Search device, search method, and computer program

Abstract

This method efficiently obtains the optimal or approximate solution to combinatorial optimization problems using large-scale language models. [Solution] The search device comprises: a first acquisition unit that acquires information related to the problem setting of a combinatorial optimization problem as natural language; a second acquisition unit that acquires an image of a directed graph composed of nodes and edges, in which the orientation of the edges connecting the nodes is set, and update instructions for the image of the directed graph in an iterative process for searching for the optimal solution; a search unit that inputs the natural language acquired by the first acquisition unit, the image of the directed graph and the update instructions acquired by the second acquisition unit into a large-scale language model, and searches for the optimal solution of the problem setting expressed in the input natural language using the orientation of the edges included in the directed graph and an iterative process accompanied by update instructions; and an output unit that outputs at least one of the optimal solution obtained by the search unit and an approximate solution of the optimal solution.

Description

Technical Field

[0001] The present invention relates to a search device, a search method, and a computer program.

Background Art

[0002] In various problems of production and logistics in factories, combinatorial optimization problems frequently occur. A method using a large language model (LLM; Large Language Model) to solve such combinatorial optimization problems is known (for example, see Non-Patent Document 1). Approaches to combinatorial optimization problems can be considered in terms of formulating the problem setting as a system of linear inequalities, implementing source code according to the grammar of a solver, and improving the algorithm for the purpose of accelerating the calculation speed. Specialized knowledge is required to perform this formulation stage, implementation stage, and acceleration stage. Therefore, it is difficult to say that the task of obtaining a solution to a combinatorial optimization problem can be easily carried out by anyone.

Prior Art Documents

Non-Patent Documents

[0003]

Non-Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] The approach to solving combinatorial optimization problems can be understood as the process of inputting a given problem setting into a computer (translating it into a language the computer can understand) using systems of inequalities and programming languages. The method described in Non-Patent Literature 1 is a strategy that sequentially obtains better solutions by iteratively providing generation instructions using only natural language. However, since large-scale language models can accept inputs other than natural language, there is room to make it easier to search for the optimal solution or an approximate solution to the optimal solution of a combinatorial optimization problem by using inputs other than natural language. On the other hand, if the language input to the large-scale language model is not considered, it may be more difficult to search for the optimal solution or an approximate solution than with generation instructions using only natural language.

[0005] This invention was made to solve at least some of the problems described above, and aims to efficiently obtain the optimal or approximate solution to a combinatorial optimization problem using a large-scale language model. [Means for solving the problem]

[0006] The present invention has been made to solve at least some of the above-mentioned problems and can be realized in the following forms.

[0007] (1) According to one embodiment of the present invention, a search device for searching for the optimal solution of a combinatorial optimization problem is provided. The search device comprises: a first acquisition unit that acquires information related to the problem setting of the combinatorial optimization problem as natural language; a second acquisition unit that acquires an image of a directed graph composed of nodes and edges, in which the orientation of edges connecting the nodes is set, and an update instruction for the image of the directed graph in an iterative process for searching for the optimal solution; a search unit that inputs the natural language acquired by the first acquisition unit, the image of the directed graph acquired by the second acquisition unit, and the update instruction into a large-scale language model, and searches for the optimal solution of the problem setting expressed by the input natural language using the orientation of edges included in the directed graph and an iterative process accompanied by the update instruction; and an output unit that outputs at least one of the optimal solution obtained by the search of the search unit and an approximate solution of the optimal solution.

[0008] In this configuration, information about the problem setting is obtained as natural language, along with an image of a directed graph with edge orientations set, and instructions for updating the image during the iterative process to find the optimal solution. In other words, the directed graph is obtained as an image distinct from natural language. There are combinatorial optimization problems where the edge orientations are not constrained. In contrast, this configuration uses an image of a directed graph with edge orientations constrained by an image distinct from natural language, and a large-scale language model searches for the optimal solution using the order information obtained from the edge orientations. As a result, by constraining the order information, the number of iterations performed by the search of the large-scale language model is reduced, and the optimal or approximate solution is derived. In other words, with this configuration, the optimal or approximate solution to a combinatorial optimization problem can be efficiently obtained using a large-scale language model.

