodiment of the present invention. In FIG. 8, an active learning system 9 is configured by incorporating a simulator 8 into an active learning machine 4. By adopting such a system, an active learning system including simulation can be constructed, and optimal prediction simulation can be automatically performed without human intervention. This In other words, the user can obtain the desired result automatically without any trouble if only the initial conditions are input, so that the burden on the user can be reduced.

 Although the purpose of the active learning system 9 is limited, the system user only needs to prepare the storage device 6, and the simulator 8 and the like are different from those of the above-described fifth embodiment of the present invention. It is possible to reduce the trouble of preparing an external system.

 FIG. 9 is a block diagram showing a configuration of an active learning system according to a seventh embodiment of the present invention. In FIG. 9, the active learning system 10

8 and a storage device 6. By adopting such a system, the burden on the user is drastically reduced as compared with the above-described first to sixth embodiments of the present invention, and the person is required only to combine the attribute of the input candidate data with the result value. If you give, you will be able to automatically perform the experiment that meets your purpose.

 FIG. 10 is a block diagram showing a configuration of an active learning machine according to an eighth embodiment of the present invention. In FIG. 10, an active learning machine 40 according to an eighth embodiment of the present invention includes an input data generation machine 4 of a lower learning machine 42 between a preprocessing device 41 and a lower learning machine 42. The configuration is the same as that of the active learning machine 4 shown in FIG. 2 except that 6 is provided, and the same components are denoted by the same reference numerals.

 The input data generation machine 46 includes an input data generation machine 46 1 to 46 N. The data stored in the storage device 6 is processed and supplied to the lower learning machine 42, so that the lower learning is performed. The mechanical device 42 can learn from data provided from the input data generating mechanical device 46, and can be applied to a wider range of problems.

 The active learning machine 40 includes a computer and software that causes the computer to function as the active learning machine 40. In this case, when the program recorded on the recording medium 45 is read into the computer, the program and the hardware resources of the computer cooperate with each other, and the preprocessing device 41 and the input data generation device 46 are connected to the computer. The lower learning machine device 42, the hypothesis 'predicted value storage device 43' and the post-processing device 44 are realized.

Thus, according to the present invention, in passive learning, moderate accuracy can be obtained. Boosting technology that can improve the prediction accuracy to an arbitrary accuracy if it is a possible learning algorithm is combined with group query learning, which is an active learning method, by using an appropriate method to improve accuracy. It is possible to realize a high active learning method.

 In addition, in the present invention, the above-described method is used to formulate an experiment plan for experiments in which it is impossible to perform all trials realistically because the number of experiments is large or the load of one trial is large. By applying this, it is possible to realize an efficient and accurate experimental design. As described above, the active learning system of the present invention is an active learning system that executes an active learning method for estimating a functional relationship that at least approximately holds between input and output data based on the selected input and output data. A functional relationship is estimated from past input / output data using a certain expression form, and a plurality of predetermined learning algorithms, one of a single type or a plurality of types, are used as lower-level algorithms, and the next lower-level algorithms are used. A learning means for learning input / output data up to the point of selecting input data as training data, and a final hypothesis obtained from a plurality of hypotheses obtained from each algorithm of the learning means, to obtain a function value for a plurality of input candidate data using a final hypothesis. Function to make predictions Predictive means and input based on function values using the variance of predicted values of multiple learned hypotheses By providing input data designating means for selecting input data from among them, when the input / output data is a continuous value, when performing an experiment in which it is virtually impossible to perform all the trials in question, a small number of experiments can be performed. The effect that the objective function can be estimated by the number of times is obtained.

Another active learning system of the present invention further includes a boosting means for improving the learning accuracy of the learning means by a boosting technique for improving the accuracy of a passive learning algorithm by repeating resampling of data. In the case of passive learning, a boosting technique that can improve the prediction accuracy to an arbitrary level as long as it is a learning algorithm that can achieve a moderate level of accuracy is called a group learning method that is an active learning method. By combining these methods with other methods, it is possible to achieve an effect that highly accurate active learning can be realized. As described above, the active learning system according to the present invention and the active learning method used in the active learning system perform an experiment in which it is practically impossible to perform all the trials when the input and output data are continuous values. It is useful as a system and method for estimating an objective function with a small number of experiments, and is particularly suitable for making a plan for efficiently performing experiments and designs in fields such as molecular biology.

