Optimization device, optimization method, and program

The use of ZDD with the Frank-Wolfe method addresses the inefficiencies in solving combinatorial optimization problems with large combinations by achieving linear convergence and fast optimization under less stringent convex conditions.

JP2024180022APending Publication Date: 2024-12-26NIPPON TELEGRAPH & TELEPHONE CORP +1
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
JP2023099427
Authority / Receiving Office
JP · JP
Patent Type
Applications
Current Assignee / Owner
Filing Date
2023-06-16
Publication Date
2024-12-26

AI Technical Summary

Technical Problem

Existing combinatorial optimization methods struggle to solve problems with exponentially large numbers of selectable combinations efficiently, especially when the objective function is less than strongly convex, leading to slow convergence and impractical computation times.

Method used

The method employs a zero-suppressed binary decision diagram (ZDD) to represent the set of selectable combinations and integrates it with the Frank-Wolfe method, allowing for linear convergence of the error with respect to the optimal value even when the objective function is less strongly convex, using a tolerance ε to achieve fast optimization.

Benefits of technology

This approach enables quick solution of combinatorial optimization problems with exponentially large combinations by ensuring linear convergence of the error with respect to the number of iterations, even under less stringent convex conditions, thus overcoming the limitations of previous methods.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 2024180022000001_ABST
    Figure 2024180022000001_ABST
Patent Text Reader

Abstract

To provide an optimization device, an optimization method, and a program for solving a combinatorial optimization problem at high speed even when errors of an objective function against an optimum value linearly converge under the condition where the objective function is looser than a strong convex and the number of selectable combinations is exponentially large.SOLUTION: Provided is an optimization device for determining a vector x that indicates that a combination Ci is included with a probability pi for an objective function f which is a convex function and a combinatorial set scrC representing a set of selectable combinations C1, ... , Ct belonging to X regarding an item set X, and that minimizes f(x). The optimization device comprises: an input unit that accepts, as inputs, the objective function f, a zero-suppressed binary decision diagram (ZDD) representing the combinatorial set scrC, and a permissible error ε from the optimum value of the objective function f; and an optimization unit that determines, using the ZDD, a next solution xT+1 from a solution xT determined in a last iteration in iterations of the Frank-Wolfe method, on the basis of the objective function f, the ZDD, and the permissible error ε, thereby determining the vector x where the error from the optimum value of the objective function f is smaller than or equal to the permissible error ε.SELECTED DRAWING: Figure 2
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present disclosure relates to an optimization device, an optimization method, and a program. [Background technology]

[0002] A typical combinatorial optimization problem aims to find a single combination that optimizes (minimizes or maximizes) a desired objective function. However, there are also combinatorial optimization problems that seek to find an optimal portfolio of combinations, i.e., a mix of several combinations. Some of these types of problems can be formulated using an objective function that is a convex function.

[0003] Combinatorial optimization problems that can be formulated using an objective function that is a convex function among problems of the type that finds a portfolio of optimal combinations can be solved using the Frank-Wolfe method described in Non-Patent Document 1 when the number of selectable combinations is small. In addition, when the objective function belongs to a good class called a strongly convex function, it can be solved using the fully-corrective Frank-Wolfe method or the away-step Frank-Wolfe method described in Non-Patent Document 2. However, in the Frank-Wolfe method described in Non-Patent Document 1, the error from the optimal value of the objective function decreases only at a rate proportional to the inverse of the number of iterations, and in the fully-corrective Frank-Wolfe method and the away-step Frank-Wolfe method described in Non-Patent Document 2, the error from the optimal value of the objective function decreases exponentially with respect to the number of iterations, but the condition that the objective function is a strongly convex function must be satisfied. In contrast, Non-Patent Document 3 describes the Frank-Wolfe method in which the error from the optimal value of the objective function decreases exponentially with respect to the number of iterations even under conditions where the objective function is looser than the strongly convex function. Hereinafter, the exponential decrease in the error from the optimal value of the objective function with respect to the number of iterations will be referred to as "linear convergence."

