CN114912647A

CN114912647A - Apparatus, method and machine-readable storage medium for decision making

Info

Publication number: CN114912647A
Application number: CN202110177334.6A
Authority: CN
Inventors: 杨铭; 夏迎炬; 刘汝杰
Original assignee: Fujitsu Ltd
Current assignee: Fujitsu Ltd
Priority date: 2021-02-07
Filing date: 2021-02-07
Publication date: 2022-08-16
Also published as: JP2022121390A

Abstract

The present disclosure relates to an apparatus, method, and machine-readable storage medium for deciding a decision. The device includes: an acquisition unit that acquires M prediction sequences in M scenes associated with decision of a decision and probabilities corresponding to the respective prediction sequences; an allocation unit that allocates N specific decision schemes for each scene, respectively; a first calculation unit that obtains, for each scene, a group optimal decision scheme corresponding to each scene; a merging unit that obtains a scene combination decision scheme based on a group optimal decision scheme and M prediction sequences of M scenes; and a second calculation unit that obtains a final decision scheme based on the scenario combination decision scheme and the probability.

Description

Apparatus, method and machine-readable storage medium for decision making

Technical Field

The present disclosure relates to the field of information processing, and in particular, to an apparatus, method, and machine-readable storage medium for decision making.

Background

This section provides background information related to the present disclosure that is not necessarily prior art.

Making decisions based on information and constraint conditions is a common behavioral pattern for humans. In practice, however, the information is not always accurate.

The problem of information determination can be referred to as a deterministic planning problem. And for the problem of uncertain information, the method can be called a random programming problem. Using inaccurate information necessarily results in poor decisions. In order to alleviate the decision problem caused by the inaccuracy of the information, people naturally think of enumerating a plurality of possible values of the information and estimating the occurrence probability of the information so as to make a decision that the optimization target expects to be optimal. Therefore, how to make an optimal decision according to various possible values and occurrence probabilities thereof becomes an important topic of research in the field.

Disclosure of Invention

This section provides a general summary of the disclosure, and is not a comprehensive disclosure of its full scope or all of its features.

It is an object of the present disclosure to provide an apparatus, method and machine-readable storage medium for optimizing decision making.

According to an aspect of the present disclosure, there is provided an apparatus for deciding a decision, comprising: an acquisition unit that acquires M prediction sequences in M scenes associated with decision of a decision and probabilities corresponding to the respective prediction sequences; an allocation unit that allocates N specific decision schemes for each scene, respectively; a first calculation unit that obtains, for each scene, a group optimal decision scheme corresponding to each scene; a merging unit that obtains a scene combination decision scheme based on a group optimal decision scheme and M prediction sequences of M scenes; and a second calculation unit which obtains a final decision scheme based on the scene combination decision scheme and the probability, wherein M and N are natural numbers greater than 0.

According to another aspect of the present disclosure, there is provided a method for deciding a decision, comprising: obtaining M prediction sequences under M scenes associated with decision of the decision and probabilities respectively corresponding to the prediction sequences; respectively distributing N specific decision schemes for each scene; obtaining a group optimal decision scheme corresponding to each scene for each scene; obtaining a scene combination decision scheme based on the group optimal decision scheme and the M prediction sequences of the M scenes; and obtaining a final decision scheme based on the scene combination decision scheme and the probability, wherein M and N are natural numbers larger than 0.

According to another aspect of the present disclosure, there is provided a machine-readable storage medium having embodied thereon a program product comprising machine-readable instruction code stored therein, wherein the instruction code, when read and executed by a computer, is capable of causing the computer to perform a method according to the present disclosure.

The apparatus, method, and machine-readable storage medium for deciding a decision according to the present disclosure may form a multi-level calculation structure (a first calculation unit and a second calculation unit) that first optimizes a specific decision scheme of each scene by the first calculation unit, and then further optimizes a scene combination decision scheme related to the optimized result of each scene by the second calculation unit based on a prediction sequence and a corresponding probability to obtain a final decision scheme. Here, through the optimization of the first calculation unit, an optimal decision scheme for each scene may be obtained, and through the optimization of the second calculation unit, a final decision scheme that is optimal for the entire decision-making problem and takes into account the probability of each scene may be obtained, thereby implementing multi-level optimization for the final decision scheme.

Further areas of applicability will become apparent from the description provided herein. The description and specific examples in this summary are intended for purposes of illustration only and are not intended to limit the scope of the present disclosure.

Drawings

The drawings described herein are for illustrative purposes only of selected embodiments and not all possible implementations, and are not intended to limit the scope of the present disclosure. In the drawings:

fig. 1 is a block diagram illustrating the structure of a decision-making device according to an embodiment of the present disclosure;

fig. 2 is a flow diagram illustrating a method according to an embodiment of the present disclosure;

FIG. 3 is a schematic table illustrating relevant parameters in an order plan application case in accordance with an embodiment of the present disclosure;

FIG. 4 is a schematic diagram illustrating the generation of a scene tree used in embodiments according to the present disclosure;

FIG. 5 is a schematic diagram illustrating a particle swarm optimization algorithm used in embodiments according to the present disclosure;

fig. 6 is a block diagram illustrating the structure of a decision-making device according to a first embodiment of the present disclosure;

FIG. 7 is a schematic diagram illustrating a decision-making device according to a first embodiment of the present disclosure;

fig. 8 is a schematic diagram illustrating the operation of a merging unit in a decision making device according to a first specific embodiment of the present disclosure;

FIG. 9 is a flow chart illustrating a method according to a first specific embodiment of the present disclosure;

fig. 10 is a block diagram illustrating the structure of a decision-making device according to a second embodiment of the present disclosure;

FIG. 11 is a schematic diagram illustrating a decision-making device according to a second embodiment of the present disclosure;

fig. 12 is a schematic diagram illustrating the operation of a reallocation unit in a decision making apparatus according to a second specific embodiment of the present disclosure;

FIG. 13 is a flow chart illustrating a method according to a second particular embodiment of the present disclosure; and

fig. 14 is a block diagram of an exemplary structure of a general-purpose personal computer in which the decision-making apparatus and method according to the embodiments of the present disclosure can be implemented.

While the disclosure is susceptible to various modifications and alternative forms, specific embodiments thereof have been shown by way of example in the drawings and are herein described in detail. It should be understood, however, that the description herein of specific embodiments is not intended to limit the disclosure to the particular forms disclosed, but on the contrary, the intention is to cover all modifications, equivalents, and alternatives falling within the spirit and scope of the disclosure. It is noted that throughout the several views, corresponding reference numerals indicate corresponding parts.

Detailed Description

Examples of the present disclosure will now be described more fully with reference to the accompanying drawings. The following description is merely exemplary in nature and is not intended to limit the present disclosure, application, or uses.

Example embodiments are provided so that this disclosure will be thorough, and will fully convey the scope to those skilled in the art. Numerous specific details are set forth such as examples of specific components, devices, and methods to provide a thorough understanding of embodiments of the present disclosure. It will be apparent to those skilled in the art that specific details need not be employed, that example embodiments may be embodied in many different forms, and that neither should be construed to limit the scope of the disclosure. In certain example embodiments, well-known processes, well-known structures, and well-known techniques have not been described in detail.

Hereinafter, the description will be given in the following order.

1. Description of the technical solution of the present application

2. Application case

3. Generation of scene trees

4. Introduction to particle swarm optimization algorithm

5. First embodiment

6. Second embodiment

7. Computer configuration examples of decision-making devices

8. Supplementary note

<1 > description of technical solution of the present application >

(example of configuration of decision-making device)

Fig. 1 shows a block diagram of the structure of an apparatus for deciding a decision (hereinafter also referred to as decision deciding apparatus) according to an embodiment of the present disclosure. As shown in fig. 1, a decision making apparatus 100 according to an embodiment of the present disclosure may include an obtaining unit 110, an allocating unit 120, a first calculating unit 130, a combining unit 140, and a second calculating unit 150.

The obtaining unit 110 may obtain M prediction sequences in M scenarios associated with decision of the decision and probabilities respectively corresponding to each prediction sequence, where M is a natural number greater than 0. Specifically, the obtaining unit 110 may obtain prediction results at a plurality of times based on historical data related to decision of the decision, and obtain M prediction sequences and corresponding probabilities based on the prediction results.

