Information processing device, control system, information processing method, and program

Abstract

This technology provides the ability to generate the training data necessary for building learning models to solve combinatorial optimization problems. [Solution] An information processing device comprising a generation means for generating training data for a learning model to solve a combinatorial optimization problem, the training data having a graph structure including nodes and edges.

Description

[Technical Field]

[0001] This disclosure relates to an information processing device, a control system, an information processing method, and a program. [Background technology]

[0002] Numerous combinatorial optimization problems are known, such as the traveling salesperson problem and the knapsack problem. Combinatorial optimization problems are often solved using solvers. However, this method tends to have a huge processing load, especially when the problem is large. In contrast, Patent Document 1 discloses a method for calculating the predicted cost between two nodes selected from a plurality of nodes, based on graph information relating to a connected graph consisting of a plurality of nodes corresponding to points on a map, edges connecting the nodes, and costs attached to each edge. This method involves obtaining the actual cost between two points on the map corresponding to the two selected nodes, generating a learning model that learns the relationship between the predicted cost and the actual cost, and calculating the cost between any two points on the map based on the generated learning model. Furthermore, Patent Document 1 discloses a method for reducing the processing load by using this learning model to determine the travel cost between two points in the traveling salesperson problem.

[0003] By the way, building a learning model using machine learning or deep learning requires a large amount of training data (a set of problems and solutions), but collecting the kind of training data needed for combinatorial optimization problems incurs human, economic, and time costs. For example, the training data for the Traveling Salesperson Problem requires maps, traffic information, delivery records, and so on. [Prior art documents] [Patent Documents]

[0004] [Patent Document 1] Japanese Patent Publication No. 2020-046894 [Overview of the project] [Problems that the invention aims to solve]

[0005] One of the objectives of this disclosure is to provide a method for generating training data necessary for constructing a learning model to solve combinatorial optimization problems. [Means for solving the problem]

[0006] According to one aspect of this disclosure, the information processing apparatus comprises a generation means for generating training data for a learning model to solve a combinatorial optimization problem, the training data having a graph structure including nodes and edges.

[0007] According to one aspect of the present disclosure, the above-mentioned information processing device further comprises: learning means for learning the learning data using a deep learning model and constructing the learning model which serves as a solver for solving the combinatorial optimization problem; and a control system comprising: planning means for planning the patrol route of a mobile body based on the learning model; and means for controlling the mobile body based on the planned route.

[0008] According to one aspect of this disclosure, the computer generates training data for a learning model to solve a combinatorial optimization problem, the training data having a graph structure including nodes and edges.

[0009] According to one aspect of this disclosure, the present invention provides a program that causes a computer to perform a process to generate training data for a learning model for solving a combinatorial optimization problem, the training data having a graph structure including nodes and edges. [Effects of the Invention]

[0010] According to this disclosure, it is possible to generate training data necessary for constructing a learning model for solving combinatorial optimization problems. [Brief explanation of the drawing]

[0011] [Figure 1]This is a block diagram showing an example of a solver creation device according to the first embodiment. [Figure 2] This flowchart shows an example of a method for constructing a learning model according to the first embodiment. [Figure 3] This figure shows an example of a combinatorial optimization problem to which the first embodiment can be applied. [Figure 4] This figure shows an overview of the learning model construction process according to the first embodiment. [Figure 5] This figure illustrates an example of applying the first embodiment to a delivery control system. [Figure 6] This is a block diagram showing an example of an information processing device according to a second embodiment. [Figure 7] This is an example of an information processing method according to the second embodiment. [Figure 8] This figure shows an example of the hardware configuration of the device according to each embodiment. [Modes for carrying out the invention]

[0012] The training data generation method and training model construction method of this disclosure will be described below with reference to the drawings. In the drawings used in the following description, components of parts not related to this disclosure may be omitted from the description and may not be shown. In all drawings, the same reference numerals are used for identical or equivalent components, and common descriptions may be omitted.

[0013] <First Embodiment> (composition) FIG. 1 is a block diagram showing an example of a solver creation device according to a first embodiment. The solver creation device 10 generates learning data necessary for constructing a learning model for solving a combinatorial optimization problem, learns the generated learning data by a deep learning model, and constructs a learning model (solver) that can solve the combinatorial optimization problem. In the present disclosure, this learning model is also referred to as a neural solver. Examples of combinatorial optimization problems include TSP (traveling salesman problem), VRP (Vehicle Routing Problem), FSSP (flow shop scheduling problem), JSSP (Job shop Scheduling Problem), FairDivision, and the like. Hereinafter, the functions of the solver creation device 10 will be described by taking the case of TSP as an example.