[0004] On the other hand, when the number of selectable combinations is exponentially large, one iteration of the Frank-Wolfe method takes an exponential amount of time, making it difficult to solve the problem in a realistic time. In response to this, Non-Patent Document 4 describes a method that uses a data structure that can compactly express a set of selectable combinations for calculations within the Frank-Wolfe method, thereby enabling practically high-speed optimization even when the number of selectable combinations is exponentially large. In the method described in Non-Patent Document 4, when the objective function is a strongly convex function, the error between the objective function and the optimal value converges linearly. [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] Marguerite Frank and Philip Wolfe. An algorithm for quadratic programming. Naval Research Logistics Quarterly, Vol. 3, pp. 95-110. [Non-Patent Document 2] Simon Lacoste-Julien and Martin Jaggi. On the global linear convergence of Frank--Wolfe optimization variants. In Proceedings of the 28th International Conference on Neural Information Processing Systems, Vol. 1, pp. 496-504, 2015. [Non-Patent Document 3] Dan Garber and Noam Wolf. Frank-Wolfe with a nearest extreme point oracle, In Proceedings of the 34th Annual Conference on Learning Theory, PMLR Vol. 134, pp. 2103-2132, 2021. [Non-Patent Document 4] Kengo Nakamura, Shinsaku Sakaue, and Norihito Yasuda. Practical Frank-Wolfe method with decision diagrams for computing Wardrop equilibrium of combinatorial congestion games. In Proceedings of the 34th AAAI Conference on Artificial Intelligence, pp. 2200-2209, 2020. Summary of the Invention [Problem to be solved by the invention]

[0006] However, although the method described in Non-Patent Document 4 linearly converges the error from the optimal value of the objective function, in theory, the speed is very slow, and the condition that the objective function is strongly convex may not hold in practical applications.

[0007] The present disclosure has been made in consideration of the above points, and provides a technology that enables a combinatorial optimization problem to be solved quickly even when the error with respect to the optimal value of an objective function converges linearly under conditions looser than the strongly convex objective function, and the number of selectable combinations is exponentially large. [Means for solving the problem]

[0008] The optimization device according to one aspect of the present disclosure is an optimization device that optimizes an objective function f, which is a convex function, and an itemset X={a1, . . . , a d}, the possible combinations C1, ,C t ⊆X, and for the set of combinations scrC, i is the probability p i (However, p i ≧0 and p1+···+p t1 is a block diagram showing an optimization method for minimizing a solution x obtained in a previous iteration of a Frank-Wolfe method based on the objective function f, the ZDD obtained by expressing the combination set scrC as a zero-suppressed binary decision diagram, and an allowable error ε with respect to an optimal value of the objective function f. T Using the ZDD, we get the following solution x T+1 and an optimization unit that calculates a vector x such that the error from the optimal value of the objective function f is equal to or smaller than the allowable error ε by calculating Effect of the Invention

[0009] A technology is provided that enables a combinatorial optimization problem to be solved quickly even when the error for the optimal value of an objective function converges linearly under conditions looser than those of a strongly convex objective function and the number of selectable combinations is exponentially large. [Brief description of the drawings]

[0010] [Figure 1] FIG. 2 is a diagram illustrating an example of a hardware configuration of an optimization device according to the present embodiment. [Diagram 2] FIG. 2 is a diagram illustrating an example of a functional configuration of an optimization device according to the present embodiment. [Diagram 3] 1 is a flowchart illustrating an example of a process for solving a combinatorial optimization problem. [Figure 4] FIG. 13 is a diagram illustrating an example of a flow of processing executed by an optimization unit. [Diagram 5] FIG. 13 is a diagram illustrating an example of the processing flow of LinearMin(w). [Figure 6] FIG. 11 is a diagram illustrating an example of the processing flow of NearPoint(x, w, ρ). DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0011] Hereinafter, one embodiment of the present invention will be described.

[0012] <Combinatorial optimization problem to find the optimal portfolio> Among combinatorial optimization problems, there are problems that seek the optimal portfolio of combinations, i.e., a mix of several combinations. In this type of problem, we consider an itemset X={a1, ,a d} and its possible combinations C1, ,C t ⊆X, the probability values ​​p1, . . . , p t This is the combination C i is the proportion p i These values ​​are probabilities, so p i ≧0 and p1+···+p t = 1.

