JPH09311895A - Distribution schedule support system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To decide how to set the assigned area of each driver to be most efficient. SOLUTION: Through the use of a two-layer probability field model defining the state of a lower layer by the set of distribution schedules in distribution examples for learning, defining the energy function of the lower layer by a limiting item, a cost item and a matching item and defining the limiting condition satisfying degree of an upper layer by a reaching degree to the desired value of each item of these three kinds, the most efficient area of the driver is decided with respect to prepared distributing destination data. A parameter in the skill function of the driver is previously learned from the distribution example for learning (a skill function deciding part 4), the correction by a system operator is added (a parameter finally deciding part 5) if necessary, and the schedule of the distribution example being a planning object is generated (a distribution schedule deciding part 3). The deciding part 3 is given an energy function and automatically adjusts weight to obtain an optimum solution.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a delivery plan support system in which a driver's specialty is taken into consideration, and more specifically, a delivery plan problem is formulated as a certain function (energy function) minimization problem and simulated. The present invention relates to a learning method for preliminarily learning the optimum sharing area of each driver from past delivery cases in a delivery planning system that searches for an optimal solution using an annealing method (hereinafter, SA method).

[0002]

2. Description of the Related Art In the case of a delivery planning problem, for example, when a package such as food, beverages, clothes, sundries, and raw materials is delivered to a restaurant, a store, a factory, or a general household by a truck,
It is necessary to formulate an optimal delivery plan, but it is especially difficult for all drivers to become familiar with the roads and geography of the entire delivery area of the delivery center because the delivery area is wide such as the entire 23 wards of Tokyo. Is. Conventionally, many inventions related to delivery planning have been proposed, but there is no proposal that reasonably considers the driver's specialty area or unfamiliar area, and if the driver is familiar with which area, the total efficiency is the most efficient. No delivery plan support system has been proposed that has a means for determining whether there is.

(Prior Art 1) For example, JP-A-6-176.
Japanese Patent No. 040 (delivery scheduling device) proposes a solution algorithm for a delivery planning problem based on a typical two-step method. The vehicle is temporarily dispatched to the delivery destination of the minutes, and the driver's specialty area is not taken into consideration at all.

(Prior Art 2) Further, Japanese Patent Application No. 6-0000 (a method and apparatus for finding an optimal solution for a delivery planning problem) proposed by the present applicant based on a simulated annealing method. Inventions relating to a method of solving a delivery planning problem and its apparatus have been proposed. The invention previously proposed by the applicant is characterized in that it is provided with an automatic adjuster for weighting factors in the energy function. Like the automatic adjustment mechanism of this weighting coefficient, it is also based on the idea of the two-layer random field model. However, the invention previously proposed by the applicant does not consider the driver's specialty area at all.

[0005]

The applicant of the present invention has previously found that in Japanese Patent Application No. 6-000 (a method and apparatus for finding an optimal solution for a delivery planning problem), a system operator finds a solution having a desired property. In order to obtain it, we have proposed an optimal solution search method and its apparatus for the delivery planning problem, which is equipped with a mechanism for automatically adjusting the weighting coefficient in the energy function. However, in actual large-scale delivery cases, it is often a big problem whether or not the truck driver who is involved in the actual delivery work is familiar with the geography of a given delivery destination area.
This is because, as the information about the delivery destination, there are many cases where the delivery is instructed by giving only the delivery destination name, telephone number, and address described in the delivery instruction slip. Therefore, normally, when planning a delivery plan,
As a general rule, each driver's area is set in advance, and each driver is assigned only the delivery destination included in the area. If each driver's specialty area is known, even in the invention previously proposed by the applicant, by adding an evaluation term (hereinafter, matching term) considering the driver's specialty area as an energy function. Although it is possible to respond to a certain extent, the function to determine the sharing area of each driver most efficiently and reasonably in the previous application and in the patents and research announcements filed by other companies so far. No delivery planning system with a

The present invention has been made in view of the above-mentioned circumstances, and an object of the invention of claim 1 is to determine how to set the sharing area of each driver to be most efficient. The invention of claim 2 is most effective as a total of a plurality of cases by preparing a plurality of past delivery cases and using these delivery cases as learning data when determining the driver's sharing area. According to the invention of claim 3, which is intended to allocate a shared area with a certain degree, an analog function is used to finely express how familiar the driver is with regard to knowledge about the road at a certain point, and the shared area is evaluated. The purpose is to do more appropriately.