[0014] The solver creation device 10 includes a learning data generation unit 11, a learning unit 12, and an output unit 13.

[0015] The learning data generation unit 11 generates learning data. The learning data for TSP consists of the coordinates of each location and the movement cost (distance or time required for movement) between locations. The movement cost has properties such as symmetric or asymmetric, and whether it satisfies the triangle inequality (metric) or not (non-metric). That the movement cost between point A and point B is symmetric means that the movement cost when moving from point A to point B is equal to the movement cost when moving from point B to point A, and asymmetric means that these movement costs do not match. A metric means that when considering points A to C, the relationship that the sum of the movement cost from point A to point B and the movement cost from point B to point C is greater than the movement cost from point A to point C holds, and non-metric means that this relationship does not hold. In the real world, the movement cost between locations is generally asymmetric and metric. In order to generate learning data (coordinates and movement costs) that satisfy the properties of being asymmetric and metric, it is often generated by the method of "randomly generating the coordinates of multiple locations and the movement costs between locations and modifying the movement costs to satisfy the properties of being asymmetric and metric", but in this method, when the number of locations is n, a computational complexity of n cubed is required, and as the number of locations n increases, it becomes unrealistic to generate learning data by this method. In contrast, in this embodiment, learning data (coordinates and movement costs) is generated in the following steps 1 to step 5.

[0016] [Method for generating learning data] (Step 1) Generate symmetric and metric movement costs. For example, determine the coordinates of multiple locations by generating random numbers, and calculate the distances between two points among the multiple locations respectively. Let the movement cost between location i and location j be S ij and normalize it so that the maximum value of the movement cost becomes 1.

[0017] (Step 2) Arbitrarily determine the value of γ from the range of 0 < γ ≤ 1 / 2.

[0018] (Step 3) For one movement cost, select a value from the range of a uniform distribution U[γ, 2γ] by a random number and use it as the asymmetric component. Let the asymmetric component between location i and location j be Aij Let it be so. For each movement cost, obtain the asymmetric component.

[0019] (Step 4) Add the components of (Step 1) and (Step 3) to each movement cost. For example, let the movement cost from point i to point j be S ij +A ij Let it be so. For the movement cost from point j to point i, it may be S ij Let it be so, or calculate another asymmetric component A ji in Step 3, and let it be S ij +A ji Let it be so.

[0020] (Step 5) Normalize each movement cost. For example, let the movement cost from point i to point j be (S ij +A ij )÷(1 + 2γ).

[0021] By repeating Steps 1 to 5 as necessary, learning data can be efficiently generated with a computational complexity of about the square of the number of points n. Specifically, for each of the n points, it is only necessary to perform the process of calculating the movement costs between that point and the other n - 1 points, and an asymmetric and metric movement cost can be generated without checking and correcting whether each movement cost is a metric. Although omitted in this disclosure, it has been proven by the inventor that the movement costs generated by the method of Steps 1 to 5 are metrics.

[0022] The learning unit 12 learns from the learning data generation unit 11 and constructs a learning model (neural solver) capable of solving combinatorial optimization problems (e.g., TSP). A deep learning model such as a transformer can be used for learning. A transformer is a neural network architecture used in LLMs (Large-Scale Language Models). By providing the transformer with learning data and instructing it to learn a route (a way to traverse n locations) that minimizes the objective function, for example, the travel cost when visiting all locations once, the transformer constructs a learning model that searches for such a route, that is, a learning model (neural solver) capable of solving TSP. By using a transformer, it is not necessary to include ground truth data in the learning data; the Transformer itself searches for the correct route by simply providing location information (locations and travel costs). This eliminates the need to prepare ground truth data (paths with the minimum cost) as learning data, and furthermore, since location information can be generated using the methods described in steps 1 to 5 above, the processing load required for generating learning data can be reduced. It is already known that a learning model for finding solutions can be obtained by applying a transformer to a TSP or similar object. The learning unit 12 stores the constructed learning model.

[0023] The output unit 13 outputs the learning model constructed by the learning unit 12 to other devices. The learning model, constructed by learning artificially generated learning data by the learning data generation unit 11, can be used in a general-purpose manner. For example, if domestic map information is provided to a learning model constructed for TSP and any point is specified, the learning model will output the route that visits the specified point at the lowest cost.