The predicted sequence may be a demand sequence related to demand for a particular item or transaction, and deciding a decision refers to deciding a decision variable that optimizes a particular objective related to the particular item or transaction based on the M demand sequences and corresponding probabilities. For example, the prediction sequence may be a multi-stage demand sequence relating to demand for a quantity of a particular item, and in this case, deciding a decision means deciding a decision variable that minimizes the overall cost based on the M demand sequences and corresponding probabilities. Further, the specific item may be, for example, a vehicle part, electric power, or the like, and the specific matter may be a service such as a repair service, a maintenance service, or the like of a 4S shop of the vehicle.

As an example, the acquisition unit 110 may derive M prediction sequences and corresponding occurrence probabilities in M scenes based on a tree data structure generated in the generation of a scene tree. Of course, the obtaining unit 110 may also obtain M prediction sequences and probabilities respectively corresponding to each prediction sequence based on other methods.

Further, the allocating unit 120 may allocate N specific decision schemes for each of the M scenes, where N is a natural number greater than 0. The N specific decision schemes may represent N possible decision schemes in respective scenarios. For example, the allocation unit 120 may randomly allocate N specific decision schemes within a certain range. Furthermore, the allocating unit 120 may provide the respective N specific decision schemes in the M scenarios to the first calculating unit 130.

Further, the first calculation unit 130 may obtain a group optimal decision scheme corresponding to each scene for each scene among the M scenes. The "group optimal decision scheme" herein means: assuming that the N specific decision schemes in each scene are regarded as one group, the optimal decision scheme obtained based on calculation of the N specific decision schemes in the group is referred to as a group optimal decision scheme.

Specifically, for each of the M scenes, the first computing unit 130 may perform multiple optimizations on N specific decision schemes in each scene, record N specific decision schemes after each optimization, and select a group of optimal decision schemes from all the specific decision schemes obtained in the optimization process. For example, the first calculation unit 130 may select a group optimal decision scheme corresponding to each of the M scenes in accordance with a first objective function corresponding to the M prediction sequences. It should be noted that, since the first computing unit 130 obtains the group optimal decision scheme for each scene, the first computing unit 130 does not consider the probability of each scene, i.e. the first objective function is independent of the probability of each scene.

Thus, the first calculation unit 130 may obtain M group optimal decision schemes corresponding to the M scenes and provide them to the merging unit 140.

Further, the merging unit 140 may obtain a scene combination decision scheme based on the group optimal decision scheme and the M prediction sequences of the M scenes. As an example, the merging unit 140 may combine the group optimal decision schemes of the M scenes based on a tree structure constructed by the M prediction sequences to obtain a scene combination decision scheme. The scenario combination decision scheme resulting from the merging unit 140 is substantially associated with the probability under each scenario. Further, the merging unit 140 may provide the scene combination decision scheme to the second calculation unit 150.

Further, the second calculation unit 150 may obtain a final decision scheme based on the scenario combination decision scheme and the probability. Specifically, the second computing unit 150 may perform multiple optimizations on the scene combination decision-making scheme, record the scene combination decision-making scheme after each optimization, select an optimal scene combination decision-making scheme from all the scene combination decision-making schemes obtained in the optimization process, and obtain a final decision-making scheme based on the optimal scene combination decision-making scheme. For example, the second calculation unit 150 may select an optimal scene combination decision scheme according to a second objective function corresponding to the M prediction sequences and the corresponding probabilities.

Thus, the decision-making device according to the present disclosure may form a multi-level calculation structure (first calculation unit and second calculation unit) that optimizes a specific decision-making scheme of each scene by the first calculation unit first, and then further optimizes a scene combination decision-making scheme related to the result after optimization of each scene by the second calculation unit based on the prediction sequence and the corresponding probability to obtain a final decision-making scheme. Here, through the optimization of the first calculation unit, an optimal decision scheme for each scene may be obtained, and through the optimization of the second calculation unit, a final decision scheme that is optimal for the entire decision-making problem and takes into account the probability of each scene may be obtained, thereby implementing multi-level optimization for the final decision scheme.

(example of the Process flow)

Fig. 2 illustrates a flow diagram of a method according to an embodiment of the present disclosure.

As shown in fig. 2, a method according to an embodiment of the present disclosure begins at step S110. In step S110, M prediction sequences in M scenarios associated with decision making and probabilities respectively corresponding to the prediction sequences are obtained.

Next, in step S120, N specific decision schemes are respectively assigned for each scene.

Next, in step S130, a group optimal decision scheme corresponding to each scene is obtained for each scene.

Next, in step S140, a scene combination decision scheme is obtained based on the group optimal decision scheme and the M prediction sequences of the M scenes.

Next, in step S150, a final decision scheme is obtained based on the scenario combination decision scheme and the probability. After that, the process ends.

The above steps of the method according to the embodiment of the present disclosure may be performed, for example, by the decision making apparatus 100 described above with reference to fig. 1, and the details thereof have been described in detail above and will not be repeated here.

Thus, the decision-making method according to the present disclosure may enable multi-level optimization of the final decision-making scheme.

<2. application case >

In practical applications, when making an optimal decision according to various possible values and occurrence probabilities of information, continuous decisions need to be made for many times. This scenario is a Multi-stage stochastic programming (MSSP) problem. Thus, the decision-making means according to the present disclosure may be implemented, for example, as a multi-stage stochastic programming problem solving means.

To embody the MSSP problem, order plans in inventory management are used as application cases. The order plan is a problem of determining when and how many orders are to be made for a specific article (e.g., a vehicle component) in the next T months. This requires that at most one order operation occur per month. The optimization goal of the order plan is to find an order decision that minimizes the total warehouse operating cost of T months.

Fig. 3 is a schematic table illustrating relevant parameters in an order plan application case according to an embodiment of the present disclosure. As shown in fig. 3, the parameters related to the optimization of the objective function are (assuming that the operation cost is composed of the order cost, inventory cost, and stock out cost):

■ Unit cost c of ordering, stocking and out-of-stock at future time t _t ,

(the unit cost varies with time),

■ demand d at future time t _t ，

■ decision initial inventory and initial quantity of stock in the beginning:

further, variables relevant to the optimization of the objective function are:

■ the stock and the stock shortage at the future time t are respectively:

in addition, the decision variables related to the optimization of the objective function are:

■ order quantity x at future time t _t 。

Further, constraints associated with the optimization of the objective function are:

■

(relation between stock quantity, order quantity, and demand quantity),

■

(relationship between the quantity of empty goods, the quantity of ordered goods, and the quantity of demanded goods),

■x _t not less than 0 (order quantity is not negative).

Therefore, the future demand d _t To determine the accurate value, the scene objective function of the case is:

on the other hand, if the future demand d _t For uncertain random variables, enumerate random vectors [ d ] ₁ ，...，d _T ]And taking the representative acquirable value and the occurrence probability as approximate estimation of the actual value of the random variable. Here, [ d ] ₁ ，...,d _T ]Also called prediction sequence, [ d ] ₁ ,...,d _T ]Is called a scene, denoted s _i The corresponding scene probability is denoted pb(s) _i ). Each scene corresponds to a deterministic planning problem, the objective function of which is expressed as

The desired objective function of the final stochastic programming problem is:

wherein:

in formula (2) and formula (3), S represents a set of all scenes.

Decision-making means (e.g. solving means) according to the present disclosure are used to find decision variables (order quantity sequence) that make F (x) as small as possible

It should be noted that the decision-making device described in the present application is used for solving a multi-stage stochastic programming problem, and is not related to a specific application case.

Although the above discloses the use of a decision-making means according to the present disclosure for order plan decisions in inventory management, the decision-making means according to the present disclosure may also be used in other application scenarios as long as it involves deciding decision variables based on multi-stage uncertain random variables.

For example, in the power industry, where problems are involved with the distribution of power stations to each location, uncertain random variables may be future power usage, and decision variables may be when and what size of power station needs to be built.

For example, in the area of investment of stocks, where the uncertain random variable may be a future price series of a stock and the decision variable may be when to buy how many stocks, there is a need to maximize the return on investment.

For example, in the field of chemical synthesis, it is necessary to control the multi-stage charging of the reaction furnace because the chemical reaction in the reaction furnace requires a suitable temperature, and different weather conditions (sun, snow, etc.) have a great influence on the temperature. Here, the uncertain random variable may be a future probable air temperature, and the decision variable may be when the temperature of the reaction furnace needs to be raised or lowered.

<3. creation of scene Tree >

Typically, many solvers that solve MSSP problems use a scene tree as an input. The scene tree is a tree-shaped data structure which effectively describes possible values and occurrence probabilities of uncertain parameters.