It is an object of the invention of claim 4 to express even when the shape of the area in which the driver is good is not a simple circular shape but a complicated shape or a plurality of discontinuous areas. According to the invention of claim 5, in reality, there are experienced drivers who have a large number of familiar areas and conversely, there are beginner drivers who have a small number of familiar areas. The invention of claim 6, which aims to reflect the skill level of the driver by increasing or decreasing the number of learning functions according to the above, does not require a driver to change the assigned area. The purpose of this is to reduce the calculation time during learning by removing the driver's proficiency level function from the learning target when no change is desired.

According to the invention of claim 7, for a driver,
Exclude some of the existing areas of expertise from the learning target, that is, regarding some areas of the area of expertise where the driver is familiar with the geography, as experience know-how, do not change at the time of learning and plan as it is. The purpose is to utilize.

According to the eighth aspect of the present invention, the learning target is specifically set such that learning is performed so that all delivery destinations are assigned within a circle area having a radius of 200 km according to the driver's ability. The invention of claim 9 which is intended to allow the operator to set it freely, aims to prevent the proficiency function from converging to a meaningless state during learning. Is a system operator who modifies the proficiency function after learning, confirms the effects of the delivery cases used as learning data and the delivery cases targeted for planning, and finally the proficiency function of each driver. The purpose is to determine.

[0010]

According to a first aspect of the present invention, the state of the lower layer is defined by a set of delivery plans in a delivery case for learning, and the energy function of the lower layer is defined by a constraint term, a cost term, and a matching term, The upper limit constraint satisfaction is these 3
The two-layer random field model defined by the degree of reaching the target value of each term of the type is used to determine the most efficient driver sharing area for the delivery destination data prepared in advance, Therefore, it is decided how to set the allocation area of each driver to be most efficient.

According to a second aspect of the present invention, in the first aspect of the present invention, it is possible to determine a driver sharing area so that a plurality of delivery cases (for example, a plurality of past delivery cases) can be delivered most efficiently. Therefore, when deciding the driver's sharing area, multiple past delivery cases are prepared and these delivery cases are used as learning data. Allotment of shared areas is made.

According to the invention of claim 3, in claim 1, as a matching term, the degree to which the driver is familiar with an area is "normalized" (a value obtained by integrating in a defined area is a normalized value) such as a normal distribution function. It is characterized in that it is expressed as a mountain-shaped function (a function that becomes a certain constant value) (hereinafter referred to as a proficiency function), so that the driver's level of knowledge regarding the road at a certain point can be determined. It is an analog function that is expressed in detail to allow more appropriate evaluation of shared areas.

According to a fourth aspect of the present invention, in the third aspect, each driver's specialty area is expressed by a plurality of proficiency functions. It is designed so that it can be expressed even when it has a complicated shape instead of a simple circular shape or a plurality of discontinuous areas.

According to a fifth aspect of the present invention, in the fourth aspect, a skilled driver who has a long delivery experience and a beginner driver who has little delivery experience and who is not familiar with geography are represented by the number of proficiency functions belonging to each driver. That is, that is, the skilled driver is characterized by many areas that are familiar, and conversely, that there are few beginner drivers is represented by the number of proficiency functions, so in reality, On the contrary, there are skilled drivers with a large number of familiar areas and conversely beginner drivers with a small number of familiar areas, so by increasing or decreasing the number of learning functions according to the experience of each driver, It is designed to reflect the skill level of the driver.

According to a sixth aspect of the present invention, in the first aspect, a driver who learns a proficiency level function and a driver who does not learn the learning function and is fixed can be arbitrarily selected.
Therefore, when it is not necessary or desired to change the assigned area for a driver, the proficiency function of this driver is removed from the learning target to reduce the calculation time during learning. It is something that is done.

The invention of claim 7 is characterized in that, in claim 1, the proficiency function learned by each driver and the proficiency function not learned by each driver can be arbitrarily selected.
Therefore, for a certain driver, a part of the existing good area is excluded from the learning target. That is, the driver's shared area is represented by a plurality of proficiency functions, but some proficiency functions are removed from the learning target and fixed,
Other proficiency functions were targeted for learning, and some of the areas of expertise in which the driver is familiar with the geography were used as-is as experience-based know-how without making any changes at the time of learning, and were used directly for planning. It is a thing.