[0013] Some of the above types of problems can be formulated as follows: First, the set of selectable combinations is

[0014]

number

[0015] Also, for a combination C ⊆ X, its indicator vector 1 C is a d=|X| dimensional vector, and a i Let us assume that if C ∈ C, then the i-th element is 1, otherwise the i-th element is 0. Then, consider the following optimization problem for the objective function f(x), which is a convex function.

[0016]

number

[0017]

number

[0018] This type of problem often appears when successively relaxing difficult combinatorial optimization problems: the problem of choosing one optimal combination from a set of possible combinations such as the one below is difficult to solve unless the set of combinations scrC has a special structure because the range of possible values ​​of x is discrete.

[0019]

number

[0020] Another application in which this type of problem appears is the equilibrium calculation of combinatorial congestion games. In congestion games, for a set of items X, each player chooses a combination of items C ⊆ X. However, what combinations of items can be selected is predetermined, and the set of such combinations is the strategy set scrC. Each element (combination of items) in the strategy set is called a strategy. Each item is set so that the more people who choose it, the higher the cost, and the cost that a player receives is the sum of the costs of the items in the strategy. At this time, each player does not cooperate with each other, but tries to choose a strategy with as low a cost as possible. An important state in congestion games is a state called an "equilibrium state." An equilibrium state is a state in which players are not dissatisfied, and is the state that is reached as a result of players who do not cooperate with each other trying to reach a state with the lowest cost. It is known that finding an equilibrium state is an optimization problem of the form of equation (1), which minimizes a function f called a potential function under the strategy set scrC.

[0021] <Conventional methods that can solve the combinatorial optimization problem shown in formula (1) and their issues> The combinatorial optimization problem shown in formula (1) can be solved by using the Frank-Wolfe method described in Non-Patent Document 1 when the number t of selectable combinations is small. At this time, the error from the optimal value of the objective function f decreases in proportion to the inverse 1 / T of the number of iterations T. Furthermore, when the objective function f belongs to a good class called a strongly convex function, the problem can be solved by using the fully-corrective Frank-Wolfe method or the away-step Frank-Wolfe method described in Non-Patent Document 2. At this time, the error from the optimal value of the objective function f decreases exponentially with respect to the number of iterations T (i.e., linear convergence).

[0022] On the other hand, when the number of selectable combinations t is exponentially large, one iteration of the Frank-Wolfe method takes an exponential time, making it difficult to solve the problem in a realistic time. The method described in Non-Patent Document 4 addresses this situation. This method compactly expresses the set of selectable combinations scrC using a zero-suppressed binary decision diagram (ZDD) described in Reference Document 1, and uses it for calculations inside the Frank-Wolfe method, thereby enabling practical high-speed optimization even when the number of selectable combinations t is exponentially large. When the objective function f is a strongly convex function, this method also linearly converges the error with the objective function f. Note that the method described in Non-Patent Document 4 focuses on finding the equilibrium state of a combinatorial congestion game, but the method is generally applicable to optimization problems in the form of Equation (1).

[0023] In the method described in Non-Patent Document 4, the error from the optimal value of the objective function f converges linearly with respect to the number of iterations T, but in theory, the speed is very slow, and if the number of items d in the item set X is also included in the analysis, in reality, only exponential convergence for a very small number such as T / exp(αd) can be guaranteed. In addition, the condition that the objective function f is strongly convex may not hold in practical applications. For example, when determining the equilibrium state of a combinatorial congestion game, if the cost of each item increases monotonically with respect to the proportion of people who selected that item, the potential function, which is the objective function, will be strongly convex, but if it is monotonically non-decreasing, that is, if there is a possibility that the cost does not increase with respect to the proportion of people who selected it, it will not be strongly convex. In this case, only a reduction in error proportional to the inverse 1 / T of the number of iterations T can be guaranteed.

