Information processing systems and programs

Abstract

To provide a technique for compressing data while maintaining characteristics regarding temporal and spatial continuity of the data.SOLUTION: An information processing system for performing reinforcement learning using a learning model including a deep neural network includes: a lattice generation unit which generates a lattice space having a lattice dimension which is determined based on a dimension of the deep neural network; a data dispersion unit which disperses information on original data including spatiotemporal information on lattice points in the lattice space to generate compressed data by compressing the original data; and a learning unit which performs reinforcement learning of the learning model using the compressed data to generate a learned model.SELECTED DRAWING: Figure 7

Description

Technical Field

[0001] The present invention relates to an information processing system, and particularly to an information processing system and a program for performing reinforcement learning using a learning model including a deep neural network.

Background Art

[0002] Deep reinforcement learning is widely used as a method for automatically generating AI that can adapt to various problems.

[0003] Many social problems are multi-agent problems with multiple entities involved. Also, many real-world problems are related to data containing large-scale spatio-temporal information. Therefore, many real-world problems require large-scale calculations to obtain solutions.

[0004] When the dimensionality of the data used in deep reinforcement learning is large, the dimensionality of the neural network for learning it also needs to be large, resulting in a large computational amount and making it difficult to efficiently execute learning.

[0005] In order to efficiently execute deep reinforcement learning, it is necessary to appropriately compress the input data. In particular, in order to solve real-world problems using multi-agent deep reinforcement learning methods, it is necessary to appropriately compress the input data containing large-scale spatio-temporal information.

[0006] As techniques for compressing data, general data compression techniques (such as those using the zip format) and compression techniques suitable for image data (such as those using the jpeg format) are known. Also, in Patent Document 1, a hidden Markov model using a small world network is used as a method for adjusting time-series data used in learning.

Prior Art Documents

Patent Documents

[0007]

Patent Document 1

[0008] However, conventional technologies had the problem of difficulty in maintaining the spatiotemporal continuity of the data.

[0009] For example, general data compression techniques do not reflect features related to spatiotemporal continuity in the compressed data. Furthermore, compression techniques suitable for image data involve information loss, destroying fine-grained features. Thus, applying these techniques to input data for deep reinforcement learning is difficult.

[0010] Furthermore, even with the technology described in Patent Document 1, the spatiotemporal continuity characteristics of the data are not reflected in the compressed data, making it difficult to apply to input data for deep reinforcement learning.

[0011] This invention was made to solve these problems, and aims to provide an information processing system and program that can compress data while maintaining the spatiotemporal continuity characteristics of the data. [Means for solving the problem]

[0012] An example of an information processing system according to the present invention is: An information processing system for performing reinforcement learning using a learning model including a deep neural network, A grid generation unit generates a grid space having grid dimensions determined based on the dimensions of the deep neural network, A data distribution unit generates compressed data by distributing information relating to the original data, including spatiotemporal information, on grid points in the grid space, thereby compressing the original data. A learning unit that performs reinforcement learning on the learning model using the compressed data and generates a trained model, It is equipped with.

[0013] An example of an information processing system according to the present invention is an information processing system that performs inference using a learned model including a deep neural network, a lattice generation unit that generates a lattice space having a lattice dimension determined based on the dimensions of the deep neural network, a data dispersion unit that compresses the original data to generate compressed data by dispersing information related to the original data including spatio-temporal information onto lattice points of the lattice space, an inference unit that performs inference by inputting the compressed data into the learned model and includes.

[0014] An example of a program according to the present invention causes a computer to function as the above-described information processing system.

Effects of the Invention

[0015] According to the information processing system and program according to the present invention, it is possible to compress data while maintaining characteristics regarding the spatio-temporal continuity of the data.

Brief Description of the Drawings

[0016] [Figure 1] Configuration example of an information processing system according to Embodiment 1. [Figure 2] Example of a flowchart representing the learning process of the information processing system according to Embodiment 1. [Figure 3] Example of original data according to Embodiment 1. [Figure 4] Example of a lattice space according to Embodiment 1. [Figure 5] Diagram for explaining the dispersion process according to Embodiment 1. [Figure 6] Example of compressed data obtained as a result of the dispersion process according to Embodiment 1. [Figure 7] Example of the result of performing the dispersion process according to Embodiment 1 for all elements of the original data. [Figure 8] Example of a flowchart representing the inference process of the information processing system according to Embodiment 1. [Figure 9] Example of original data according to Embodiment 2. [Figure 10] Diagram for explaining distributed processing of original data according to Embodiment 2 [Figure 11] Configuration example of information processing system according to Embodiment 3. [Figure 12] Example of lattice space and movement path of agent according to Embodiment 3. [Figure 13] Diagram for explaining relationship between cell and area according to Embodiment 3. [Figure 14] Example of flowchart representing learning process of information processing system according to Embodiment 3. [Figure 15] Diagram for explaining distributed processing according to Embodiment 3.