The solving device to which the decision-making device belongs according to the present disclosure only depends on the tree data structure of the possible values and occurrence probabilities of the uncertain parameters, and is not related to a specific scene tree generation algorithm. Therefore, a simple scenario tree generation example is given below to introduce the basic concept of a scenario tree, without involving a specific algorithm.

An example process of generating a scene tree will be described below with reference to fig. 4. Fig. 4 is a schematic diagram illustrating the generation of a scene tree used in embodiments according to the present disclosure. First, prediction results at a plurality of times are generated based on a plurality of prediction models and history data (history series). Then, one possible value at the previous moment is connected with all possible values at the next moment, and in the process, the nodes at the next moment are copied for multiple copies, so that the number of the nodes is increased. The above process is repeated in sequence along the time sequence until the last moment.

After the scene tree is obtained, each path from the root node to the leaf node of the tree is [ d ] ₁ ,...,d _T ]One possible value scenario. Here the scene probability pb(s) _i ) It can simply be assumed to be equal probability. Of course pb(s) _i ) But may also be non-equal in probability depending on the scene.

For example, in the example of fig. 4, 9 scenes s are generated ₁ To s ₉ . For example, in scene s ₁ In (1), the predicted sequence [ d ] ₁ ，d ₂ ，d ₃ ]Is [3,2,5 ]]The corresponding probability is pb(s) ₁ ) In the scene s ₂ In (b), the predicted sequence [ d ] ₁ ,d ₂ ,d ₃ ]Is [3,2,3 ]]The corresponding probability is pb(s) ₂ ) At scene s ₃ In (1), the predicted sequence [ d ] ₁ ,d ₂ ,d ₃ ]Is [3,2,4 ]]The corresponding probability is pb(s) ₃ )。

The problem for each scenario is a deterministic optimization problem whose objective function can be analogized to equation (1). And the objective function of the desired optimization problem for all possible scenarios can be analogized to equation (2).

It should be noted that fig. 4 only shows a case of constructing 9 scenes based on the ternary tree, but actually, the tree shape of the scene tree and the method of constructing the scene tree are not limited as long as the number of stages from the tree root node to the leaf nodes (the number of T, for example, in fig. 4, the number of stages from the tree root node to the leaf nodes is 3) of each scene are consistent.

<4 introduction to particle swarm optimization Algorithm >

A decision-making device according to the present disclosure may employ a Particle Swarm Optimization (PSO) algorithm to solve an optimization problem. The principle of the particle swarm optimization algorithm will be described below with reference to fig. 5. Fig. 5 illustrates a schematic diagram of a particle swarm optimization algorithm used in embodiments according to the present disclosure.

The particle swarm optimization algorithm takes the position of a series of particles as a candidate solution of a solution space by initializing, and then updates the speed of each particle and the position in the solution space through a simple rule. The motion of each particle is influenced by its own historical optimal solution spatial location and the historical optimal locations of all particles. And finally, the optimal positions in all the particles are used as the output result of the particle swarm optimization algorithm. The classical PSO is of the formula:

■ velocity update formula for each particle:

v _j (k+1)＝w*v _j (k)+c ₁ *r ₁ *(p _j (k)-x _j (k))+c ₂ *r ₂ *(p _g (k)-x _j (k))

wherein, w, c ₁ ,c ₂ Are respectively a weight parameter, r ₁ ,r ₂ To obey a uniform distribution of 0-1 uniform distribution, k is an iteration number index indicating the number of updates of the particle, x _j (k),v _j (k) Respectively showing the position and the velocity vector of the jth particle after the kth updating. p is a radical of _j (k) For the optimal history position, p, of the jth particle until the kth update _g (k) The optimal position until the k-th update for all particles.

■ location update formula for each particle:

x _j (k+1)＝x _j (k)+v _j (k+1)

■ formula for historical optimal location update for each particle:

■ optimal location update formula for all particles:

this is the optimal position for all particles at the current time.

Thus, the optimal position (group optimal position) of all particles at all times is then:

p _g (k+1)＝max(p _g (k)，p′ _g (k+1))。

specifically, fig. 5 depicts the principle of motion of each particle. As shown in fig. 5, w x v _j (k) Is the inertial force of the particle itself, c ₁ *r ₁ *(p _j (k)-x _j (k) C) attraction of particles to their own historical optimal positions, c ₂ *r ₂ *(p _g (k)-x _j (k) Attractive force to the particle for all particle history optimal positions, and the sum of the three vector moments constitutes the update speed (including direction and magnitude) of the particle.

For the convenience of the following description, the following mathematical symbols are defined in the present application:

■ PSO particle position matrix: p, the size of the matrix is [ the number of particles, the dimension of the decision variables ],

■ PSO particle velocity matrix: v, the size of the matrix is [ the number of particles, the dimension of the decision variables ],

■ PSO iterates until the current times k, and a particle position matrix P when all the particle historical optimal positions are obtained _g (optimally not necessarily at the kth iteration),

■ PSO iterates to the presentUntil the number k, obtaining a particle velocity matrix V when all the particles are in the historical optimal positions _g (optimally not necessarily at the kth iteration).

<5 > first embodiment

(framework of decision-making device)

The decision-making device according to the present disclosure is based on a scene tree and particle swarm optimization algorithm, and its overall framework is as follows.

For each scene (namely a path from a root node to a leaf node) in the scene tree, a particle swarm optimization is utilized to find out the particle position and speed matrix when the corresponding scene reaches the historical optimal position of all particles

And

for the above ordering plan case, the matrix size is [ particle number, T ]]. For example, in the case of the scene tree generated in fig. 4, the size of the matrix is [ number of particles, 3 ]]。

In the present embodiment, the PSOs on the respective scenes are solved as the first-layer PSOs. The PSO used by all scenes is required to have the same adjustable parameter, namely the particle number, and the initial positions and the velocities of all PSO particles of the first layer are initialized randomly when the solving device is initially operated. For example, taking the order plan case described above as an example, the objective function of each scenario of the first-layer PSO corresponds to equation (1) of the order plan case.

And initializing the position and the speed of the particles of the PSO algorithm of the second layer by utilizing the sharing relation between non-leaf nodes among scenes in the scene tree. The PSO algorithm of the second layer only comprises one PSO, and the corresponding particle position and speed matrix P ² And V ² Is [ particle number, number of nodes in scene tree ]]. The second layer of PSOs also has the same number of particles as all the PSOs of the first layer. The optimized location matrix obtained by optimizing the input location matrix of the second-tier PSO algorithm may be used to obtain a final decision scheme, such as a final decision scheme regarding order volume.

The configuration of the decision making means according to the present embodiment and the flow of processing performed by the decision making means will be described in detail below with continued reference to the order plan case described above with respect to fig. 3 and the scene tree structure generated in the description of fig. 4.

(example of configuration of decision-making device)

An example of the configuration of a decision making device 600 according to a first specific embodiment of the present disclosure, which is one example of the decision making device 100 shown in fig. 1, will be described in detail below with reference to fig. 6 to 8. Fig. 6 illustrates a block diagram of the structure of a decision making apparatus 600 according to a first specific embodiment of the present disclosure, fig. 7 illustrates a schematic diagram of the decision making apparatus 600, and fig. 8 illustrates a schematic diagram of the operation of a merging unit in the decision making apparatus 600.

As shown in fig. 6, the decision making apparatus 600 may include an obtaining unit 610, an allocating unit 620, a first calculating unit 630, a combining unit 640, and a second calculating unit 650. Among them, the first calculation unit 630 may include first-layer PSO units 1 to M, and the second calculation unit 650 may include a second-layer PSO unit 651 and a decision variable calculation unit 652.

The obtaining unit 610 may obtain M prediction sequences in M scenes associated with decision of the decision and probabilities respectively corresponding to each prediction sequence. For example, the obtaining unit 610 may obtain 9 prediction sequences ([3,2,5 ] in 9 scenarios described above with reference to fig. 4]，[3,2,3]，[3,2,4]，[3,5,5]，[3,5,3]，[3,5,4]，[3,3,5]，[3,3,3]，[3,3,4]) And a probability pb(s) corresponding to each predicted sequence respectively ₁ ) To pb(s) ₉ ). It is noted that the 9 prediction sequences and corresponding probabilities here are only examples. The obtaining unit 610 may construct a scene tree based on prediction results at a plurality of times generated by the prediction model and the history sequence according to actual application needs, thereby obtaining a plurality of prediction sequences in different scenes and probabilities respectively corresponding to each prediction sequence.