[0024] <Proposed method> To solve the problems existing in the method described in Non-Patent Document 4, a method of utilizing ZDD in another type of Frank-Wolfe method described in Non-Patent Document 3 is proposed. In the Frank-Wolfe method described in Non-Patent Document 3, in the optimization problem of the form of Equation (1), even under the condition that the objective function f is looser than strongly convex, the error with respect to the optimal value of the objective function f converges linearly with respect to the number of iterations T. Also, in the analysis including the number of items d, it shows exponential convergence with respect to T / O(d 4 ), and only polynomial-level decay occurs with respect to the number of items d, not exponential. On the other hand, the Frank-Wolfe method itself described in Non-Patent Document 3 requires time proportional to the number t of combinatorial selections possible in one iteration, so it cannot be applied when the number of combinatorial selections possible is exponentially large. Therefore, in the proposed method, by making a part of the calculation inside the Frank-Wolfe method a procedure using ZDD, practical high-speed optimization is realized.

[0025] Hereinafter, the optimization device 10 for obtaining the solution of the combinatorial optimization problem shown in Equation (1) by the proposed method will be described.

[0026] The differences between the methods described in Non-Patent Document 3 and Non-Patent Document 4 and the proposed method are as follows: The method described in Non-Patent Document 3 requires time proportional to the number t of combinatorial selections possible in one iteration of the calculation inside the Frank-Wolfe method, so it cannot be applied when the number of combinatorial selections possible is exponentially large. On the other hand, the proposed method takes time proportional to the size of the ZDD representing the set scrC of combinatorial selections possible in one iteration of the calculation inside the Frank-Wolfe method, and can be optimized practically at high speed even when the number t of combinatorial selections possible is exponentially large, similar to Non-Patent Document 4. Furthermore, different from the method described in Non-Patent Document 4, it has linear convergence of the error under the condition that the objective function f is looser than strongly convex, and exponential convergence with respect to T / O(d 4 ).

[0027] <Structure of ZDD> As a preparation for explaining the optimization device 10 according to this embodiment, the structure of a ZDD will be explained. A ZDD is a data structure that represents a set of combinations using a loop-free graph (Directed Acyclic Graph (DAG)) made up of a collection of nodes and oriented edges. A ZDD has two types of nodes: terminal nodes and branch nodes, which are represented by downtack symbols and uptack symbols, respectively. Hereinafter, in the text of this specification, a downtack symbol will be substituted with a capital tau "T" and an uptack symbol will be substituted with a vertical symbol "⊥".

[0028] There are no outgoing edges from each terminal node, and each branch node has exactly two outgoing edges, a 0-branch and a 1-branch. Also, in a ZDD, there is exactly one node with no incoming edges (branches), which is called the root node and is denoted by r. For a branch node v, the node pointed to by the 0-branch from that branch node v is called the 0-child node and is denoted by v0. Similarly, for a branch node v, the node pointed to by the 1-branch from that branch node v is called the 1-child node and is denoted by v1.

[0029] Each branch node v has an integer value v called a label. lv ∈{1, ,c}, which corresponds to the item number. For the terminal node, the label value is c+1. ​​In this case, each 0-branch and 1-branch must go from the node with the smaller label to the node with the larger label. That is, for any branch node v, v lv <(v0) lv , and v lv <(v1) lv Make sure that the following holds true.

[0030] The combination set represented by a ZDD is represented by a set of paths from the root node r to the terminal node T. A path R from the root node r to the terminal node T is a 1-edge that goes out from a node with label i if and only if iIt corresponds to a subset C(R) ⊆ X of the itemset X such that ∈C(R). The set of subsets corresponding to all paths from the root node r to the terminal node T is the combination set represented by the ZDD.

[0031] Here, for example, when the combination set scrC is a set of partial structures such as matching of a certain graph structure or a cycle, a tree, or a path, a ZDD representing the combination set scrC can be efficiently constructed using a method called the frontier method described in Reference 2. In addition, a ZDD representing any combination set can be constructed by using an operation called the Apply operation described in Reference 1. Therefore, in the following embodiment, it is assumed that a ZDD representing the combination set scrC is provided to the optimization device 10 as a representation of the combination set scrC.