[0009] (2) In the search apparatus according to the above embodiment, the first acquisition unit may acquire instructions to generate a plurality of solutions in each of the iterative processes, and the second acquisition unit may acquire, as the update instruction, an instruction to create an image of the directed graph of the approximate solution that is closest to the optimal solution among the plurality of solutions generated in each of the iterative processes. In this configuration, a directed graph is created for the approximate solution closest to the optimal solution among the multiple solutions generated during the iterative process, and directed graphs are not created for the other solutions. Therefore, in the next process, multiple solutions are searched for from the directed graph of the approximate solution closest to the optimal solution. As a result, directed graphs are not created for solutions that are not close to the optimal solution, making the search for the optimal solution even more efficient.

[0010] (3) In the search device according to the above embodiment, the first acquisition unit acquires at least one arbitrary solution of the combinatorial optimization problem as an initial value, and the second acquisition unit does not need to acquire the image of the directed graph as the initial value. With this configuration, only one solution is obtained as an initial value, and it is not necessary to obtain a directed graph. Therefore, users can search for the optimal and approximate solutions to combinatorial optimization problems simply by providing a simple initial value.

[0011] Furthermore, the present invention can be realized in various forms, for example, as a search device, a support device, a generation device, a search method, a support method, a generation method, and a system comprising these devices or implementing these methods, a computer program for executing these devices or methods, a server device for distributing this computer program, a non-temporary storage medium storing the computer program, and so on. [Brief explanation of the drawing]

[0012] [Figure 1] This is a schematic block diagram of a search device as an embodiment of the present invention. [Figure 2] This is an explanatory diagram of the coordinate values ​​of the 48 cities in the Traveling Salesperson Problem. [Figure 3] This is an explanatory diagram of the prompts obtained as natural language by the first and second acquisition units. [Figure 4] This is an explanatory diagram of the directed graph image acquired by the second acquisition unit. [Figure 5] This is an explanatory diagram of the graph image representing the optimal solution to the Traveling Salesperson Problem. [Figure 6] This is a flowchart of the search method in this embodiment. [Figure 7] This is an explanatory diagram of the undirected graph image in the comparative example. [Figure 8] This is an explanatory diagram of the results of the search for examples and comparative examples. [Figure 9] This is an explanatory diagram of the results of the search for examples and comparative examples. [Figure 10] This is a schematic block diagram of the search system in the second embodiment. [Modes for carrying out the invention]

[0013] <First Embodiment> FIG. 1 is a schematic block diagram of a search device 100 as an embodiment of the present invention. The search device 100 of the present embodiment reduces the formulation stage of the combinatorial optimization problem setting and the implementation stage of the source code in the solving solver by inputting natural language and an image of a graph composed of nodes and edges into a large language model.

[0014] As shown in FIG. 1, the search device 100 includes a control device 10, a storage device 20 that stores various data, an input unit 30 that receives user input, and an output unit 40 that outputs images, sounds, and the like. The storage device 20 is composed of a hard disk drive (HDD), etc. The storage device 20 includes a model database (model DB) 21 that stores a large language model. As the large language model, Gemini-1.5-flash (where "Gemini" is a registered trademark) or the like may be used, or an individually developed model may be used. The input unit 30 includes input devices such as a keyboard, a mouse, and a microphone, for example. The output unit 40 includes output devices such as a monitor that outputs various images and a speaker that outputs sound, for example. In the present embodiment, the monitor as the output unit 40 outputs the optimal solution or an approximate solution of the optimal solution obtained by the search of the search unit 13 by means of numerical values or an image of a graph.

[0015] The control device 10 is composed of a computer. The control device 10 includes a CPU (Central Processing Unit) not shown. The CPU is connected to a ROM (Read Only Memory) and a RAM (Random Access Memory) not shown, and controls each part of the control device 10 by expanding and executing the computer program stored in the ROM in the RAM. In addition, as shown in FIG. 1, the CPU also functions as a first acquisition unit 11, a second acquisition unit 12, and a search unit 13.