Mode for Carrying Out the Invention

[0017] Hereinafter, embodiments of the present invention will be described based on the accompanying drawings. [Embodiment 1] FIG. 1 is a configuration example of an information processing system 10 according to Embodiment 1 of the present invention. The information processing system 10 is an information processing system for performing reinforcement learning using a learning model including a deep neural network.

[0018] The information processing system 10 is configured using a computer. The computer has a hardware configuration as a known computer, and includes, for example, arithmetic means 20 and storage means 30. The arithmetic means 20 includes, for example, a processor, and the storage means 30 includes storage media such as a semiconductor memory device and a magnetic disk device. Part or all of the storage media may be non-transitory storage media.

[0019] Also, the computer may include input / output means (not shown). The input / output means includes, for example, input devices such as a keyboard and a mouse, output devices such as a display and a printer, and communication devices such as a network interface.

[0020] The computation means 20 includes a grid generation unit 21, a data distribution unit 22, a learning unit 23, and an inference unit 24. The storage means 30 stores the learning model 31, the trained model 32, and the program 33.

[0021] The learning model 31 is a learning model that includes a deep neural network (not shown). By learning the parameters of the learning model 31, a trained model 32 is generated.

[0022] Program 33 causes the computer to function as an information processing system 10. That is, by the processor executing program 33, the computer functions as the information processing system 10 described in this embodiment.

[0023] Figure 2 is an example flowchart representing the learning process of the information processing system 10. In the learning process, the information processing system 10 first acquires source data to generate training data used for learning (step S1).

[0024] Figure 3 shows an example of the original data D1. The original data D1 contains spatiotemporal information. In the example in Figure 3, the original data D1 is spatiotemporal data represented by a two-dimensional (9x9) array. For example, the vertical direction represents time and the horizontal direction represents space, but this is not limited to this. For example, if there are multiple original data D1s, the vertical and horizontal directions of one original data D1 may both represent space, and the index of the original data D1 may represent time. Furthermore, the original data D1 may be data of three dimensions or more.

[0025] The original data D1 may be multi-agent data. In the example in Figure 3, four of the elements (e.g., scalar quantities) in the original data D1 have a value of "1", while the other elements have a value of "0". The elements with a value of "1" may represent agents in reinforcement learning.

[0026] Next, the grid generation unit 21 generates a grid space having predetermined grid dimensions (step S2). The grid space can be configured to have the same number of dimensions as the original data D1 (2 dimensions in this embodiment). The grid dimensions are configured to be smaller than the dimensions of the original data D1 in each dimension of the grid space. The grid dimensions are determined based on the dimensions of the deep neural network of the learning model 31 and are corresponding to the dimensions of the deep neural network of the learning model 31. For example, they may be the same as the dimensions of the deep neural network. In this embodiment, the dimensions of the input layer of the deep neural network are 3 × 3, and the grid dimensions are also 3 × 3.

[0027] Figure 4 shows an example of a 3x3 grid space. In Figure 4, the grid space is represented overlaid on the original data D1, and the positions of the elements in the original data D1 shown in gray correspond to grid points L11 to L33. In the example in Figure 4, the dimensions of the original data D1 are 9x9, but the dimensions of the grid space are 3x3, meaning that the dimensions of the original data D1 have been reduced in two dimensions.

[0028] Next, the data distribution unit 22 compresses the original data D1 by distributing the information related to the original data D1 on the grid points of the grid space, thereby generating compressed data (step S3).

[0029] The distributed processing in step S3 will be explained using Figure 5. In Figure 5, the distributed processing of the value "1" within the dashed box in Figure 4 is used as an example to explain the distribution processing of the elements contained in the original data D1. For simplicity, values ​​outside the dashed box are omitted from Figure 5 and represented by "·".

[0030] Figure 5(a) shows an example of distance definitions used in distributed processing. Distances are defined along each axis in the grid space of the original data D1. From the value "1", the distance to the upper grid points L11 and L12 along the vertical direction is 1, and the distance to the lower grid points L21 and L22 is 3. Similarly, from this value "1", the distance to the left grid points L11 and L21 along the horizontal direction is 2, and the distance to the right grid points L12 and L22 is 2.