Further, the allocating unit 620 may allocate N specific decision schemes for each of the M scenarios, respectively. Here, the N specific decision schemes are represented by a matrix of positions of the N particles in the computation space, and wherein the position of each of the N particles in the computation space characterizes the corresponding specific decision scheme.

For example, the assigning unit 620 may randomly assign the position vectors of the N particles constituting the position matrix as N specific decision schemes within respective predetermined regions of the M scenes defined by the position constraint, respectively. For example, three scenes s are visually shown in two dimensions in fig. 7 ₁ 、s ₂ 、s ₃ The respective predetermined areas defined by the position constraints, i.e. the feasible areas of the three scenes, are as follows. The feasible field of a scene represents the positions where the N particles in the scene may be located. As shown in the upper diagram of fig. 7, the allocation unit 620 may randomly allocate a position matrix of N particles in the feasible region of the scene. It is to be noted that the upper diagram of fig. 7 only schematically shows the position movement of the particles having the group optimal position for the sake of explanation. The circle filled with parallel oblique lines shown in the upper diagram of fig. 7 schematically represents the initial position of the particles dispensed by the dispensing unit 620.

In addition, the form of the position matrix obtained by the allocation unit 620 is similar to the position matrix shown in the upper right part of fig. 8, and the size of the matrix is [ number of particles, T [ ]]. Since the scene tree structure on the left side of fig. 8 only exemplarily includes predictions of demand at three times, T is 3 in this example, i.e., the size of the position matrix is N,3 in this example]. Thus, the position of each particle in each scene is represented by (x) ₁ ,x ₂ ,x ₃ ) Is defined in which (x) ₁ ,x ₂ ,x ₃ ) Indicating the decision variables that need to be optimized, such as order quantity sequences.

Furthermore, the assigning unit 620 may also assign the velocities to the N particles in each scene, respectively, i.e., generate a velocity matrix of the N particles in each scene. For example, the dispensing unit 620 may randomly dispense the velocity of the particles within a suitable range.

Thus, the allocating unit 620 may obtain a position matrix and a velocity matrix of the N particles in each scene, and provide them to the first calculating unit 630.

Further, each first-layer PSO unit in the first computing unit 630 may obtain a group optimal decision scheme corresponding to each scene for one of the M scenes, respectively. For example, referring to the upper diagram of FIG. 7, at scene s ₁ 、s ₂ 、s ₃ The group optimal positions are represented by a pentagram, and can be obtained by the particle swarm optimization algorithm described above with respect to fig. 5 based on the initial position matrix in each scene.

Specifically, each of the first-tier PSO cells 1 to M may iteratively calculate and update the position matrix and the velocity matrix of the N particles of each of the M scenes a plurality of times using the particle swarm optimization algorithm (using the formula in <4. introduction to particle swarm optimization > above). After the iterative computations are performed for multiple times, the first computing unit 630 may obtain optimal positions (group optimal positions) of all particles at all times in each scene, and select a position matrix including the group optimal position from among the position matrices obtained through the iterations as a group optimal position matrix (group optimal decision scheme).

Each first layer PSO unit may obtain a group optimal position matrix corresponding to each of the M scenes using a first objective function corresponding to the M prediction sequences. For example, the first-layer PSO units may obtain the set of optimal positions in each scenario using formula (1), i.e., f (x), of the order plan case described above with reference to fig. 3, i.e., selecting an order quantity sequence (x) having f (x) as small as possible ₁ ,x ₂ ,x ₃ )。

Referring also to the upper right part of fig. 8, the first calculation unit 630 may derive for each scene s ₁ To s ₉ Group optimal position matrix of

To

(only three are shown), where superscript 1 represents the calculation of the first layer PSO in the first calculation unit.

In addition, the first calculation unit 630 may also obtain a velocity matrix corresponding to a group optimal position matrix in each scene

To is that

And provides the obtained group optimal position matrix and velocity matrix for each scene to the merging unit 640.

Further, the merging unit 640 may obtain a scene combination position matrix P based on the group optimal position matrix of M scenes and M prediction sequences ² . For example, referring to the middle diagram of FIG. 7, scene s ₁ 、s ₂ 、s ₃ Are combined into a scene combination location matrix (represented in the middle graph of fig. 7 by circles filled with grid lines).

The merging unit 640 combines the group optimal position matrices of the M scenes based on the tree structure constructed by the M prediction sequences to obtain a scene combination position matrix.

Specifically, referring to FIG. 8, in the left graph, the non-leaf nodes

At the same time as scene s ₁ ，s ₂ ,s ₃ All the above steps are carried out. While

First layer position matrices respectively corresponding to scenes

Corresponding x ₂ And (4) columns. Due to s ₁ ,s ₂ ,s ₃ Is doing x ₂ In column decision, the values of the parameters are [3,2 ]]. Depending on the same parameters, it is not reasonable that multiple decisions occur when a desired decision is given. Therefore, in the right half of fig. 8, it is necessary to merge the data together

X of (a) ₂ The column is merged as P ² X of (a) ₂₁ The initial value of the column.

Similarly, the operations may be combined by a combination operation

X of ₂ The column is merged as P ² X of ₂₂ Initial value of column, will

X of ₂ The column is merged as P ² X of ₂₃ Initial value of column, will

X of (a) ₁ The column is merged as P ² X of ₁₁ Initial value of column, will

To

X of ₃ Column is respectively taken as P ² X of ₃₁ Is listed to x ₃₉ The initial value of the column. Thus, the merging unit 640 may obtain a scene combination position matrix P ² . In the example of fig. 8, a scene combination position matrix P ² Has a size of [ N,13 ]]Where 13 corresponds to the number of nodes in the scene tree.

According to the numerical type of the decision variables, the specific merging operation strategy is as follows:

■ when x ₂₁ When the decision variable is 0-1, the voting mode (the value frequency of 0 and 1) is adopted for combined value selection,

■ when x ₂₁ When deciding variables for integers, probabilities are used (i.e. scene probabilities pb(s) _i ) ) the mean is nearest, or the mean is rounded up or down,

■ when x ₂₁ For continuous decision of variables, use is made ofThe following formula (4):

and carrying out merging operation.

Here only the merging generation x is shown ₂₁ It should be understood that the same applies to P ² Need to merge other columns generated, e.g. x ₁₁ Column, x ₂₂ Column sum x ₂₃ And (4) columns.

In addition, the merge unit 640 may check P after the merge is completed ² Whether each row (i.e., each particle) in (i.e., whether the position of the particle) satisfies the optimized constraint (i.e., whether the position of the particle satisfies the optimized constraint):

■, when a row (i.e., the position of a particle) satisfies the constraint, no post-processing is done,

■ when a row does not satisfy the constraint condition, the row is reinitialized by sampling in a loop, and the loop stop condition is that the row generated by sampling satisfies the constraint.

In addition, the merging unit 640 may also initialize the velocity matrix provided by the first calculating unit 630 to obtain a combined velocity matrix V ² 。

For example, the merging unit 640 may take the speed matrix as merging

V of (a) ₂ The column is averaged and then based on a set maximum speed parameter v _max And (4) cutting the particle speed. The formulation is described as the following formula (5).

Similarly, the merging unit 640 may be a pair

V of (a) ₂ The above process is performed (averaged and then based on the set maximum velocity parameter v _max The particle velocity is clipped) as V ² V is ₂₂ Initial value of column, will

V is ₂ The column is processed as described above as V ² V is ₂₃ Initial value of column, will

And

v of (a) ₁ The column is processed as described above as V ² V is ₁₁ Initial value of column, will

To is that

V is ₃ Columns are respectively taken as V ² V of (a) ₃₁ Column to v ₃₉ The initial value of the column. Thus, the merging unit 640 may obtain a combined position matrix V ² 。

The merging unit 640 may combine the obtained scenes into a position matrix P ² And combined velocity matrix V ² Is provided to the second calculation unit 650.

It should be noted that although the merging unit 640 is discussed to merge the velocity matrices provided by the first computing unit 630 to obtain a combined velocity matrix, the position matrix P combined with the scene may be directly combined with the scene ² And randomly distributing the speed of the corresponding N particles in a certain range, thereby obtaining a speed matrix of the N particles. In this case, the merging unit 640 may combine the obtained scenes into bitsPlace matrix P ² And randomly assigned velocity matrix V ² Is provided to the second calculation unit 650.

Further, the second calculation unit 650 may be based on the scene position matrix P ² And probabilities to obtain a final decision scheme, i.e., a final decision variable. For example, referring to the lower diagram of fig. 7, the scene combination position matrix is optimized multiple times to get a group optimal position (pentagram).