[0016] The first acquisition unit 11 acquires information related to the setting of the set combinatorial optimization problem as natural language. In the present embodiment, the traveling salesman problem called "att48" for 48 cities in the United States is handled. In the traveling salesman problem of "att48", in a single stroke path that travels through 48 cities in the United States, the problem is to search for the path with the shortest path length as an optimization.

[0017] FIG. 2 is an explanatory diagram of the coordinate values of 48 cities in the traveling salesman problem of att48. In FIG. 2, the coordinate values in two dimensions of 48 cities handled by att48 are shown as a list. The coordinate values of 48 cities are used for searching for the optimal solution by the search unit 13 described later.

[0018] The second acquisition unit 12 shown in FIG. 1 acquires an image of a directed graph composed of nodes and edges, with the direction of the edges connecting the nodes set, as information related to the combinatorial optimization problem. Also, in the iterative process for searching for the optimal solution, the second acquisition unit 12 acquires an update instruction for the image of the directed graph. The creation instruction for the directed graph and the image of the directed graph acquired by the second acquisition unit 12 will be described together with the prompt generated by the search unit 13. In the present embodiment, the first acquisition unit 11 acquires natural language from the data input by the user via the input unit 30, and the second acquisition unit 12 acquires the creation instruction for the directed graph and an image of an example of the directed graph.

[0019] The search unit 13 inputs natural language acquired by the first acquisition unit 11, instructions for creating a directed graph acquired by the second acquisition unit 12, and an image of an example of a directed graph into a large-scale language model. The search unit 13 searches for the optimal solution to the ATT48 Traveling Salesperson Problem, which is expressed in natural language as a combinatorial optimization problem, using the orientation of edges included in the directed graph and an iterative process accompanied by instructions for creating a directed graph. In this embodiment, the natural language information related to the combinatorial optimization problem acquired by the first acquisition unit 11 and the instructions for updating the directed graph are prompts created using the Python 3.12 programming language ("Python" is a registered trademark). If Gemini-1.5-flash, an example of a large-scale language model described in this embodiment, is used, when the desired prompt and graph image are sent to Python 3.12, a solution corresponding to the prompt is obtained after a certain period of time.

[0020] Figure 3 is an explanatory diagram of the prompts as natural language acquired by the first acquisition unit 11 and the second acquisition unit 12. The prompts shown in Figure 3 consist of a first prompt P1, a second prompt P2, a third prompt P3, and a fourth prompt P4. The first prompt P1 is a prompt for the coordinate values ​​of the 48 cities in the traveling salesman problem of att48 shown in Figure 2. The second prompt P2 is a prompt for the path information of the 10 types of solutions obtained in the previous process when the search unit 13 iteratively searches for solutions to the traveling salesman problem. In other words, the first acquisition unit 11 acquires instructions to create 10 types of solutions in each iterative search process as information related to the setting of the traveling salesman problem. In this embodiment, since the search unit 13 generates 10 types of solutions in one process, the second prompt P2 represents the solution obtained in the previous process.

[0021] The third prompt P3 is a prompt that instructs the system to search for a solution with a shorter path than the solution generated in the process immediately preceding the iterative search. In other words, the second acquisition unit 12 acquires the third prompt P3 as an instruction to create a directed graph of the solution closest to the optimal solution from among the 10 types of solutions created in each iterative search process. The fourth prompt P4 is a description of an image representing the solution with the shortest path length from among the 10 types of solutions in the process immediately preceding the second prompt P2.