[0032] <Hardware configuration example of optimization device 10> An example of a hardware configuration of an optimization device 10 according to this embodiment is shown in Fig. 1. As shown in Fig. 1, the optimization device 10 according to this embodiment has an input device 101, a display device 102, an external I / F 103, a communication I / F 104, a RAM (Random Access Memory) 105, a ROM (Read Only Memory) 106, an auxiliary storage device 107, and a processor 108. Each of these pieces of hardware is connected to each other via a bus 109 so as to be able to communicate with each other.

[0033] The input device 101 is, for example, a keyboard, a mouse, a touch panel, a physical button, etc. The display device 102 is, for example, a display, a display panel, etc. Note that the optimization device 10 does not necessarily have at least one of the input device 101 and the display device 102, for example.

[0034] The external I / F 103 is an interface with an external device such as a recording medium 103a. Examples of the recording medium 103a include a CD (Compact Disc), a DVD (Digital Versatile Disk), an SD memory card (Secure Digital memory card), and a USB (Universal Serial Bus) memory card.

[0035] The communication I / F 104 is an interface for connecting the optimization device 10 to a communication network. The RAM 105 is a volatile semiconductor memory (storage device) that temporarily stores programs and data. The ROM 106 is a non-volatile semiconductor memory (storage device) that can store programs and data even when the power is turned off. The auxiliary storage device 107 is a non-volatile storage device such as a hard disk drive (HDD), a solid state drive (SSD), or a flash memory. The processor 108 is various types of arithmetic devices such as a central processing unit (CPU) or a graphic processing unit (GPU).

[0036] Note that the hardware configuration shown in FIG. 1 is just an example, and the optimization device 10 may have multiple auxiliary storage devices 107 or multiple processors 108, may not have some of the hardware shown in the figure, or may have various types of hardware other than the hardware shown in the figure.

[0037] <Example of functional configuration of optimization device 10> An example of the functional configuration of the optimization device 10 according to this embodiment is shown in Fig. 2. As shown in Fig. 2, the optimization device 10 according to this embodiment has an input unit 201, an optimization unit 202, and an output unit 203. Each of these units is realized, for example, by a process in which one or more programs installed in the optimization device 10 are executed by a processor 108 or the like.

[0038] When an objective function f, which is a convex function, a ZDD expressing a combination set scrC, which is a set of selectable combinations, and an allowable error ε are given, an input unit 201 inputs the objective function f, the ZDD, and the allowable error ε.

[0039] The optimization unit 202 solves the combinatorial optimization problem shown in equation (1) within an error ε based on the objective function f, ZDD, and allowable error ε input by the input unit 201. That is, the optimization unit 202 solves the optimal solution x * For |f(x)-f(x * )|≦ε such that x∈conv({1 C |C∈scrC}.

[0040] The output unit 203 outputs the solution x obtained by the optimization unit 202 to a predetermined output destination. Examples of the predetermined output destination include the display device 102 such as a display, the auxiliary storage device 107 such as a HDD, SSD, or flash memory, and other devices, equipment, and terminals connected via a communication network.

[0041] In addition to the solution x, the output unit 203 outputs real numbers λ1, . . . , λ that can be expressed as follows for the solution x, as necessary. l and C1,...,C l ∈scrC may be output.

[0042]

number

[0043] <Solving the combinatorial optimization problem shown in formula (1)> The process of finding a solution to the combinatorial optimization problem shown in formula (1) will be described below with reference to Fig. 3. In the following, it is assumed that the optimization device 10 is given an objective function f that is a convex function, a ZDD that represents a combination set scrC that is a set of selectable combinations, and a permissible error ε.

[0044] The input unit 201 receives a given objective function f, ZDD, and allowable error ε (step S101).

[0045] Next, the optimization unit 202 solves the combinatorial optimization problem shown in formula (1) within the error ε based on the objective function f, ZDD, and allowable error ε input in the above step S101 to obtain a solution x (step S102). Note that the details of the processing in this step will be described later.