Specifically, the second-layer PSO unit 651 in the second calculation unit 650 may utilize the particle position and velocity matrix P provided by the merge unit 640 ² And V ² As an initial particle state, finding the optimal positions of all particles of the expected decision corresponding to the scene tree by using a PSO algorithm

Corresponding position and velocity matrix

The second calculation unit 650 may obtain a final decision scheme using a second objective function corresponding to the M prediction sequences and the corresponding probabilities. For example, in the second-level PSO unit 651, the second objective function corresponds to equation (2), i.e., f (x), of the order plan case. The second objective function is used to find the decision variables, such as order quantity sequence, that make F (x) as small as possible

The second-layer PSO unit 651 may optimize the scene combination position matrix for a plurality of iterations according to the second objective function, and select an optimal position from a plurality of scene combination position matrices composed of the scene combination position matrix obtained after each of the plurality of iterations, and a decision variable represented by a particle having the optimal position

Is provided to the decision variable calculation unit 652.

Due to provision to decision variable calculation unit 652Decision variables

Has a size of [1,13 ]]The decision variable calculation unit 652 will therefore perform dimension reduction on the decision variable to obtain a decision variable consistent with the number of actual prediction stages (e.g., 3 in this example), thereby obtaining the final decision scheme.

It is to be noted that the manner in which the decision variable calculation unit 652 performs the dimension reduction processing is not limited. As an example, the decision variable calculation unit 652 may calculate the final desired decision at each time instant according to the following equation (6).

x _t ＝∑ _s∈S pb(s)x _t (s) (6)

In the above formula, x _t (s) represents the decision quantity at all nodes corresponding to the time t, and pb(s) represents the probability corresponding to the nodes. As an example, the decision variable calculation unit 652 may be a general purpose computer

As a decision variable at time t-1

Will be provided with

And

the sum of products of the probabilities at the corresponding nodes is used as a decision variable at time t-2

And will be

And

the sum of the products of the probabilities at the corresponding nodes is used as the decision variable at time t-3

For example, assume that the probabilities of the 9 scenes obtained from the scene tree shown on the left in FIG. 8 are equal probabilities, namely pb(s) ₁ ) To pb(s) ₉ ) 1/9 respectively, then the node

And

the probabilities of (A) are 1/9 respectively, and the node

Is a node

And

the sum of the probabilities is 1/3. Similarly, a node may be obtained

Has a probability of being a node

And

the sum of the probabilities of (c) is 1/3. Thus, a node

The probability of (a) is 1/3.

Thus, the decision variable calculation unit 652 can obtain the decision variables at

t

1, 2, and 3

I.e. the final decision scheme.

Alternatively, in practical applications, the decision variable calculation unit 652 may not obtain the decision variable

The above-mentioned processing is performed.

For example, if the prediction sequence corresponding values at times t-1 and t-2 have become known over time, then the decision variable at time t-3 need only be calculated from the known prediction sequence corresponding values. Taking the above order plan as an example, referring to fig. 8, if the demand at time t-2 has been determined to be 2, it is only necessary for time t-3

And

expectation (1)

And

the sum of products of the probabilities at the corresponding nodes) to determine the desired decision variable (order quantity) at time t-3.

In this case, the decision variable calculation unit 652 may obtain the order quantity at the time t ═ 3, that is, the final decision scheme.

The decision-making decision device according to the present disclosure may form a multi-level calculation structure (a first calculation unit and a second calculation unit) that first optimizes a specific decision-making scheme of each scene by the first calculation unit, and then further optimizes a scene combination decision-making scheme related to the optimized result of each scene by the second calculation unit based on the prediction sequence and the corresponding probability to obtain a final decision-making scheme. Here, through the optimization of the first calculation unit, an optimal decision scheme for each scene may be obtained, and through the optimization of the second calculation unit, a final decision scheme that is optimal for the entire decision-making problem and takes into account the probability of each scene may be obtained, thereby implementing multi-level optimization for the final decision scheme. Also, since PSO is independent of mathematical expression form, the decision-making device 600 according to the present disclosure can be applied to complex large-scale optimization problems.

(example of the processing procedure)

Fig. 9 illustrates a flow chart of a method according to a first specific embodiment of the present disclosure.

As shown in fig. 9, a method according to an embodiment of the present disclosure begins at step S210. In step S210, M prediction sequences in M scenes associated with decision of the decision and probabilities respectively corresponding to each prediction sequence are acquired.

Next, in step S220, a position matrix of N particles is assigned to each scene.

Next, in step S230, a group optimal position matrix corresponding to each scene is obtained for each scene.

Next, in step S240, a scene combination position matrix is obtained based on the group optimal position matrix of M scenes and the M prediction sequences.

Next, in step S250, a final decision scheme is obtained based on the scene combination position matrix and the probability. After that, the process ends.

The above steps of the method according to the embodiment of the present disclosure may be implemented by the decision making device as described above with reference to fig. 1 to 8, for example, and the detailed description has been given above, and will not be repeated here.

The decision-making method according to the present disclosure may enable multi-level optimization of the final decision-making scheme. Also, since PSO is independent of the mathematical expression form, the decision-making method according to the present disclosure can be applied to complex large-scale optimization problems.

<6 > second embodiment

(framework of decision-making device)

The basic framework of the decision-making device according to the second embodiment of the present disclosure is the same as that according to the first embodiment, except that: when the decision-making device initially runs, the initial positions and the speeds of all PSO particles of the first layer are initialized randomly; and the subsequent iteration is iteratively initialized by the optimization result of the second layer of PSO.

The configuration of the decision making means and the flow of processing performed by the decision making means will be described in detail below.

(example of configuration of decision-making device)

An example of the configuration of the decision making device 1000 according to the second specific embodiment of the present disclosure will be described in detail below with reference to fig. 10 to 12. Fig. 10 illustrates a block diagram of a structure of a decision making apparatus 1000 according to a second specific embodiment of the present disclosure, fig. 11 illustrates a schematic diagram of the decision making apparatus 1000 according to the second specific embodiment of the present disclosure, and fig. 12 illustrates a schematic diagram of an operation of a re-allocation unit in the decision making apparatus 1000 according to the second specific embodiment of the present disclosure.

As shown in fig. 10, the decision making apparatus 1000 may include an obtaining unit 1010, an allocating unit 1020, a first calculating unit 1030, a combining unit 1040, a second calculating unit 1050, a reallocating unit 1060, and a sampling unit 1070. Similar to the decision making apparatus 600, the first calculation unit 1030 may include first-layer PSO units 1 to M, and the second calculation unit 1050 may include a second-layer PSO unit 1051 and a decision variable calculation unit 1052.

It is to be noted that the acquisition unit 1010, the allocation unit 1020, the first calculation unit 1030, the merging unit 1040, the second calculation unit 1050, the first-layer PSO units 1 to M, the second-layer PSO units 1051, and the decision variable calculation unit 1052 are similar to the acquisition unit 610, the allocation unit 620, the first calculation unit 630, the merging unit 640, the second calculation unit 650, the first-layer PSO units 1 to M, the second-layer PSO unit 651, and the decision variable calculation unit 652 in fig. 6, and thus descriptions thereof will be appropriately omitted.

The acquisition unit 1010 may acquire M prediction sequences in M scenes associated with decision of a decision and probabilities respectively corresponding to each prediction sequence. For example, the obtaining unit 610 may obtain 9 prediction sequences ([3,2,5 ] for the 9 scenarios described above with reference to fig. 4]，[3,2,3]，[3,2,4]，[3,5,5]，[3,5,3]，[3,5,4]，[3,3,5]，[3,3,3]，[3,3,4]) And a probability pb(s) corresponding to each predicted sequence respectively ₁ ) To pb(s) ₉ )。

Further, the allocating unit 1020 may allocate N specific decision schemes for each of the M scenarios, respectively. Here, the N specific decision schemes are represented by a matrix of positions of the N particles in the computation space, and wherein the position of each of the N particles in the computation space characterizes the corresponding specific decision scheme.

Furthermore, the assigning unit 1020 may assign the speeds to the N particles in each scene respectively, that is, generate a speed matrix of the N particles in each scene.

Thus, the allocation unit 1020 may obtain the position matrix and the velocity matrix of the N particles in each scene and provide them to the first calculation unit 1030.