[0024] (operation) Figure 2 is a flowchart showing an example of a method for constructing a learning model according to the first embodiment. The learning data generation unit 11 generates learning data (coordinates of multiple locations and travel costs between each location) by executing the processes described in steps 1 to 5 above (step S11). The learning data generation unit 11 outputs the generated learning data to the learning unit 12. Next, the learning unit 12 constructs a learning model (neural solver) (step S12). The user sets the objective function to the learning unit 12 and instructs it to learn the learning data and construct a learning model that optimizes the objective function. The learning unit 12 constructs a learning model that uses a transformer to find the lowest cost route to visit each location. Next, the output unit 13 outputs the learning model constructed in step S12 to another device (step S13). This other device may be, for example, a device that performs delivery planning.

[0025] Figure 3 shows an example of a combinatorial optimization problem to which the data generation method and learning model construction method according to this embodiment can be applied. Item 1 in the table in Figure 3 is TSP and VRP. This is as explained using Figures 1 and 2. A large amount of asymmetric metric data is generated, with the points to be visited as nodes and the edges between nodes as travel costs, to create learning data in a graph structure. Then, the transformer is instructed to learn this learning data and search for a path that minimizes the travel cost, thereby constructing a learning model capable of solving TSP and VRP. In the case of using multiple vehicles in VRP, the travel cost when multiple vehicles divide the task of visiting each point is set as the objective function, and by learning a route that minimizes this objective function, a learning model that can handle multiple vehicles can be constructed.

[0026] Item 2 in the table in Figure 3 is FSSP. When multiple jobs are completed, each processed by multiple machines, for example, the goal is to find the machine operation schedule (the order in which each machine processes work) that minimizes the operating time of each machine. In this case, training data is generated in the form of a graph structure where the processing time for each job and machine is represented by nodes, the difference in processing time with the previous job in the flow order is represented by edges between nodes, and the total processing time of the last job is represented by edges. The transformer is then instructed to learn from this training data and search for the order in which jobs are assigned to each machine (flow order) that minimizes the total processing time, thereby constructing a learning model capable of solving FSSP.

[0027] Item 3 in the table in Figure 3 is FairDivision. FairDivision is a problem such as fairly distributing property to heirs, fairly distributing relief supplies and services to disaster victims, or equally assigning work to employees. In this case, a learning model capable of solving the FairDivision problem is constructed by generating training data in a graph structure where the utility of each player (a unit of property, relief supplies, work) is represented by nodes, and the utility of each player is represented by edges (if good a is a node, the edges connected to the node of good a represent the utility when player A obtains good a, the utility when player B obtains good a, etc.). The transformer is instructed to learn from this training data to find a distribution method that maximizes the utility (nodes) obtained by each player and does not create a sense of unfairness even when exchanged with other players.

[0028] What these methods have in common is that they generate training data in the form of a graph structure (only the problem, no correct answer is needed), and then train the transformer to learn how to traverse nodes and how to distribute nodes (how to partition the graph) in that graph structure that satisfies a certain objective function. As shown in Figure 4, a solver (neural solver) for the combinatorial optimization problem is obtained by generating training data in the form of a graph structure (the training data corresponds to the "problem" of the combinatorial optimization problem), and creating a learning model that can efficiently learn and solve the problem using the transformer. At this time, the focus is on generating training data for creating a learning model that can solve the combinatorial optimization problem. In the case of TSP and VRP, the training data for creating a learning model that can solve the problem is training data in which the movement cost is asymmetric and has the property of a metric. By devising the process of generating training data, for example, as in steps 1 to 5 above, a large amount of training data can be generated by a relatively simple process with little computational load. In this embodiment, for each combinatorial optimization problem, the type of training data needed to create a solvable learning model is analyzed, and a method (for example, steps 1-5) is developed and devised to obtain training data suitable for that optimization problem. Then, data is generated using the developed and devised method, and this is used as training data in the form of a graph structure with nodes and edges. If training data in the form of a graph structure can be generated, a learning model for solving can be created by having the transformer learn how to traverse and distribute nodes to achieve the objective. In other words, if a method for generating training data suitable for combinatorial optimization problems can be developed, it is expected that solvers for any combinatorial optimization problem can be created. That is, the method for generating training data and constructing a learning model according to this embodiment is a systematic methodology for solving combinatorial optimization problems.