[0046] Then, the output unit 203 outputs the solution x obtained in the above step S102 to a predetermined output destination (step S103).

[0047] <Processing flow executed by the optimization unit 202> Hereinafter, the details of the flow of the process executed by the optimization unit 202 in the above step S102 will be described with reference to FIG.

[0048] First, the optimization unit 202 performs the optimization for x1←1 using an arbitrarily selected combination C∈scrC. C Then, the optimization unit 202 initializes the solution x T f(x T )-f(x * The process of lines 4 to 7 is repeated until it is determined that the condition x )|≦ε is satisfied (line 3). T f(x T )-f(x * If it is determined that |≦ε is satisfied (line 4), the solution x T is output as the solution x of the combinatorial optimization problem shown in equation (1) (line 5). Note that T is a variable that represents the number of iterations of the processing in lines 4 to 7.

[0049] Here, the solution x T f(x T )-f(x * As a method for judging whether |f(x )|≦ε is satisfied, it is possible to use the duality gap described in Reference 3. In the fourth line of Figure 4, we use this duality gap to T )-f(x * )|≦ε.

[0050] x T The duality gap in is the value defined by the following formula.

[0051]

number

[0052] In the sixth line, the optimization unit 202 T and ∇f(x T ) and βρ T By calling the NearPoint function with and as arguments, the optimal combination C' is selected from scrC using ZDD, and v T+1 The instruction vector 1 C' The details of the NearPoint function are described later. T The method of determining this will be described later.

[0053] In line 7, the optimization unit 202 calculates the vectors v1, . . . , v T+1 Find the x that minimizes the following function in the convex hull of x T+1 Assign to .

[0054]

number

[0055] <Constants β and ρ T How to determine Constants β and ρ T First, β is set to 1. Next, ρ T There are various ways to define it, but two ways of defining it will be explained below as examples.

[0056] The first method is to determine the objective function f to be optimized when its properties are well known. The objective function f satisfies the quadratic growth property in Definition 1 of Non-Patent Document 3 with a constant α>0. In addition, the optimal value f of the combinatorial optimization problem shown in Equation (1) is * The lower bound L of f * Assume that we know the value that satisfies ≧L. Then, C=f(x1)-L, M=max{(1 / α)(4+8d 4 ),1 / 2}, and then ρ T Determine.

[0057]

number

[0058] The second method is to determine the objective function f when the nature of the function is not well known. In this case, ρ T Instead of setting a single value for TConsider a procedure of trying the values of ρ and adopting the one that gives the smallest objective function value. More specifically, consider a procedure where k is set as follows, and for ρ T = 1, 1 / 2, 1 / 4, ···, 1 / 2 k the processing in the 4th to 5th lines of Figure 4 is repeated, and v T+1 and x T+1 when the value of f(x T+1 ) is the smallest are adopted.

[0059]

Number

[0060] However, the first and second definitions above are examples of specific definitions of ρ T and are not limited to these definitions.

[0061] <Flow of processing of the LinearMin function> The flow of processing of the LinearMin function when w is given as an argument will be described below with reference to Figure 5. In LinearMin(w), using the ZDD representing scrC, the minimum value of the product (inner product) of the transpose of 1 C when C ∈ scrC and w is calculated.

[0062] First, the optimization unit 202 initializes D Τ ← 0 and D ⊥ ← ∞ (line 1). Note that ∞ can be set as the maximum value that can be obtained in the implementation.

[0063] Next, the optimization unit 202 scans each branch node v in bottom-up order, that is, in descending order of the label (line 2), and calculates the value of D v as follows (line 3).

[0064]

Number

[0065]

Number

[0066] <Flow of NearPoint function processing> Hereinafter, the flow of the NearPoint function when x, w, and ρ are given as arguments will be described with reference to FIG. 6. In NearPoint(x, w, ρ), a process of calculating C ∈ scrC that minimizes the following function is performed.

[0067]

Number

[0068] First, the optimization unit 202 initializes D Τ ← 0, D ⊥ ← ∞ (the 1st line). Note that ∞ may be the maximum value that can be taken in the implementation.