[0022] Figure 4 is an explanatory diagram of the directed graph image acquired by the second acquisition unit 12. Figure 5 is an explanatory diagram of the graph image of the optimal solution to the Traveling Salesperson Problem for att48. Figure 4 shows 48 cities represented by black circles as 48 nodes, straight lines as edges connecting two of the 48 cities, and arrows DR1 representing the relationship between the start and end points for each edge. Figure 4 shows an image representing all 48 cities and a magnified view of the vicinity of edge ED1 connecting city 38 and city 31. As shown in the magnified view of Figure 4, arrow DR1 indicates that the start point of edge ED1 is city 38 and the end point of edge ED1 is city 31. Figure 5 shows the graph image of the optimal solution to the Traveling Salesperson Problem for att48, which has been determined in advance.

[0023] When the search unit 13 receives the prompt shown in Figure 3 and the instruction to create a graph image shown in Figure 4, it uses the prompt shown in Figure 3 to generate an initial prompt as an initial value. The search unit 13 generates an initial prompt consisting of the first prompt P1 in Figure 3, the third prompt P3, and the path information of one randomly generated solution. In other words, the initial prompt does not include the graph image shown in Figure 4 or the third prompt P3, which is a description of the graph image.

[0024] The search unit 13 uses the prompt shown in Figure 3, the directed graph image shown in Figure 4, and the generated initial prompt to search for the optimal path to solve the Traveling Salesperson Problem of att48. In this embodiment, the number of search operations is input via the input unit 30, and the search unit 13 searches for a solution until the number of search operations is reached. In other embodiments, the search condition may be given as search time instead of the number of search operations.

[0025] Figure 6 is a flowchart of the search method in this embodiment. In the search flow shown in Figure 6, first, the first acquisition unit 11 performs a first acquisition step (step S1) in which it acquires information related to the setting of the combinatorial optimization problem in natural language. The first acquisition unit 11 acquires information related to the problem setting, specifically the information of the first prompt P1 and the second prompt P2 shown in Figure 3.

[0026] Next, the second acquisition unit 12 performs a second acquisition step (step S2) to acquire update instructions for the directed graph in the iterative process for solving the Traveling Salesperson Problem. The second acquisition unit 12 acquires the third prompt P3 and the fourth prompt P4 shown in Figure 3 as update instructions for the directed graph. The search unit 13 creates an initial prompt using the prompts shown in Figure 3 (step S3). The search unit 13 generates an initial prompt consisting of the first prompt P1, the third prompt P3, and path information for multiple solutions created randomly.

[0027] The search unit 13 performs an iterative search process using the five prompts P1-P5 shown in Figure 3 on the generated initial prompt (step S4). The search unit 13 determines whether or not to terminate the search process (step S5). In this embodiment, the search unit 13 determines whether the total number of search processes has reached the threshold number of search processes input via the input unit 30. If it is determined that the number of search processes has not reached the threshold (step S5: NO), the search process continues (step S4). If it is determined that the number of search processes has reached the threshold (step S5: YES), the output unit 40 performs an output process to output the search process results (step S6), and the search flow ends. The output process can also be described as a process in which the large-scale language model outputs the optimal solution or approximate solution as a search result obtained using the arrow DR1 of edge ED1 included in the acquired directed graph and the iterative search process which is an update instruction.

[0028] The shortest path length of the searched solution was evaluated according to the number of search processes in the example, which is the search result by the search device 100 of the above embodiment, and in two comparative examples 1 and 2. Comparative example 1 is the search result when the same large-scale language model as in the example is input with a first prompt P1 for the coordinate values ​​of 48 cities shown in Figure 3, a second prompt P2 for path information of 10 types of solutions, and a third prompt P3 that searches for a solution with a shorter path than the solution generated in the previous process. In other words, in comparative example 1, compared to the example, the image of the directed graph shown in Figure 4 and the fourth prompt P4, which is a description of the directed graph, are not input.