According to an eighth aspect of the present invention, in the first aspect, the degree of matching between the driver's shared area (expressed by a proficiency level function) and the delivery destination is arbitrarily set by the system operator with the learning level of the proficiency level function as a target value. The goal of learning is specified according to the driver's ability, for example, learning is performed so that all delivery destinations are assigned within a circle area with a radius of 200 km. It can be freely set by the system operator.

The invention of claim 9 is characterized in that, in claim 3, the parameters such as the center and the spread of the proficiency function are set to be restricted. Therefore, at the time of learning, the proficiency function is meaningless. This is to prevent convergence to the state. For example, when assigning a plurality of proficiency functions to a driver, the centers of the proficiency functions should not be close to each other within a certain range to prevent two proficiency functions from overlapping. It is the one.

According to a tenth aspect of the present invention, a learning delivery case or a delivery case to be planned is simulated by using a learning function or a proficiency function further modified by a system operator. According to the above, while comparing and confirming the effect of learning of the proficiency function and the effect of modification by the system operator, the area in which each driver is assigned is finally determined, and thus, the proficiency after learning is learned. The system operator corrects the degree function, confirms the effect of the delivery case used as learning data or the delivery case that is the planning target, and finally determines the proficiency function of each driver. This allows the system operator to interactively modify the proficiency function after learning by adding his or her experience know-how. Those were Unishi.

[0020]

1 is a schematic block diagram of a delivery planning system having a learning function determining function (learning function) according to the present invention. In FIG. 1, 1 is a learning function determining function (learning function). 2) is a control unit for searching for an optimal solution of a delivery problem (determination of delivery plan), and 2 is an input unit for data required for search and learning. Reference numeral 3 is a delivery plan determining unit that determines a delivery plan, 4 is a proficiency function determining unit that determines a driver's proficiency function, and 5 is one of the proficiency functions calculated by learning in the proficiency function determining unit 4. The system operator manually corrects and adjusts the parameter, and the effect is calculated by the delivery plan determination unit 3
This is a parameter final decision unit that finally decides the parameters in the proficiency function while simulating with. The delivery plan determination unit 3 uses a proficiency level function learned in advance by the proficiency level function determination unit 4 or a proficiency level function finalized by further modification in the parameter final determination unit 5. The calculation result output unit 6 is a calculation result output unit that outputs the optimum solution obtained by the delivery plan determination unit 3 and the proficiency level function obtained by the proficiency level determination unit 4.

FIG. 2 shows the 3 delivery plan determining unit 3 shown in FIG.
11 is a block diagram of a delivery plan sheet for generating a delivery plan by repeating trial transformation based on a random number.
Is a decision unit that decides the cost term, constraint term, and matching term that make up the energy function. Reference numeral 13 is an arithmetic unit that defines an energy function based on the information created and determined by the delivery plan drafting unit 11 and the energy function decision unit 12 and calculates the energy value of the delivery plan. Reference numeral 14 is an energy minimization unit that obtains the minimum state of the energy function defined by the calculation unit 13 by using the simulated annealing method, and 15 is an energy function in the energy minimization process of the energy function minimization unit 14. It is an adjusting unit that adjusts the value of the weighting coefficient of the.

(Overall Process Flow) The parameters in the driver's proficiency function are learned in advance from the learning delivery case (proficiency function determining unit 4 in FIG. 1). If necessary, the system operator is modified (the parameter final determination unit 5 in FIG. 1) to make a plan for the delivery case to be planned (delivery plan determination unit 3 in FIG. 1). The delivery plan determination unit 3 in FIG. 1 is given an energy function and automatically adjusts weights to obtain an optimum solution. The search method therefor has already been proposed in Prior Art 2 and will not be described in detail here.