[0031] The data distribution unit 22 distributes the value of each element (the "1" within the dashed box in the example of Figure 5) contained in the original data D1 to each grid point based on the distance between that element and each grid point. Here, it is preferable to distribute the values ​​so that the distributed values ​​decrease as the distance increases. In this embodiment, the values ​​of the elements are distributed inversely proportional to the distance.

[0032] In the example in Figure 5(a), the distance to the upper grid points L11 and L12 is 1, and the distance to the lower grid points L21 and L22 is 3. Therefore, the values ​​distributed to the upper grid points L11 and L12 (more precisely, the sum of the values ​​distributed to these two grid points; the same applies hereafter) and the values ​​distributed to the lower grid points L21 and L22 are distributed in a ratio of 3:1.

[0033] Therefore, the sum of the values ​​distributed to the upper grid points L11 and L12 is 0.75, and the sum of the values ​​distributed to the lower grid points L21 and L22 is 0.25. Similarly, the sum of the values ​​distributed to the left grid points L11 and L21 is 0.5, and the sum of the values ​​distributed to the right grid points L12 and L22 is 0.5.

[0034] When the value "1" is distributed among the four grid points L11 to L22 in a manner that satisfies these conditions, the result is as shown in Figure 5(b). Specifically, 0.375 is distributed to grid point L11, 0.375 to grid point L12, 0.125 to grid point L21, and 0.125 to grid point L22.

[0035] Figure 6 shows an example of compressed data D2 obtained as a result of this distributed processing. In the example in Figure 6, the "1" within the dashed frame in Figure 4 is distributed to each grid point in Figure 4. In this embodiment, the distributed value is set to 0 when the distance is 4 or greater (this applies to grid points L13, L23, L31, L32, and L33 in Figure 4).

[0036] In this way, by distributing the data in a grid space with smaller dimensions than the original data such that the dispersed values ​​decrease as the distance increases, it is possible to reduce the dimensions of the original data D1 and compress it while maintaining the spatiotemporal continuity features of the original data D1.

[0037] Furthermore, the distributed processing according to this embodiment does not involve any loss of information (i.e., the original data D1 can be completely restored from the compressed data D2). Therefore, the characteristics of the original data D1 can be preserved more completely.

[0038] Figure 7 shows an example of the result of performing this distributed processing on all elements of the original data D1. Compressed data D2 is also generated for the other elements whose value is "1" in Figure 4. In this embodiment, distributed processing is not performed on elements whose value is "0", and therefore compressed data D2 is not generated for them. However, compressed data D2 may also be generated for elements whose value is "0" (in which case all elements of the generated compressed data D2 will be 0).

[0039] Here, as shown in Figure 7, multiple compressed data sets D2 are generated from a single original data set D1, and the number of compressed data sets D2 may exceed the number of original data sets D1. However, since the dimensions of the compressed data sets D2 (3x3) are smaller than the dimensions of the original data sets D1 (9x9), by appropriately designing the grid dimensions (i.e., the dimensions of the input layer), the computational load related to the learning process can be reduced. This is one of the effects of compression in this embodiment.

[0040] Returning to Figure 2, after generating compressed data D2 in this manner, the training unit 23 uses the generated compressed data D2 as training data to perform reinforcement training on the training model 31 (step S4). In this way, the trained model 32 is generated.

[0041] While specific examples of reinforcement learning are not provided, it can be applied, for example, to information related to transportation systems. For instance, the source data may include information representing the locations of numerous automobiles in a given area. Furthermore, the information processing system 10 according to this embodiment is not limited to such examples and can be used for any learning process, including known reinforcement learning processes.

[0042] As described above, the information processing system 10 according to this embodiment can compress two-dimensional or more original data representing spatiotemporal information (for example, sparse binary data as shown in Figure 3) into denser data (for example, continuous value data as shown in Figure 6) while maintaining its spatiotemporal continuity and features, thereby reducing the size of the data.

[0043] Furthermore, this streamlines the learning process, dramatically expanding the scale of social problems that can be handled by multi-agent reinforcement learning methods.

[0044] The above describes the learning process of the information processing system 10. The information processing system 10 also functions as an information processing system that performs inference using a trained model 32, which includes a deep neural network. The inference process will be described below.