[0029] Comparative Example 2 differs from Example 1 in that, instead of a directed graph, an image of an undirected graph is generated in which the relationship between the start and end points of the edge connecting the two cities is not represented by arrows. Note that the fourth prompt P4, which describes the directed graph, is the same in Example 1 and Comparative Example 2. Figure 7 is an explanatory diagram of the undirected graph image in Comparative Example 2. In the undirected graph image shown in Figure 7, unlike the directed graph image shown in Figure 4, arrows DR1, etc., of edge ED1 are not shown. Here, in the fourth prompt P4 shown in Figure 3...<trace> from< / trace> Between them are instructions to arrange the routes between each city in order. Therefore, in Comparative Example 2, although no arrows are shown in the image of the undirected graph, it is implicitly indicated that the orientation of the edges of the 48 cities should be considered.

[0030] Figure 8 is an explanatory diagram of the search results for the Examples and Comparative Examples 1 and 2. In Figure 8, for each of the Examples and Comparative Examples 1 and 2, when the search process was performed 50 times, the path lengths of the approximate solutions are shown among the 10 types of path information searched in one process. In addition, the path with the shortest path length, which is the optimal solution, is shown as a dashed line in Figure 8.

[0031] In Figure 8, the shortest path length for the Example is shown by a diamond, Comparative Example 1 by a circle, and Comparative Example 2 by a triangle. As shown in Figure 8, when the number of searches exceeds approximately 20, the path length of the Example is the smallest among the three: the Example and Comparative Examples 1 and 2. In other words, the approximate solution of the Example is closest to the optimal solution. On the other hand, the search results of Comparative Example 2, which includes a fourth prompt P4 and an undirected graph update instruction in addition to the prompts of Comparative Example 1, are inferior to those of Comparative Example 1.

[0032] Figure 9 is an explanatory diagram of the average value, standard deviation, and approximation rate of the search results for the example and comparative examples 1 and 2. Figure 9 shows the average value, standard deviation, and approximation rate of each of the 10 types of path information searched in one search process when each path length is searched 50 times. As shown in Figure 9, the approximation rate is the value obtained by dividing the average value of the path lengths in the 50 search processes by the shortest path length, which is the optimal solution. In this embodiment, the search results shown in Figures 8 and 9 are output as images to the monitor by the output unit 40.

[0033] As described above, in the search device 100 of this embodiment, the first acquisition unit 11 acquires information related to the setting of the Traveling Salesperson Problem as a set combinatorial optimization problem in natural language. The second acquisition unit 12 acquires an image of a directed graph (Figure 4) as information related to the Traveling Salesperson Problem, which consists of nodes and edges, and has the orientation of the edges connecting the nodes set. The second acquisition unit 12 also acquires instructions to create an image of the directed graph in the iterative processing for searching for the optimal solution. The search unit 13 inputs the natural language acquired by the first acquisition unit 11, the directed graph update instructions acquired by the second acquisition unit 12, and an example image of the directed graph into a large-scale language model. The search unit 13 searches for the optimal solution to the Traveling Salesperson Problem of att48 as a combinatorial optimization problem expressed in natural language, using the orientation of the edges included in the directed graph and iterative processing accompanied by instructions to create the directed graph. The output unit 40 outputs the search results shown in Figures 8 and 9. In this embodiment, the directed graph is acquired as an image different from natural language. In the Traveling Salesperson Problem, the direction of the edges is not constrained, meaning that the direction of the edges is the same as the direction of the arrow DR1 in the directed graph, i.e., the counter-direction of a given approximate solution. In this embodiment, for the Traveling Salesperson Problem where the direction of the edges is not constrained, an image of a directed graph with constrained edge directions is used as an image different from natural language, and a large-scale language model searches for the optimal solution using the order information obtained from the direction of the edges. As a result, by constraining the order information, the number of iterations performed by the search of the large-scale language model is suppressed, and the optimal or approximate solution is derived. On the other hand, Comparative Example 2, which uses an image of an undirected graph (Figure 7) as an update prompt for the natural language of Comparative Example 1, shows lower solution-finding performance than Comparative Example 1, which does not use an image of an undirected graph, as shown in Figures 8 and 9. Here, in the Traveling Salesperson Problem path, the path that goes back to a given approximate solution is the same approximate solution. This is not inconsistent with the solution to the Traveling Salesperson Problem, but it is inconsistent with the order information attached to the prompt expressed in natural language.Therefore, the undirected graph contains path information that does not correspond to the order information implicitly assigned in the generation instructions using natural language. This uncorresponding path information is thought to have hindered the search for the optimal solution in the large-scale language model. In this embodiment, the directed graph does not show a path that goes backward in reverse order. Although this is not inconsistent with the solution to the Traveling Salesperson Problem, it does not include a reverse path as an approximate solution. However, it does match the order information assigned to the prompts expressed in natural language. As a result, it is thought that the update instructions for the directed graph did not cause confusion in the large-scale language model, but rather strengthened the order information. Thus, according to this embodiment, the optimal or approximate solution to the combinatorial optimization problem can be efficiently obtained using a large-scale language model.