Further, each first-layer PSO unit in the first computing unit 1030 may obtain a group optimal decision scheme corresponding to each scene for one of the M scenes, respectively. For example, referring to the upper left diagram of FIG. 11, at scene s ₁ 、s ₂ 、s ₃ The group optimal positions are represented by a pentagram, and the group optimal positions (pentagram) in each scene can be obtained by the particle swarm optimization algorithm described above with respect to fig. 5 based on the initial position matrix in each scene (in the upper left diagram of fig. 11, the circles filled with parallel oblique lines schematically represent the initial positions of the particles having the group optimal positions).

Specifically, each of the first-layer PSO cells 1 to M may iteratively calculate and update the position matrix and the velocity matrix of the N particles of each of the M scenes a plurality of times using a particle swarm optimization algorithm (using a formula in <4. introduction to particle swarm optimization >). After performing multiple iterative computations, the first computing unit 1030 may obtain optimal positions (group optimal positions) of all particles at all times in each scene, and select a position matrix including the group optimal position in position matrices obtained through multiple iterations as a group optimal position matrix (group optimal decision scheme).

Each first layer PSO unit may obtain a group optimal position matrix corresponding to each of the M scenes using a first objective function corresponding to the M prediction sequences. For example, the first-layer PSO unit may obtain the group optimal position in each scene by using the formula (1) of the order planning case, i.e. f (x), i.e. selecting the order quantity sequence (x) with f (x) as small as possible ₁ ,x ₂ ,x ₃ )。

Thereby, the first calculation unit 1030 may derive for each scene s ₁ To s ₉ Group optimal position matrix of

To is that

In addition, the first calculation unit 1030 may also obtain a velocity matrix corresponding to a group optimal location matrix in each scene

To

And provides the obtained group optimal position matrix and velocity matrix for each scene to the merging unit 1040.

Further, the merging unit 1040 may obtain the scene combination position matrix P based on the group optimal position matrix of the M scenes and the M prediction sequences ² . For example, referring to the upper right diagram of FIG. 11, scene s ₁ 、s ₂ 、s ₃ Is combined into one scene combination position matrix (indicated by circles filled with grid lines in the upper right diagram of fig. 11). The operation of the merging unit has been described in detail above with respect to fig. 8, which applies equally to the merging unit 1040.

Furthermore, similarly, the merging unit 1040 may also initialize the velocity matrix provided by the first calculating unit 1030 (refer to the initialization of the velocity matrix by the merging unit 640) to obtain the combined velocity matrix V ² And combining the obtained scenes into a position matrix P ² And combined velocity matrix V ² Is provided to a second calculation unit 1050.

Similarly, the merging unit 1040 may directly combine the position matrix P with the scene ² And randomly distributing the speed of the corresponding N particles in a certain range, thereby obtaining a speed matrix of the N particles. In this case, the merging unit 1040 may combine the obtained scenes into the position matrix P ² And randomly assigned velocity matrix V ² To a second calculation unit 1050.

Further, the second-layer PSO unit 1051 in the second calculation unit 1050 may combine the position matrix P based on the scene ² And probabilities to obtain a scene combination location matrix comprising group optimal locations. For example, referring to the bottom right diagram of fig. 11, the scene combination position matrix is optimized multiple times to get a group optimal position (pentagram).

Specifically, the second-layer PSO unit 1051 in the second calculation unit 1050 may utilize the particle position and velocity matrix P provided by the merge unit 1040 ² And V ² As an initial particle state, finding the optimal positions of all particles of the expected decision corresponding to the scene tree by using a PSO algorithm

Corresponding position and velocity matrix

The second layer PSO unit 1051 may use a second target corresponding to the M prediction sequences and corresponding probabilitiesThe function obtains a scene combination position matrix comprising group optimal positions. For example, in the second level PSO unit 1051, the second objective function corresponds to equation (2) of the order plan case, i.e., F (x). The second objective function is used to find the decision variables, such as order quantity sequence, that make F (x) as small as possible

The second-layer PSO unit 1051 may iteratively optimize the scene combination position matrix a plurality of times according to a second objective function, and select an optimal position and a scene combination position matrix including the optimal position from among a plurality of scene combination position matrices composed of the scene combination position matrix obtained after each of the plurality of iterations

In addition, the second layer PSO unit 1051 may combine the scenes including the optimal positions with a position matrix

And corresponding velocity matrix

To a redistribution unit 1060.

Further, the re-allocation unit 1060 may combine the position matrix based on the scene including the optimal position

The position matrix of N particles is reassigned to each scene. For example, referring to the lower left diagram of FIG. 11, a position matrix is combined to a scene s based on the scene including the optimal position ₁ ,s ₂ ,s ₃ The initial states of the N particles for each scene are reassigned.

Specifically, the operation of the re-allocation unit 1060 is described in detail with reference to fig. 12.

The reallocation unit 1060 may allocate a part of a scene combination position matrix including the optimal position as a position matrix corresponding to the scene, respectively, for each scene based on a tree structure constructed by the M prediction sequences.

In particular, the redistribution unit 1060 may distribute the scene according to a scene tree structure

Corresponding columns in (1) are decomposed into a position and velocity matrix P for each scene ¹ (s _i ) And V ¹ (s _i ) And combining P ¹ (s _i ) And V ¹ (s _i ) The position and velocity matrix of the initial particle for the first layer of PSO is iterated as a new round. As an example, FIG. 12 illustrates a second layer PSO particle position matrix

X of ₂₁ The column decomposition process is schematically shown.

As shown in FIG. 12, in the left graph, the non-leaf nodes

At the same time as scene s ₁ ,s ₂ ,s ₃ All the steps are carried out. Then, will

X of ₂₁ Column is disassembled to the corresponding position matrix P of the affiliated scene ¹ (s ₁ )、P ¹ (s ₂ ) And P ¹ (s ₃ ) Corresponding x ₂ And (4) columns. Furthermore, although not shown, non-leaf nodes

At the same time as scene s ₄ ,s ₅ ,s ₆ All the above steps are carried out. Then, will

X of (a) ₂₂ Column is disassembled to the corresponding position matrix P of the affiliated scene ¹ (s ₄ )、P ¹ (s ₅ ) And P ¹ (s ₆ ) Corresponding x ₂ And (4) columns. Further, although not shown, the root node

At the same time as scene s ₁ To s ₉ All the steps are carried out. Then, will

X of ₁₁ Column is disassembled to the corresponding position matrix P of the affiliated scene ¹ (s ₁ ) To P ¹ (s ₉ ) Corresponding x ₁ And (4) columns. Furthermore, although not shown, the leaf nodes

Is a scene s ₁ All the above steps are carried out. Then, will

X of ₃₁ Column is disassembled to the corresponding position matrix P of the affiliated scene ¹ (s ₁ ) Corresponding x ₃ And (4) columns.

According to the rule will

All columns in the scene are assigned to corresponding columns of the corresponding scene particle swarm location matrix. Thus, a position matrix P of N particles of each scene can be obtained ¹ (s _i )。

Further, obtaining the initial matrix position P of the particle position of all scene particle swarm optimization ¹ (s _i ) Then, the sampling unit 1070 may resample the re-allocated position matrix of each scene to obtain a new position matrix of each scene. When the elements in the reallocated position matrix do not comply with the position constraint of the corresponding scene, the elements not complying with the position constraint are resampled, thereby obtaining a new position matrix of each scene. For example, the sampling unit 1070 may randomly assign values to elements that are not subject to the location constraint within respective predetermined regions of the respective scene defined by the location constraint.

Specifically, the sampling unit 1070 may couple P according to the constraint conditions in each scenario ¹ (s _i ) Adjusting the position of the particles corresponding to each row, wherein the specific adjusting steps are as follows:

■ when a row (i.e., the position of a particle) satisfies the constraint (i.e., falls within the feasible region), no post-processing is done,

Thus obtaining the final initial position matrix P of the particle swarm algorithm particles in each scene ¹ (s _i )。

In addition, the particle velocity matrix of the redistribution unit 1060 for the second layer PSO

X of ₂₁ The decomposition performed on the columns is similar to that performed on the position matrix

X of ₂₁ The decomposition performed by the column. According to the decomposition rule

All columns in (a) are assigned to corresponding columns of the corresponding scene particle group velocity matrix. Thus, a velocity matrix V of N particles for each scene can be obtained ¹ (s _i ). Only after decomposition, the sampling unit 1070 uses the set maximum velocity parameter v _max Cutting the particle speed to obtain the final initial speed matrix V of the particle swarm algorithm particles in each scene ¹ (s _i )。

The sampling unit 1070 may apply the processed position matrix P of each scene ¹ (s _i ) And velocity matrix V ¹ (s _i ) Is provided to the first calculation unit 1030 to initialize the PSO of the first layer corresponding scene. For example, referring to the lower left and upper left diagrams of FIG. 11, based on scene s ₁ ,s ₂ ,s ₃ The new position matrix and velocity matrix to initialize the first layer PSO.