Claims

The scope of the claims 1. An active learning system that executes an active learning method that estimates a functional relationship that at least approximately holds between input and output data based on the selected input and output data. A function relationship is estimated using an expression, and a plurality of learning algorithms, one of a predetermined single type or a plurality of types, are used as lower-order algorithms. Learning means for learning input / output data as training data;  Function predicting means for predicting a function value for a plurality of input candidate data using a final hypothesis obtained from a plurality of hypotheses obtained from each algorithm of the learning means; Input data designating means for selecting input data from the input candidate data based on a value of a function using variance; and An active learning system comprising:  2. Boosting means for improving the learning accuracy of the learning means by a boosting technique for improving the accuracy of the passive learning algorithm by repeating resampling of data is further provided. The active learning system according to claim 1, wherein:  3. In the learning means and the function prediction means,  The lower algorithm performs learning on the plurality of input candidate data selected at random. 3. The active learning system according to claim 1 or 2, wherein: 4. In the input data designating means,  Selecting the data with the largest variance of the predicted values of the plurality of hypotheses as the input candidate data An active learning system according to any one of claims 1 to 3, characterized in that: 5. Includes data updating means for adding the input candidate data and the function value to the past data The active learning system according to any one of claims 1 to 4, characterized in that:  6. The active learning system according to any one of claims 1 to 5, further comprising an experiment means for experimentally finding a function value for the input candidate data.  7. Includes simulation means for obtaining a function value for the input candidate data by simulation The active learning system according to any one of claims 1 to 4, characterized in that: '  8. In the learning means,  The active learning according to any one of claims 1 to 7, wherein any one of a conclusion according to an arbitrary regularity and a result obtained by a simulation calculation is used as input data of the lower algorithm. System.  9. An active learning method for estimating a functional relationship that at least approximately holds between input and output data based on the selected input and output data,  A functional relationship is estimated from a past input / output data using a certain expression, and a plurality of learning algorithms of a predetermined single type or a plurality of types are used as a lower algorithm, and the next lower algorithm is used. A learning step of learning input / output data up to the point of selecting input data as training data;  A function prediction step of predicting a function value for a plurality of input candidate data using a final hypothesis obtained from a plurality of hypotheses obtained from each algorithm of the learning step; An input data specifying step of selecting input data from the input candidate data based on a value of a function using variance; 10. The active learning method according to claim 10, wherein:  10. Further includes a boosting step for improving the learning accuracy of the learning step by a boosting technique for improving the accuracy of the passive learning algorithm by repeating data resampling. 10. The active learning method according to claim 9, wherein: 1 1. In the learning step and the function prediction step,  The lower algorithm performs learning on the plurality of input candidate data selected at random. The active learning method according to claim 9 or 10, wherein:  1 2. In the input data designating step,  Selecting the data with the largest variance of the predicted values of the plurality of hypotheses as the input candidate data The active learning method according to any one of claims 9 to 11, wherein: 1 3. Based on the selected input and output data, a program for causing the computer to function as an active learning system that executes an active learning method for estimating a functional relationship that at least approximately holds between the input and output data. A computer-readable recording medium on which  On the computer,  A functional relationship is estimated from a past input / output data using a certain expression, and a plurality of learning algorithms of a predetermined single type or a plurality of types are used as a lower algorithm, and the next lower algorithm is used. A learning process of learning input / output data up to the point of selecting input data as training data;  A function prediction process of predicting a function value for a plurality of input candidate data using a final hypothesis obtained from a plurality of hypotheses obtained from the respective algorithms of the learning process; and a prediction value of the learned plurality of hypotheses. Input data designation processing for selecting input data from the input candidate data based on a value of a function using variance; and A computer-readable recording medium on which a program for executing the program is recorded. 1 4. On the computer,  A program is recorded for further executing a boosting process for improving the learning accuracy of the learning process by a boosting technique for improving the accuracy of a passive learning algorithm by repeating data resampling. The computer-readable recording medium according to claim 13, wherein: 1 5. In the learning process and the function prediction process, 15. The program according to claim 13, wherein a program for causing said lower algorithm to execute a process of performing learning on each of said plurality of input candidate data selected at random is recorded. Computer readable recording medium.  1 6. On the computer,  In the input data designation process,  16. A program for executing a process of selecting data having a maximum variance of predicted values of the plurality of hypotheses as the input candidate data is recorded. A computer-readable recording medium according to any one of the above.  17 7. A program that causes a computer to function as an active learning system that executes an active learning method that estimates a functional relationship that at least approximately holds between input and output data based on the selected input and output data. So,  On the computer,  A functional relationship is estimated from a past input / output data using a certain expression, and a plurality of learning algorithms of a predetermined single type or a plurality of types are used as the lower algorithm, and the lower algorithm is used as the lower algorithm. A learning process of learning input / output data up to the point of selecting the next input data as training data;  A function prediction process of predicting a function value for a plurality of input candidate data using a final hypothesis obtained from a plurality of hypotheses obtained from the respective algorithms of the learning process, and a prediction value of the learned plurality of hypotheses. Input data designation processing for selecting input data from the input candidate data based on a value of a function using variance; and A program for executing  1 8. On the computer,  17. The method according to claim 17, further comprising performing a boosting process for improving the learning accuracy of the learning process by a boosting technique for improving the accuracy of a passive learning algorithm by repeating data resampling. Program described in section. 1 9. On the computer, In the learning process and the function prediction process,  The program according to claim 17 or 18, wherein the lower-order algorithm performs a process of performing learning on each of the plurality of input catch data selected at random. 20. On the computer,  In the input data designation process,  20. The method according to claim 17, further comprising: executing a process of selecting, as said input candidate data, data having a maximum variance of predicted values of said plurality of hypotheses. Program.