[0034] Furthermore, the first acquisition unit 11 of this embodiment acquires instructions to create 10 types of solutions in each iterative search process as information related to the setting of the Traveling Salesperson Problem. The second acquisition unit 12 acquires a third prompt P3 as an instruction to update the directed graph of the solution closest to the optimal solution among the 10 types of solutions created in each iterative search process. In this embodiment, among the multiple solutions generated in the iterative process, a directed graph of the approximate solution closest to the optimal solution is created, and directed graphs of the other solutions are not created. Therefore, in the next process in the iterative process, multiple solutions are searched from the directed graph of the approximate solution closest to the optimal solution. As a result, directed graphs are not created from solutions that are not close to the optimal solution, and the search for the optimal solution is performed more efficiently.

[0035] Furthermore, the search unit 13 of this embodiment generates a prompt consisting of the first prompt P1 in Figure 3, the third prompt P3, and the path information of a randomly generated solution, as the initial prompt, and does not acquire an image of the directed graph. In this embodiment, only one solution is acquired as the initial value, and it is not necessary to acquire a directed graph. Therefore, the user can search for the optimal solution and approximate solution of a combinatorial optimization problem simply by preparing a simple initial value.

[0036] <Second Embodiment> Figure 10 is a schematic block diagram of the search system (search device) 100a in the second embodiment. The search system 100a in the second embodiment differs from the search device 100 in the first embodiment in that the functions performed by the control device 10 in the first embodiment are divided into two control devices 10A and 10B. Therefore, in the second embodiment, configurations and processes that differ from those in the first embodiment will be described, as will configurations and processes that are the same as those in the first embodiment.

[0037] The search system 100a of the second embodiment, as shown in Figure 10, comprises two control devices 10A and 10B, a storage device 20, an input unit 30, and an output unit 40. Of the two control devices 10A and 10B, control device 10A is a so-called personal computer and functions as a first acquisition unit 11 and a second acquisition unit 12 using information input via the input unit 30. Control device 10B is a large computer known as a mainframe, host computer, or general-purpose computer. In the second embodiment, the search unit 13 of control device 10B searches for a solution to a combinatorial optimization problem using a large language model stored in the model DB 21 of the storage device 20. Because the solution search process is performed by control device 10B, which is a large computer with high computing power, the optimal solution or an approximate solution to the optimal solution is found quickly. The user can easily search for a solution to a combinatorial optimization problem by inputting natural language as information regarding the problem setting of the combinatorial optimization problem, an update instruction for the directed graph, and an image of an example of the directed graph to control device 10A.

[0038] <Modified examples of embodiments> The present invention is not limited to the embodiments described above, and can be implemented in various forms without departing from its spirit. For example, the following modifications are possible. Furthermore, in the above embodiments, some of the configurations implemented by hardware may be replaced with software, and conversely, some of the configurations implemented by software may be replaced with hardware.

[0039] <Example 1> The search device 100 of the first embodiment and the search system 100a of the second embodiment described above are examples of devices and systems for searching for the optimal solution to a combinatorial optimization problem. The search device can be modified to the extent that it acquires information about the problem setting of a combinatorial optimization problem in natural language, acquires instructions to update a directed graph and an image of the directed graph in the search for a solution, and inputs the acquired information into a large-scale language model to search for a solution. The search device may also search for solutions to combinatorial optimization problems other than the traveling salesman problem, such as vehicle routing problems and scheduling problems.