[0045] Figure 8 is an example flowchart representing the inference process of the information processing system 10. In the inference process, the information processing system 10 first acquires source data to generate test data used for inference (step S11). The format of the source data for the test data can be the same as, for example, the source data D1 of the training data shown in Figure 3.

[0046] The grid generation unit 21 generates a grid space (step S12), and the data distribution unit 22 generates compressed data (step S13). These processes can be performed in the same way as steps S2 and S3 in Figure 2, respectively.

[0047] Next, the generated compressed data D2 is used as test data, and the inference unit 24 performs inference by inputting it into the trained model 32 (step S14). The output from the trained model 32 becomes the inference result.

[0048] Thus, even in inference processing, by distributing the information of the original data across grid points, it is possible to compress the original data by reducing its dimensions while maintaining the spatiotemporal continuity features of the original data.

[0049] In Embodiment 1, the information processing system 10 includes both a learning unit 23 and an inference unit 24, and can perform both learning and inference. As a modified example, the system may include only the learning unit 23 from the learning unit 23 and have only a learning function. Alternatively, the system may include only the inference unit 24 from the learning unit 23 and have only an inference function.

[0050] [Embodiment 2] In Embodiment 1, one agent in the original data was associated with a single element. In Embodiment 2, one agent in the original data is associated with multiple elements. Embodiment 2 will be described below, but parts common to Embodiment 1 may be omitted from the explanation.

[0051] Figure 9 shows an example of the source data D11 according to this embodiment. Figures 9(a) and 9(b) illustrate the same data in different ways. The source data D11 is three-dimensional data, with elements arranged in the X, Y, and Z directions, and has dimensions of 9 × 9 × 3. The Z direction may represent the time direction.

[0052] The original data D11 contains information relating to five agents A through E. Each agent is represented by multiple elements arranged in the Z direction. Figure 9(a) shows the positions of agents A through E in the XY plane and the content of the data relating to each agent. For example, the data relating to agent A has the value "101," meaning that, as shown in Figure 9(b), at agent A's position, "1," "0," and "1" are arranged sequentially in the Z direction, as indicated by the dashed box. The same applies to agents B through E.

[0053] The distributed processing of the original data D11 will be explained using Figure 10. Figure 10(a) shows the original data D11, and Figure 10(b) shows the compressed data D12. The distributed processing of the original data D11 can be performed in the same manner as in Embodiment 1. However, since each agent contains three elements in the Z direction, the number of compressed data D12 elements will be three times the number of agents (i.e., 15 elements). Each element of agent A is shown in Figure 10(a) using a dashed frame, and the correspondence with the compressed data D12 in Figure 10(b) is shown with a solid line. The correspondence for agents B to E is similar, although it is not specifically illustrated.

[0054] As described above, the information processing system according to this embodiment can also be used when each agent in a multi-agent model includes multiple elements.

[0055] Although the inference process in Embodiment 2 will not be described in detail, compressed data is generated by performing the above-described distributed processing on the original test data, and inference can be performed by using the compressed data as test data and inputting it into the trained model 32 by the inference unit 24.

[0056] [Embodiment 3] Embodiment 3 is a modification of Embodiment 1 in which the information processing system 10 performs grouping based on the spatiotemporal distribution of agents. Embodiment 3 will be described below, but parts common to Embodiment 1 may be omitted from the explanation.

[0057] Figure 11 shows an example of the configuration of the information processing system 110 according to Embodiment 3. In addition to the configuration shown in Figure 1, the calculation means 20 of the information processing system 110 further includes a group classification unit 25.

[0058] Figure 12 shows an example of a grid space and agent movement path according to Embodiment 3. In this embodiment, the grid dimensions are 5 × 5. A region enclosed by grid points is called a cell. For example, cell A is enclosed by grid points L11, L12, L21 and L22. Since the grid dimensions are 5 × 5, the grid space contains 4 × 4 = 16 cells.

[0059] Within the data space, areas are defined based on cells. In this embodiment, for each cell, an area centered on that cell is defined.

[0060] Figure 13 is used to explain the relationship between cells and regions. For cell F (the area shown by the diagonal lines), the region centered on cell F is shown as the area within the dashed-dotted frame, i.e., the area shown by the diagonal lines and dots. The region consists of a total of nine cells: the central cell F and the cells adjacent to cell F (for example, cells that share one or more grid points with cell F), namely cells A, B, C, E, G, I, J, and K. Hereafter, this will be denoted as "region ABCEFGIJK".