[0029] (Example of use) Next, an example of planning a delivery plan and controlling a delivery plan based on a learning model for VRP constructed using the method of this embodiment will be described with reference to Figure 5. The delivery control system 100 includes a solver creation device 10, a delivery planning device 20, a vehicle control device 30, and vehicles 40a, 40b, ... The solver creation device 10 generates learning data using the method of steps 1 to 5, learns the learning data using transformer, and constructs a learning model for VRP that enables multiple vehicles 40a, 40b to divide the work and deliver goods at the lowest cost. The solver creation device 10 outputs the constructed learning model to the delivery planning device 20. The delivery planning device 20 applies a real-world delivery planning problem to the learning model obtained from the solver creation device 10 and solves it. For example, the delivery planning device 20 is given map information including the delivery route to the destination and the number of vehicles that can be used for delivery as calculation conditions, and has a learning model solve a delivery planning problem (VRP) that determines how to deliver goods to multiple destinations using multiple vehicles 40a, 40b, ... to minimize travel costs. By solving the VRP, a solution is obtained in which vehicle 40a visits point 1, point 2, point 3, ..., vehicle 40b visits point 4, point 5, point 6, ..., ..., and so on. The delivery planning device 20 outputs the obtained solution to the vehicle control device 30. Vehicles 40a, 40b, ... are vehicles capable of autonomous or remote driving, and the vehicle control device 30 can control vehicles 40a, 40b, .... The vehicle control device 30 controls vehicles 40a, 40b, ... based on the information of the delivery destinations that each vehicle should visit, obtained from the delivery planning device 20. For example, the vehicle control device 30 sets points 1, 2, 3, etc. as delivery destinations for vehicle 40a and instructs vehicle 40a to automatically patrol in the order of points 1, 2, 3, etc. Similarly, the vehicle control device 30 sets points 4, 5, 6, etc. as delivery destinations for vehicle 40b and instructs vehicle 40a to automatically patrol in the order of points 4, 5, 6, etc. Following these instructions, vehicle 40a patrols points 1, 2, 3, etc. and delivers the goods, and vehicle 40b patrols points 4, 5, 6, etc. and delivers the goods.The learning model created by this embodiment can be used in actual control applications.

[0030] <Second Embodiment> Figure 6 is a block diagram showing an example of an information processing device according to the second embodiment. The information processing device 800 includes a generation means 801 that generates training data for a learning model to solve combinatorial optimization problems, the training data having a graph structure including nodes and edges. The training data generation unit 11 of the solver creation device 10 is an example of the generation means 801.

[0031] Figure 7 is a flowchart showing an example of an information processing method according to the second embodiment. The computer generates training data for a learning model to solve a combinatorial optimization problem, the training data having a graph structure including nodes and edges (step S801).

[0032] Figure 8 shows an example of the hardware configuration of the device according to each embodiment. The computer 900 comprises a CPU 901, main memory 902, auxiliary memory 903, input / output interface 904, and communication interface 905. The solver creation device 10 and information processing device 800 described above are implemented in the computer 900. The functions described above are stored in the auxiliary memory 903 in the form of programs. The CPU 901 reads the program from the auxiliary memory 903, expands it into the main memory 92, and executes the above processing according to the program. The CPU 901 also allocates memory space in the main memory 902 according to the program. The CPU 901 also allocates memory space in the auxiliary memory 903 to store data being processed according to the program.

[0033] A program for realizing all or part of the functions of the solver creation device 10 and the information processing device 800 may be recorded on a computer-readable recording medium, and the program recorded on this recording medium may be loaded into a computer system and executed to perform processing by each functional unit. Here, "computer system" includes hardware such as the OS and peripheral devices. Furthermore, if a WWW system is used, "computer system" also includes the homepage provisioning environment (or display environment). Furthermore, "computer-readable recording medium" refers to portable media such as CDs, DVDs, USBs, and storage devices such as hard disks built into the computer system. Furthermore, if this program is distributed to the computer 900 via a communication line, the computer 900 that receives the program may load it into the main memory 902 and execute the above processing. Furthermore, the above program may be for realizing only a part of the functions described above, and may also be for realizing the above functions in combination with programs already recorded in the computer system.

[0034] Although one embodiment of this disclosure has been described in detail above with reference to the drawings, the specific configuration is not limited to that described above, and various design changes can be made without departing from the spirit of this invention. Furthermore, one aspect of this disclosure can be modified in various ways within the scope of the claims, and embodiments obtained by appropriately combining the technical means disclosed in different embodiments are also included in the technical scope of this disclosure. In addition, configurations in which elements described in each of the above embodiments and modifications that produce similar effects are substituted for each other are also included. Moreover, each embodiment can be appropriately combined with other embodiments.

[0035] Some or all of the above embodiments may be described as follows, but are not limited to the following:

[0036] (Note 1) Training data for a learning model to solve a combinatorial optimization problem, wherein the training data has a graph structure including nodes and edges, and a generation means for generating the training data. An information processing device equipped with the following features.