Alternatively, the redistribution unit 1060 may alsoRandomly distributing the speed of the N particles of each scene in a certain range to obtain a speed matrix V of the N particles of each scene ¹ (s _i ). In this case, the sampling unit 1070 may assign the randomly assigned velocity matrix V provided by the redistribution unit 1060 ¹ (s _i ) Is provided to the first calculation unit 1030.

The first calculation unit 1030 will calculate a position matrix P for the particles of the input M scenes ¹ (s _i ) And velocity matrix V ¹ (s _i ) And respectively carrying out particle swarm optimization treatment.

That is, the decision making device 1000 repeats the process of obtaining the group optimal position matrix performed by the first calculation unit 1030, the process of obtaining the scene combination position matrix performed by the merging unit 1040, the process of acquiring the scene combination position matrix including the optimal position performed by the second calculation unit 1050, the reallocation process performed by the reallocation unit 1060, and the process of resampling performed by the sampling unit 1070, for the new position matrix and velocity matrix until the set number of repetitions or the variation in the value of the objective function formula (2) is less than a certain threshold, that is, the optimal position converges to a predetermined degree.

Suppose that the decision variable represented by the particle with the optimal position finally obtained by the second-layer PSO unit 1051 in the second calculation unit 1050 is

The second layer PSO unit 1051 will decide the variable

Is supplied to the decision variable calculation unit 1052.

Further, the decision variable calculation unit 1052 derives a final decision scheme by the optimal position obtained after the repeated operations. The operation of the decision variable calculation unit has been described in detail above for the first specific embodiment, and the same applies to the decision variable calculation unit 1052.

The decision-making device 1000 according to the second embodiment may initialize the input of the first stage PSO with the optimization result of the second stage PSO and form a mechanism for mutual initialization between the two stages, so as to optimize the final decision scheme through multiple iterations of calculation.

(example of the Process flow)

Fig. 13 illustrates a flow chart of a method according to a second specific embodiment of the present disclosure.

As shown in fig. 13, a method according to an embodiment of the present disclosure begins at step S310. In step S310, M prediction sequences in M scenes associated with decision of the decision and probabilities respectively corresponding to the prediction sequences are acquired.

Next, in step S320, a position matrix of N particles is assigned for each scene.

Next, in step S330, a group optimal position matrix corresponding to each scene is obtained for each scene.

Next, in step S340, a scene combination position matrix is obtained based on the group optimal position matrix of M scenes and the M prediction sequences.

Next, in step S350, an optimal position and a scene combination position matrix including the optimal position are obtained based on the scene combination position matrix and the probability.

Next, in step S360, it is determined whether the optimum position converges to a predetermined degree or reaches a predetermined number of repetitions.

In the case where it is determined that the optimum position does not converge to the predetermined degree and the predetermined number of repetitions is not reached, in step S370, the position matrix of N particles is newly assigned to each scene.

Next, in step S380, the re-allocated position matrix of each scene is re-sampled to obtain a new position matrix of each scene.

Next, the steps of obtaining the group optimal position matrix in step S330, obtaining the scene combination position matrix in step S340, obtaining the optimal position and the scene combination position matrix including the optimal position in step S350, re-allocation in step S370, and re-sampling in step S380 are repeated for the new position matrix until it is determined in S360 that the optimal position converges to a predetermined degree or a predetermined number of repetitions.

Further, in step S390, a final decision scheme is obtained by the optimal position obtained after repeating the operation a plurality of times.

After that, the process ends.

The above steps of the method according to the embodiment of the present disclosure may be implemented by the decision making means, for example, as described above with reference to fig. 10 to 12, and the detailed description has been given above, and will not be repeated here.

The decision-making method according to the second embodiment may initialize the input of the first-stage PSO with the optimization result of the second-stage PSO and form a mechanism for mutual initialization between the two stages, thereby optimizing the final decision-making scheme through multiple iterative computations.

<7. computer configuration example of decision-making device >

It should be apparent that the various operational procedures of the methods according to the present disclosure may be embodied in the form of computer-executable programs stored in a variety of machine-readable storage media.

Moreover, the object of the present disclosure can also be achieved by: a storage medium storing the above executable program code is directly or indirectly supplied to a system or an apparatus, and a computer or a Central Processing Unit (CPU) in the system or the apparatus reads out and executes the program code. At this time, as long as the system or the apparatus has a function of executing a program, the embodiments of the present disclosure are not limited to the program, and the program may also be in any form, for example, an object program, a program executed by an interpreter, a script program provided to an operating system, or the like.

Such machine-readable storage media include, but are not limited to: various memories and storage units, semiconductor devices, magnetic disk units such as optical, magnetic, and magneto-optical disks, and other media suitable for storing information, etc.

In addition, the computer can also implement the technical solution of the present disclosure by connecting to a corresponding website on the internet, downloading and installing the computer program code according to the present disclosure into the computer and then executing the program.

As shown in fig. 14, a CPU 1401 executes various processes in accordance with a program stored in a Read Only Memory (ROM)1402 or a program loaded from a storage portion 1408 to a Random Access Memory (RAM) 1403. In the RAM 1403, data necessary when the CPU 1401 executes various kinds of processing and the like is also stored as necessary. The CPU 1401, the ROM 1402, and the RAM 1403 are connected to each other via a bus 1404. An input/output interface 1405 is also connected to the bus 1404.

The following components are connected to the input/output interface 1405: an input portion 1406 (including a keyboard, a mouse, and the like), an output portion 1407 (including a display such as a Cathode Ray Tube (CRT), a Liquid Crystal Display (LCD), and the like, and a speaker and the like), a storage portion 1408 (including a hard disk and the like), a communication portion 1409 (including a network interface card such as a LAN card, a modem, and the like). The communication section 1409 performs communication processing via a network such as the internet. The driver 1410 may also be connected to the input/output interface 1405 as necessary. A removable medium 1411 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, or the like is mounted on the drive 1410 as needed, so that a computer program read out therefrom is installed into the storage section 1408 as needed.

In the case where the above-described series of processes is realized by software, a program constituting the software is installed from a network such as the internet or a storage medium such as the removable medium 1411.

It should be understood by those skilled in the art that such a storage medium is not limited to the removable medium 1411 shown in fig. 14 in which the program is stored, distributed separately from the apparatus to provide the program to the user. Examples of the removable medium 1411 include a magnetic disk (including a floppy disk (registered trademark)), an optical disk (including a compact disk read only memory (CD-ROM) and a Digital Versatile Disk (DVD)), a magneto-optical disk (including a Mini Disk (MD) (registered trademark)), and a semiconductor memory. Alternatively, the storage medium may be the ROM 1402, the hard disk included in the storage section 1408, or the like, in which the programs are stored, and which is distributed to the user together with the apparatus including them.

In the systems and methods of the present disclosure, it is apparent that individual components or steps may be broken down and/or recombined. These decompositions and/or recombinations are to be considered equivalents of the present disclosure. Also, the steps of executing the series of processes described above may naturally be executed chronologically in the order described, but need not necessarily be executed chronologically. Some steps may be performed in parallel or independently of each other.

Although the embodiments of the present disclosure have been described in detail with reference to the accompanying drawings, it should be understood that the above-described embodiments are merely illustrative of the present disclosure and do not constitute a limitation of the present disclosure. It will be apparent to those skilled in the art that various modifications and variations can be made in the above-described embodiments without departing from the spirit and scope of the disclosure. Accordingly, the scope of the disclosure is to be defined only by the claims appended hereto, and by their equivalents.

<8. supplementary notes >

With respect to the embodiments including the above embodiments, the following remarks are also disclosed:

supplementary note 1. an apparatus for deciding a decision, comprising:

an acquisition unit that acquires M prediction sequences in M scenes associated with the decision of the decision and probabilities corresponding to the respective prediction sequences;

an allocation unit that allocates N specific decision schemes for each scene, respectively;

a first calculation unit that obtains, for each scene, a group optimal decision scheme corresponding to each scene;

a merging unit that obtains a scene combination decision scheme based on the group optimal decision scheme of the M scenes and the M prediction sequences; and

a second calculation unit that obtains a final decision scheme based on the scenario combination decision scheme and the probability,

wherein M and N are natural numbers greater than 0.