[0061] In the example in Figure 13, one region contains nine cells, but a region may also contain only eight or fewer cells. For example, the region centered on cell A is region ABEF, which is composed of four cells A, B, E, and F.

[0062] In this embodiment, the source data contains information about multiple agents, and each agent is associated with different elements of the source data over time. That is, each agent moves within the data space of the source data over time. Accordingly, the agents move within the grid space.

[0063] For example, suppose there are five agents, namely Agents 1 through 5. Agents 1 through 5 each move along the movement paths P1 through P5 shown in Figure 12. Movement path P1 starts at cell A, passes through cell A, cell B, cell C, cell B, cell F in that order, and ends at cell F. Movement path P2 passes through cell A, cell E, cell F, cell J, cell I in that order. Movement path P3 passes through cell I, cell M, cell N, cell O, cell N in that order. Movement path P4 passes through cell C, cell D, cell C, cell G, cell C in that order. Movement path P5 is fixed at cell L, that is, it passes through cell L, cell L, cell L, cell L, cell L in that order.

[0064] Figure 14 is an example flowchart showing the learning process of the information processing system 110 according to Embodiment 3. In the learning process, first the information processing system 110 acquires the raw data (step S21), and the grid generation unit 21 generates a grid space (step S22). These processes can be the same as steps S1 and S2 in Figure 2, respectively.

[0065] Next, the group classification unit 25 generates a sequence representing the regions of the grid space that each agent traverses along the time axis (step S23). For example, the first agent traverses cells in the order of cell A → cell B → cell C → cell B → cell F along the movement path P1, so it traverses regions ABEF → region ABCEFG → region BCDFGH → region ABCEFG → region ABCEFGIJK. Thus, in this embodiment, the original data can be said to include information representing the regions of the grid space that each agent traverses along the time axis.

[0066] In this way, the group classification unit 25 generates a sequence representing the areas that each agent passes through in order. Note that the relationship between cells and areas is not limited to that shown in Figure 13, and can be designed as appropriate by those skilled in the art. For example, an area corresponding to a certain cell may include only that cell. Alternatively, it may include not only cells adjacent to the central cell, but also cells adjacent to those cells.

[0067] Next, the group classification unit 25 calculates the similarity between the sequences of each agent (step S24). For example, it calculates the similarity between the sequence corresponding to the travel path P1 generated as described above and the sequence corresponding to the travel path P2 generated in the same way. In this embodiment, the group classification unit 25 calculates the similarity of the spatiotemporal distribution of agents as the similarity of sequences that sequentially represent the areas that each agent passes through. In this way, the similarity is calculated for all agent pairs (pairs of travel paths).

[0068] While specific methods for calculating similarity are not given as examples, those skilled in the art can design methods for calculating similarity appropriately based on prior art. For example, it can be expressed as a series of values ​​or characters that identify a cell or area, and similarity can be calculated based on such a series.

[0069] Next, the group classification unit 25 clusters the agents based on their similarity (step S25). That is, it classifies each agent into one of several clusters. Agents with high similarity between their series belong to the same cluster. Specific clustering methods are not given as examples, but those skilled in the art can design them appropriately based on prior art.

[0070] In this embodiment, clustering is performed based on the similarity of the movement paths of each agent as described above. However, the similarity may be calculated based on other criteria, and the similarity can be calculated based on any index that represents the spatiotemporal distribution of each agent in the original data. In this way, the group classification unit 25 clusters each agent into one of several clusters based on the similarity of the spatiotemporal distribution of each agent.

[0071] Next, the group classification unit 25 groups the agents (step S26). That is, it classifies each agent into one of several groups based on the similarity of the spatiotemporal distribution of each agent in the original data.

[0072] Here, it is preferable to classify agents belonging to the same cluster into different groups. Such grouping is possible by preparing as many groups as there are agents in the largest cluster. Other grouping rules can be arbitrarily designed, as long as multiple agents from the same cluster are not classified into the same group. For example, during the grouping process, the next agent may be preferentially classified into the group with the fewest agents at that time.

[0073] Next, the data distribution unit 22 compresses the original data and generates compressed data by distributing the information related to each group in the original data onto the grid points in the grid space (step S27).

[0074] The distributed processing according to this embodiment will be explained using Figure 15. First, compressed data is calculated for each agent in the same manner as in Embodiment 2. Next, the sum of the compressed data of agents belonging to the same group is calculated. In the example in Figure 15, the first agent and the third agent belong to the first group, so the compressed data for the first group is calculated by summing the compressed data of the first agent and the compressed data of the third agent.