[0069] Next, the optimization unit 202 scans each branch node v in bottom-up order, that is, in descending order of the label (the 2nd line), and calculates the value of D v as follows (the 3rd line).

[0070]

number

[0071]

number

[0072]

number

[0073] Next, the optimization unit 202 assigns the root node r of the ZDD to v and assigns an empty set to C (lines 4 and 5). After that, the optimization unit 202 repeatedly traces from v to its child nodes until v=T (line 6). At this time, the optimization unit 202 calculates D v =D v_1 +w v_lv +ρ(1-2x v_lv ) (line 7), then C has a v_lv (line 8), and update v to v1 (line 9). On the other hand, the optimization unit 202 adds v =D v_1 +w v_lv +ρ(1-2x v_lv ), update v to v0 (line 11). This is the same as D in line 3. vThis means that if the first item of min is smaller in the calculation of , the 0-child node v0 is traced, and if the second item is smaller, the 1-child node v1 is traced. Finally, the optimization unit 202 performs C is output to the caller as the result of NearPoint(x,w,ρ) (line 12).

[0074] <Summary> As described above, the optimization device 10 according to this embodiment can quickly solve the combinatorial optimization problem shown in formula (1) even when the number of selectable combinations is exponentially large by expressing the combination set in ZDD. Furthermore, the optimization device 10 according to this embodiment can linearly converge the error for the optimal value of the objective function f with respect to the number of iterations T under conditions looser than the strongly convex condition.

[0075] As a case where the error linearly converges with respect to the number of iterations T even under conditions looser than those of the objective function f being strongly convex, more precise conditions are described in Definition 1 of Non-Patent Document 3. For example, the case where the objective function f is expressed as follows using another strongly convex function g, a matrix A, and a vector b corresponds to this case.

[0076]

number

[0077] When calculating the equilibrium state of a combinatorial congestion game, linear convergence can be guaranteed even in the following case. A combinatorial congestion game is a game where each item a i Usage rate (the percentage of players who choose a combination that includes that item) x i A cost function c that is monotonically non-decreasing with respect to i (x i) is given. If the potential function is defined as follows, then solving the combinatorial optimization problem shown in formula (1), where scrC is a set of selectable combinations, is equivalent to finding an equilibrium state.

[0078]

number

[0079]

number

[0080] The present invention is not limited to the above-described embodiments specifically disclosed, and various modifications, changes, combinations with known technologies, etc. are possible without departing from the scope of the claims.

[0081] [Reference 1] Reference 1: Shin-ichi Minato. Zero-suppressed BDDs for set manipulation in combinatorial problems. In Proceedings of the 30th ACM / IEEE Design Automation Conference, pp. 272-277, 1993. Reference 2: Jun Kawahara, Takeru Inoue, Hiroaki Iwashita, and Shin-ichi Minato. Frontier-based search for enumerating all constrained subgraphs with compressed representation. IEICE Transactions on Fundamentals of Electronics, Communications and Computer Sciences, Vol. E100.A, pp. 1773-1784, 2007. Reference 3: Martin Jaggi. Revisiting Frank-Wolfe: projection-free sparse convex optimization. In Proceedings of the 30th International Conference on Machine Learning, PMLR Vol. 28, pp. 427-435, 2013. [Explanation of symbols]

[0082] 10 Optimization Device 101 Input Device 102 Display device 103 External I / F 103a Recording media 104 Communication I / F 105 RAM 106 ROM 107 Auxiliary storage 108 processors 109 Bus 201 Input section 202 Optimization Department 203 Output section

Claims

1. Given a convex objective function f and an item set X = {a 1 , ..., a d }, selectable combination C 1 , ..., C t ⊆X, and a combination set scrC represents the set of combinations C i is the probability p i (However, p i ≧0 and p 1 +...+p t An optimization device for determining a vector x that represents that f(x) is included in the equation (f(x)=1), and that minimizes f(x), an input unit for inputting the objective function f, a ZDD obtained by expressing the combination set scrC as a zero-suppressed binary decision diagram, and a tolerance ε between the optimal value of the objective function f and the ZDD; Based on the objective function f, the ZDD, and the allowable error ε, the solution x obtained in the previous iteration is used in each iteration of the Frank-Wolfe method. T Using the ZDD, the following solution x T+1 an optimization unit for determining a vector x having an error from an optimal value of the objective function f that is equal to or smaller than the allowable error ε by determining An optimization device having the above configuration.