2. The apparatus of claim 1, wherein,

the predicted sequence is a demand sequence related to demand for a particular item or transaction, and deciding the decision refers to deciding a decision variable that optimizes a particular objective related to the particular item or transaction based on the M demand sequences and corresponding probabilities.

Note 3. the apparatus according to

note

1 or 2, wherein,

representing the N specific decision schemes with a matrix of positions of N particles in a computation space, and wherein the position of each of the N particles in the computation space characterizes the corresponding specific decision scheme.

Supplementary note 4. the apparatus according to supplementary note 3, wherein,

the first calculation unit obtains a group optimal position matrix corresponding to each of the M scenes using a first objective function corresponding to the M prediction sequences, an

The second calculation unit obtains the final decision scheme using a second objective function corresponding to the M prediction sequences and corresponding probabilities.

Supplementary note 5. the apparatus according to supplementary note 4, wherein the second calculation unit optimizes a scene combination position matrix by the second objective function for a plurality of iterations, and selects an optimal position among a plurality of scene combination position matrices composed of the scene combination position matrix obtained after each of the plurality of iterations, and obtains the final decision scheme by the optimal position.

Supplementary note 6. the apparatus according to supplementary note 4, wherein the second calculation unit optimizes the scene combination position matrix by the second objective function for a plurality of iterations, and selects an optimum position and a scene combination position matrix including the optimum position among a plurality of scene combination position matrices composed of the scene combination position matrix obtained after each of the plurality of iterations, and

the device further comprises:

a reassignment unit that reassigns a position matrix of N particles to each scene based on a scene combination position matrix including the optimal position; and

a sampling unit for resampling the re-distributed position matrix of each scene to obtain a new position matrix of each scene, and

the apparatus repeats the following process for the new position matrix: a process of obtaining a group optimal position matrix performed by the first calculation unit, a process of obtaining a scene combination position matrix performed by the merging unit, a process of acquiring a scene combination position matrix including the optimal position performed by the second calculation unit, a re-allocation process performed by the re-allocation unit, and a re-sampling process performed by the sampling unit,

until the optimum position converges to a predetermined degree or a predetermined number of repetitions, and

the second calculation unit derives the final decision scheme from the optimal position obtained after stopping the above-described repetitive processing.

Note 7 the apparatus according to note 6, wherein the reallocation unit allocates a part of a scene combination position matrix including the optimal position as a position matrix corresponding to a scene, respectively, for each scene based on a tree structure constructed by the M prediction sequences.

Note 8. the apparatus according to note 3, wherein the merging unit combines the group optimal position matrices of the M scenes based on a tree structure constructed by the M prediction sequences to obtain a scene combination position matrix.

Note 9. a method for deciding a decision, comprising:

obtaining M prediction sequences under M scenes associated with the decision of the decision and probabilities respectively corresponding to each prediction sequence;

respectively distributing N specific decision schemes for each scene;

obtaining a group optimal decision scheme corresponding to each scene for each scene;

obtaining a scene combination decision scheme based on the group optimal decision scheme for the M scenes and the M prediction sequences; and

obtaining a final decision scheme based on the scenario combination decision scheme and the probability,

wherein M and N are natural numbers greater than 0.

Supplementary note 10. the method according to supplementary note 9, wherein,

the N specific decision schemes are represented by a matrix of positions of N particles in a computation space, and wherein a position of each of the N particles in the computation space characterizes the corresponding specific decision scheme.

Supplementary note 11. the method according to supplementary note 10, wherein the step of assigning N specific decision schemes for each scene comprises: randomly assigning position vectors of the N particles constituting the position matrix as the N specific decision schemes within respective predetermined regions of the M scenes defined by position constraints, respectively.

Reference 12. the method according to reference 10, wherein,

obtaining a set optimal position matrix corresponding to each of the M scenes using a first objective function corresponding to the M prediction sequences, an

Obtaining the final decision scheme using a second objective function corresponding to the M prediction sequences and corresponding probabilities based on the set of optimal location matrices.

Reference 13. the method of reference 12, wherein the step of obtaining the final decision scheme comprises:

and iteratively optimizing a scene combination position matrix for a plurality of times according to the second objective function, selecting an optimal position from a plurality of scene combination position matrixes consisting of the scene combination position matrix obtained after each iteration of the plurality of iterations, and obtaining the final decision scheme through the optimal position.

Reference 14. the method of reference 12, wherein the step of obtaining the final decision scheme comprises:

iteratively optimizing a scene combination position matrix a plurality of times in accordance with the second objective function, and selecting an optimal position and a scene combination position matrix including the optimal position among a plurality of scene combination position matrices composed of the scene combination position matrix obtained after each of the plurality of iterations, and

the method further comprises the following steps:

reassigning a location matrix of N particles to each scene based on a scene combination location matrix including the optimal location;

resampling the redistributed position matrix of each scene to obtain a new position matrix of each scene; and

repeating the steps of obtaining a group optimal position matrix, obtaining a scene combination position matrix including the optimal position, reallocating, and resampling for a new position matrix until the optimal position converges to a predetermined degree or a predetermined number of repetitions, and

the step of obtaining the final decision scheme further comprises: the final decision scheme is derived from the optimal position obtained after a number of iterations.

Supplementary note 15. the method according to supplementary note 14, wherein the step of reassigning the position matrix of N particles to each scene comprises:

a part of a scene combination position matrix including the optimal position is allocated as a position matrix corresponding to a scene for each scene, respectively, based on a tree structure constructed by the M prediction sequences.

Supplementary note 16. the method according to supplementary note 14, wherein the step of resampling comprises:

when the elements in the reallocated position matrix do not comply with the position constraint of the corresponding scene, the elements not complying with the position constraint are resampled, thereby obtaining a new position matrix of each scene.

Supplementary note 17 the method according to supplementary note 16, wherein resampling the elements not subject to the position constraint comprises randomly assigning values to the elements not subject to the position constraint within respective predetermined regions of the respective scenes defined by the position constraint.

Supplementary note 18. the method according to supplementary note 12, wherein the step of obtaining a group optimal position matrix for each scene comprises:

for each of the M scenes, iteratively optimizing a position matrix of the N particles a plurality of times according to the first objective function, and selecting an optimal position from a plurality of position matrices consisting of position matrices obtained after each of the plurality of iterations; and

and acquiring a position matrix comprising the optimal position as a group optimal position matrix of each scene.

Supplementary note 19. the method according to supplementary note 10, wherein the step of obtaining a scene combination position matrix comprises:

combining the set optimal position matrices of the M scenes based on a tree structure constructed by the M prediction sequences to obtain a scene combination position matrix.

Reference 20 a machine-readable storage medium having embodied thereon a program product comprising machine-readable instruction code stored therein, wherein the instruction code, when read and executed by a computer, is capable of causing the computer to perform the method according to reference 9-19.

Claims

1. An apparatus for deciding a decision, comprising:

wherein M and N are natural numbers greater than 0.

2. The apparatus of claim 1, wherein,

3. The apparatus of claim 1 or 2,

4. The apparatus of claim 3, wherein,

5. The apparatus according to claim 4, wherein the second calculation unit optimizes a scene combination position matrix for a plurality of iterations depending on the second objective function, and selects an optimal position among a plurality of scene combination position matrices composed of the scene combination position matrix obtained after each of the plurality of iterations, and obtains the final decision scheme by the optimal position.

6. The apparatus according to claim 4, wherein the second calculation unit optimizes a scene combination position matrix in accordance with the second objective function for a plurality of iterations, and selects an optimum position and a scene combination position matrix including the optimum position among a plurality of scene combination position matrices composed of scene combination position matrices obtained after each of the plurality of iterations, and

the device further comprises:

the second calculation unit obtains the final decision scheme from the optimum position obtained after stopping the above-described repetitive processing.

7. The apparatus according to claim 6, wherein the reallocation unit allocates a part of a scene combination position matrix including the optimal position as a position matrix corresponding to a scene, respectively, for each scene based on a tree structure constructed by the M prediction sequences.

8. The apparatus of claim 3, wherein the merging unit combines the group optimal position matrices of the M scenes based on a tree structure constructed through the M prediction sequences to obtain a scene combination position matrix.

9. A method for deciding a decision, comprising:

respectively distributing N specific decision schemes for each scene;

wherein M and N are natural numbers greater than 0.

10. A machine-readable storage medium having embodied thereon a program product comprising machine-readable instruction code stored therein, wherein the instruction code, when read and executed by a computer, is capable of causing the computer to perform the method of claim 9.