[0075] Although not specifically illustrated, the compressed data for the second group is calculated in a similar manner by summing the compressed data of the second agent, the fourth agent, and the fifth agent.

[0076] In this embodiment, the compressed data for a group is calculated using the "sum," but the function for calculating the compressed data for a group can be arbitrarily designed, and for example, the mean or other statistics may be used.

[0077] Thus, in step S27, the original data is ultimately distributed across groups of agents. In other words, multiple elements corresponding to multiple agents are distributed within the same compressed data, thus reducing the number of compressed data items compared to the case without grouping.

[0078] Returning to Figure 14, after generating compressed data in this manner, the training unit 23 uses the generated compressed data as training data to perform reinforcement training on the training model 31 (step S28). This process can be the same as step S4 in Figure 2. In this way, the trained model 32 is generated.

[0079] In the third embodiment, although the inference process will not be described in detail, compressed data is generated by performing the above-described distributed processing on the original test data, and inference can be performed by using the compressed data as test data and inputting it into the trained model 32 by the inference unit 24.

[0080] As explained above, this embodiment makes it possible to further reduce the number of compressed data items. However, in this embodiment, since multiple elements corresponding to multiple agents are distributed within the same compressed data, there is a possibility that some information from the original data may be lost (in other words, it may not be possible to fully recover the original data from the compressed data).

[0081] However, as mentioned above, by classifying agents belonging to the same cluster (i.e., agents with similar spatiotemporal distributions) into different groups, the likelihood of multiple agents being located close together within the same group is reduced. Therefore, the likelihood of multiple agent values ​​being distributed at the same grid point in the same compressed data is reduced, and the loss of information is kept to a minimum.

[0082] In particular, as in this embodiment, by performing clustering and grouping based on the sequence of areas that each agent passes through, multiple agents passing through similar paths can be classified into different groups, that is, distributed into different compressed data, thereby minimizing the loss of information.

[0083] Thus, according to this embodiment, by considering the spatiotemporal distribution of agents, the dimensions of the original data can be reduced and compressed more efficiently. [Explanation of Symbols]

[0084] 10… Information Processing Systems 20...Calculation means 21...Lattice generator 22...Data distribution unit 23…Learning Department 24… Reasoning part 25…Group Classification Department 30…Memory means 31…Learning Model 32…Trained Model 33…Program 110... Information Processing Systems D1...Original data D2... Compressed data P1...Travel route P2... Travel route P3... Travel Route P4... Travel Route P5... Travel route D11...Original data D12... Compressed data L11…lattice point L12…lattice point L13…lattice point L21…lattice point L22…lattice point

Claims

1. An information processing system for performing reinforcement learning using a learning model including a deep neural network, A grid generation unit generates a grid space having grid dimensions determined based on the dimensions of the deep neural network, A data distribution unit generates compressed data by distributing information relating to the original data, including spatiotemporal information, on grid points in the grid space, thereby compressing the original data. A learning unit that performs reinforcement learning on the learning model using the compressed data and generates a trained model, Equipped with, The aforementioned source data includes information about multiple agents, The information processing system includes a group classification unit that classifies each agent into one of a plurality of groups based on the similarity of the spatiotemporal distribution of each agent in the original data. Information processing system.

2. The information processing system further: The information processing system according to claim 1, further comprising an inference unit that performs inference by inputting the compressed data into the trained model.

3. The information processing system according to claim 1, wherein the data distribution unit generates the compressed data by distributing the information relating to each group in the original data onto the grid points of the grid space.

4. The aforementioned raw data includes information representing the region of the grid space that each agent traverses along the time axis, The information processing system according to claim 1, wherein the group classification unit calculates the similarity as the similarity of a series representing the areas that each agent passes through in order.

5. The aforementioned group classification unit is Based on the similarity of the spatiotemporal distribution of each agent, each agent is clustered into one of a plurality of clusters. Each agent belonging to the same cluster is classified into a different group. The information processing system according to claim 1.

6. The information processing system according to claim 1, wherein the grid space has the same number of dimensions as the original data, and the grid dimensions are smaller than the dimensions of the original data in each dimension.

7. The data distribution unit distributes the values ​​of each element included in the original data to each of the grid points based on the distance between the element and each of the grid points, as described in claim 1.

8. A program that causes a computer to function as the information processing system described in claim 1.

Images