[0040] In the first embodiment described above, the first acquisition unit 11 acquired a second prompt P2 as an instruction to create 10 types of solutions in the search process, but the number of iterations created in one iteration of the search process can be changed. For example, 11 or more types of solutions may be created in one iteration, or fewer than 10 types of solutions may be created. In the first embodiment described above, the output unit 40 output the search results to the monitor as images shown in Figures 8 and 9. However, the output method of the output unit 40 can be changed, and for example, the results may be sent as data to another device. In the process of step S5 of the search flow shown in Figure 6, the termination was determined by the number of iterations of the search process, but the termination may be determined using the calculation time or the like.

[0041] <Modification 2> The input form used by the first acquisition unit 11 to acquire natural language and the input form used by the second acquisition unit 12 to acquire an image of a directed graph may be provided separately. By providing separate input forms for natural language and images, the search unit 13 can efficiently perform the search by separately recognizing the problem setting and the directed graph used to search for the optimal or approximate solution to the problem.

[0042] The control device 10 of the first embodiment described above may function as a discrimination unit that distinguishes between natural language about the setting of a combinatorial optimization problem and natural language about instructions for solving the combinatorial optimization problem, based on the natural language acquired by the first acquisition unit 11 and the second acquisition unit 12. Furthermore, separate input forms may be provided for acquiring natural language about the setting of a combinatorial optimization problem and for acquiring natural language about instructions for solving the combinatorial optimization problem.

[0043] The embodiments of this specification have been described above based on the embodiments and modifications described above. The embodiments described above are for the purpose of facilitating understanding of this specification and do not limit it. This specification may be modified and improved without departing from its spirit and the scope of the claims, and equivalents thereof are included in this specification. Furthermore, any technical features that are not described as essential in this specification may be deleted as appropriate.

[0044] The present invention can also be realized in the following forms. [Application Example 1] A search device for finding the optimal solution to a combinatorial optimization problem, A first acquisition unit that acquires information related to the problem setting of the aforementioned combinatorial optimization problem in natural language, A second acquisition unit acquires an image of a directed graph composed of nodes and edges, in which the orientation of the edges connecting the nodes is set, and an update instruction for the image of the directed graph in an iterative process for searching for the optimal solution. A search unit inputs the natural language acquired by the first acquisition unit, the image of the directed graph acquired by the second acquisition unit, and the update instructions into a large-scale language model, and searches for the optimal solution to the problem setting expressed by the input natural language using the orientation of the edges included in the directed graph and iterative processing accompanied by the update instructions. An output unit that outputs at least one of the optimal solution obtained by the search unit and an approximate solution of the optimal solution, A search device equipped with the following features. [Application Example 2] The search device described in Application Example 1, The first acquisition unit acquires a plurality of solution generation instructions in each of the iterative processes, The second acquisition unit is a search device that, as an update instruction, acquires an instruction to create an image of the directed graph of the approximate solution that is closest to the optimal solution among the multiple solutions generated in each of the iterative processes. [Application Example 3] A search device as described in Application Example 1 or Application Example 2, The first acquisition unit acquires at least one arbitrary solution to the combinatorial optimization problem as an initial value, The second acquisition unit is a search device that does not acquire the image of the directed graph as the initial value. [Application Example 4] A search method for finding the optimal solution to a combinatorial optimization problem, wherein a computer... A first acquisition step involves having a large-scale language model acquire information related to the problem setting of the aforementioned combinatorial optimization problem as natural language, A directed graph composed of nodes and edges, wherein the orientation of the edges connecting the nodes is set, and a second acquisition step of causing the large-scale language model to acquire an image of the directed graph and an update instruction for the image of the directed graph in an iterative process for searching for the optimal solution, The large-scale language model includes an output step that outputs at least one of the orientation of edges included in the directed graph, an optimal solution obtained using the iterative process with update instructions, and an approximate solution of the optimal solution. A search method that performs this task. [Application Example 5] A computer program that searches for the optimal solution to a combinatorial optimization problem, A first acquisition function that acquires information related to the problem setting of the aforementioned combinatorial optimization problem in natural language, A directed graph composed of nodes and edges, wherein the orientation of the edges connecting the nodes is set, and a second acquisition function that acquires an image of the directed graph and an update instruction for the image of the directed graph in an iterative process for searching for the optimal solution, A search function inputs the natural language obtained by the first acquisition function, the image of the directed graph obtained by the second acquisition function, and the update instructions into a large-scale language model, and searches for the optimal solution to the problem setting expressed by the input natural language using the orientation of edges included in the directed graph and iterative processing accompanied by the update instructions. An output function that outputs at least one of the optimal solution obtained by the search function and an approximate solution of the optimal solution, A computer program that causes a computer to execute something. [Explanation of Symbols]