2. The optimization unit is At each iteration of the Frank-Wolfe method, the ZDD is used to calculate x T τ ∇f(x T ) + ρ||x T +x || 2 (where τ is the transpose and ρ is a constant) T+1 Calculate as: v 1 , ..., v T+1 From the next solution x T+1 The optimization device according to claim 1 , further comprising:

3. The optimization unit is For each branch node v of the ZDD, D v ←min{D v_0 , D v_1 + (∇f(x T )) v_lv +ρ(1-2(x T ) v_lv ) (where v 0 is the 0-child node of v, v 1 is the 1-child node of v, v lv is the label of v, (x T ) v_lv xxx T No.V lv The th element, (∇f(x T )) v_lv is ∇f(x T )v lv The th element, D Τ = 0, D ⊥ = maximum value) and From the root node r of the ZDD to the terminal node T, v =D v_1 + (∇f(x T )) v_lv +ρ(1-2(x T ) v_lv ) then 1-child node, D v ≠D v_1 + (∇f(x T )) v_lv +ρ(1-2(x T ) v_lv ) if 0-The combination represented by the path when branched to the child node is T+1 The optimization device according to claim 2 , wherein the optimization is performed as follows:

4. The optimization unit is v 1 , ..., v T+1 In the convex hull of x τ ∇f(x T )+(β / 2)||x−x T || 2 (where β is a constant) is minimized by the next solution x T+1 The optimization device according to claim 2 or 3, wherein the optimization device calculates the following:

5. The optimization unit is The solution x T 2. The optimization device according to claim 1, wherein the vector x is obtained by repeating the iterations of the Frank-Wolfe method until a gap in duality between x and x is equal to or smaller than the tolerance ε.

6. Given a convex objective function f and an item set X = {a 1 , ..., a d }, selectable combination C 1 , ..., C t ⊆X, and a combination set scrC represents the set of combinations C i is the probability p i (However, p i ≧0 and p 1 +...+p t An optimization device for determining a vector x that represents that f(x) is included in the equation (1) and that minimizes f(x), an input step of inputting the objective function f, a ZDD obtained by expressing the combination set scrC as a zero-suppressed binary decision diagram, and a tolerance ε between the optimal value of the objective function f and the ZDD; Based on the objective function f, the ZDD, and the allowable error ε, the solution x obtained in the previous iteration is used in each iteration of the Frank-Wolfe method. T Using the ZDD, the following solution x T+1 an optimization procedure for obtaining a vector x whose error from an optimal value of the objective function f is equal to or smaller than the allowable error ε by obtaining Optimization method to perform.

7. Given a convex objective function f and an item set X = {a 1 , ..., a d }, selectable combination C 1 , ..., C t ⊆X, and a combination set scrC represents the set of combinations C i is the probability p i (However, p i ≧0 and p 1 +...+p t An optimization device for finding a vector x that represents that f(x) is included in the equation (1) and that minimizes f(x), an input step of inputting the objective function f, a ZDD obtained by expressing the combination set scrC as a zero-suppressed binary decision diagram, and a tolerance ε between the optimal value of the objective function f and the ZDD; Based on the objective function f, the ZDD, and the allowable error ε, the solution x obtained in the previous iteration is used in each iteration of the Frank-Wolfe method. T Using the ZDD, the following solution x T+1 an optimization procedure for obtaining a vector x whose error from an optimal value of the objective function f is equal to or smaller than the allowable error ε by obtaining A program that executes the following.

Citation Information

Patent Citations

  • Optimization method, optimization apparatus, and program

    JP2022178099A

  • Equilibrium state calculation device, equilibrium state calculation method and program

    JP7279820B2