[0037] (Note 2) The information processing apparatus described in Appendix 1, wherein the generation means normalizes the distance between two predetermined points, determines an arbitrary value γ from a range greater than 0 and less than or equal to 1 / 2, selects an arbitrary value from a range greater than or equal to γ ​​and less than or equal to γ ​​× 2, adds the selected value to the normalized distance, calculates the movement cost between the two points by normalizing the added value, and generates the learning data in which the two points are the nodes and the movement cost is the edge.

[0038] (Note 3) The information processing apparatus described in Appendix 2, wherein the generation means generates the coordinates of each of the two predetermined points and calculates the distance between the two points based on the generated coordinates.

[0039] (Note 4) The information processing device according to Appendix 2 or Appendix 3, wherein the generation means randomly generates the coordinates of a plurality of points, selects two of the generated plurality of points, calculates the distance between the two selected points, and normalizes the calculated distance.

[0040] (Note 5) The information processing apparatus according to appendices 1 to 4, further comprising: a learning means for learning the aforementioned training data using a deep learning model and constructing the training model which serves as a solver for solving the aforementioned combinatorial optimization problem.

[0041] (Note 6) The deep learning model is a transformer, as described in Appendix 5.

[0042] (Note 7) The information processing device according to Appendix 5 or Appendix 6, wherein the learning means constructs the learning model which serves as a solver for the traveling salesman problem or the vehicle routing problem by learning a path that minimizes the cost of visiting the nodes included in the graph structure based on the learning data.

[0043] (Note 8) A control system comprising an information processing device described in any of Appendix 3 to Appendix 7, a planning means for planning a mobile body's patrol route based on the learning model, and a means for controlling the mobile body based on the planned route.

[0044] (Note 9) An information processing method for a computer to generate training data for a learning model to solve a combinatorial optimization problem, the training data having a graph structure including nodes and edges.

[0045] (Note 10) A program that causes a computer to perform a process to generate training data for a learning model to solve combinatorial optimization problems, the training data having a graph structure including nodes and edges. [Explanation of symbols]

[0046] 10. Solver creation device 11. Learning Data Generation Unit 12. Learning Department Output section of 13... 20. Delivery planning device 30. Vehicle control device 40a, 40b... Vehicles 100...Delivery control system 800... Information Processing Equipment 801...Generation means 900... Computer 901···CPU 902...Main memory 903...Auxiliary storage device 904... Input / Output Interface 905...Communication Interface

Claims

1. Training data for a learning model to solve a combinatorial optimization problem, wherein the training data has a graph structure including nodes and edges, and a generation means for generating such training data. An information processing device equipped with the following features.

2. The generating means is The distance between two predetermined points is normalized, an arbitrary value γ is determined from a range greater than 0 and less than or equal to 1 / 2, an arbitrary value is selected from a range greater than or equal to γ ​​and less than or equal to γ ​​× 2, the selected value is added to the normalized distance, and the movement cost of the two points is normalized to calculate the movement cost of the two points, and the learning data is generated in which the two points are the nodes and the movement cost is the edge. The information processing apparatus according to claim 1.

3. The generation means generates coordinates for the two predetermined points and calculates the distance between the two points based on the generated coordinates. The information processing apparatus according to claim 2.

4. The generation means randomly generates coordinates of multiple points, selects two of the generated points, calculates the distance between the two selected points, and normalizes the calculated distance. The information processing apparatus according to claim 2.

5. A learning means that learns the aforementioned training data using a deep learning model and constructs the learning model which serves as a solver for solving the aforementioned combinatorial optimization problem. The information processing apparatus according to claim 1 or claim 2, further comprising:

6. The aforementioned deep learning model is a transformer. The information processing apparatus according to claim 5.

7. The learning means constructs a learning model that serves as a solver for the traveling salesman problem or the vehicle routing problem by learning a path that minimizes the cost of visiting the nodes included in the graph structure, based on the learning data. The information processing apparatus according to claim 5.

8. The information processing apparatus according to claim 7, A planning means for planning the patrol route of a mobile object based on the aforementioned learning model, Means for controlling the moving body based on the planned schedule, A control system equipped with the following features.

9. Computers Training data for a learning model to solve a combinatorial optimization problem, wherein the training data has a graph structure including nodes and edges, Information processing methods.

10. On the computer, A process for generating training data for a learning model to solve a combinatorial optimization problem, the training data having a graph structure including nodes and edges. A program that executes the command.

Images