[0045] 10, 10A, 10B… Control devices 11…First acquisition part 12…Second acquisition part 13…Exploration Department 20…Storage device 21…Model Database 30...Input section 40…Output section 100…Exploration device 100a... Search system (search device)

Claims

1. A search device for finding the optimal solution to a combinatorial optimization problem, A first acquisition unit that acquires information related to the problem setting of the aforementioned combinatorial optimization problem in natural language, A second acquisition unit acquires an image of a directed graph composed of nodes and edges, in which the orientation of the edges connecting the nodes is set, and an update instruction for the image of the directed graph in an iterative process for searching for the optimal solution. A search unit inputs the natural language acquired by the first acquisition unit, the image of the directed graph acquired by the second acquisition unit, and the update instructions into a large-scale language model, and searches for the optimal solution to the problem setting expressed by the input natural language using the orientation of the edges included in the directed graph and iterative processing accompanied by the update instructions. An output unit that outputs at least one of the optimal solution obtained by the search unit and an approximate solution of the optimal solution, A search device equipped with the following features.

2. A search device according to claim 1, The first acquisition unit acquires a plurality of solution generation instructions in each of the iterative processes, The second acquisition unit is a search device that, as an update instruction, acquires an instruction to create an image of the directed graph of the approximate solution that is closest to the optimal solution among the multiple solutions generated in each of the iterative processes.

3. A search device according to claim 1 or claim 2, The first acquisition unit acquires at least one arbitrary solution to the combinatorial optimization problem as an initial value, The second acquisition unit is a search device that does not acquire the image of the directed graph as the initial value.

4. A search method for finding the optimal solution to a combinatorial optimization problem, wherein a computer... A first acquisition step involves having a large-scale language model acquire information related to the problem setting of the aforementioned combinatorial optimization problem as natural language, A directed graph composed of nodes and edges, wherein the orientation of the edges connecting the nodes is set, and a second acquisition step of causing the large-scale language model to acquire an image of the directed graph and an update instruction for the image of the directed graph in an iterative process for searching for the optimal solution, The large-scale language model includes an output step that outputs at least one of the orientation of edges included in the directed graph, an optimal solution obtained using the iterative process with update instructions, and an approximate solution of the optimal solution. A search method that performs this task.

5. A computer program that searches for the optimal solution to a combinatorial optimization problem, A first acquisition function that acquires information related to the problem setting of the aforementioned combinatorial optimization problem in natural language, A directed graph composed of nodes and edges, wherein the orientation of the edges connecting the nodes is set, and a second acquisition function that acquires an image of the directed graph and an update instruction for the image of the directed graph in an iterative process for searching for the optimal solution, A search function inputs the natural language obtained by the first acquisition function, the image of the directed graph obtained by the second acquisition function, and the update instructions into a large-scale language model, and searches for the optimal solution to the problem setting expressed by the input natural language using the orientation of edges included in the directed graph and iterative processing accompanied by the update instructions. An output function that outputs at least one of the optimal solution obtained by the search function and an approximate solution of the optimal solution, A computer program that causes a computer to execute something.

Images