(Explanation of the Invention Using Examples) Various kinds of energy functions can be considered depending on the definition of the delivery planning problem to be solved. Here, as an example of the delivery planning problem, [Example 1] "I want to deliver to N destinations in the shortest total delivery time. However, the number of vehicles used is minimized and the delivery time of each vehicle is reduced. I want to make it as even as possible. Some of the destination stores have a designated delivery time in the morning, and all vehicles must depart from the same depot and return to this depot after the delivery is complete. ” Think about the problem. However, the delivery time here is
It is the time until one vehicle departs from the depot and returns. Now, assume that the delivery plan x is expressed by a one-dimensional matrix as shown in FIG. 3 (the same applies to other solution expressions such as a variable-length two-dimensional list in which delivery destinations are arranged in order of delivery for each vehicle).

FIG. 3 shows an example in which three vehicles deliver to nine delivery destinations. The nine delivery destinations are represented by numbers 1 to 9, and the depot (package delivery center) is represented by 0. The number of 0 is the maximum number of vehicles + 1.
In this embodiment, all vehicles receive parcels from the depot and deliver according to the given delivery order. However, it is assumed that the number of vehicles used is not fixed and only the maximum number is given. In the example of FIG. 3, the maximum number of vehicles = 3. From the one-dimensional matrix shown in FIG. 3B, the symbol string between 0 and 0 represents the allocation (vehicle allocation) of the delivery destinations for one vehicle and the delivery order. If 0 and 0 are consecutive, it means that there is an unused vehicle,
This represents a plan that reduces the number of vehicles used by one. An example of the initial state is also shown in FIG.

FIG. 4 is a block diagram of the proficiency function determining unit 4 of FIG. 1. In the drawing, reference numeral 21 is a creating unit which repeats trial transformation based on a random number to generate a total delivery plan in a delivery case for learning. is there. Reference numeral 22 is a determination unit that determines the cost term, the constraint term, and the matching term that make up the total energy function. 23 is an operation for defining the total energy function based on the information created and determined by the total delivery plan drafting unit 21 and the total energy function decision unit 22 and calculating the total energy value of the delivery plan in the delivery case for learning It is a department. Reference numeral 24 is a total energy minimization unit that obtains the minimum state of the total energy function defined by the calculation unit 23 by using the simulated annealing method. Reference numeral 25 is a parameter adjusting unit that adjusts the weighting coefficient in the total energy function and the value of the parameter included in the proficiency function to appropriate values in the process of minimizing the total energy of the total energy function minimizing unit 24.

The proficiency function determining unit (the proficiency function determining unit 4 in FIG. 1) represented by the block diagram of FIG. 4 is the core of the present invention. The proficiency function determining unit will be described in detail below. First, N _ds sets of delivery cases for learning are prepared, and a delivery plan for each delivery case is x _l (l = 1, ...,
N _ds ). The delivery plan (hereinafter, total delivery plan) in all learning delivery cases is represented by X _total = {x _l }, and X
_total energy function E _total defined above
(Hereinafter, total energy function)

[0027]

[Equation 1]

Defined by α _l is a weight of each learning data set (delivery case data), and is adjusted and fixed according to the importance of the data set (all will be described below as 1). E _l is an energy function defined on the delivery plan x _l in the l-th learning delivery case. The energy function represents the degree to which the delivery plan is undesirable. Here, the following energy function E _l corresponding to Example 1 will be described.

[0029]

[Equation 2]

{ _{L, i} E _cnst } (i = 1 _{, ...,} 4) is called a constraint term, _l E _cost is called a cost term, and _l E _match is called a matching term.

The constraint term is a term that quantitatively expresses how much the constraint imposed on the delivery plan is satisfied.
Alternatively, it is a term in which the delivery planner specifically sets the request in detail. On the other hand, the cost term is a term that expresses the cost desired to be minimized as much as possible, although the requirement cannot be specifically expressed. For example, the conditions for specifying delivery time are
If you are 5 minutes late, you will be a little troubled.
Since it is possible to express the demands of the delivery planner quantitatively, such that it is absolutely troublesome if the delay is 5 minutes or more, it can be included in the constraint term. As for the total delivery time, it is often desirable that the delivery time is as short as possible without any specific request from the delivery planner, so the cost term is used in this example. In the following, the cost term coefficient b _l is set to b _l = 1. The last matching term is originally a term that should be included in the constraint term, but here it is specially named as a matching term. Here, with respect to the delivery plan generated by the system, the degree of matching between each driver's specialty area and the delivery destination actually assigned is shown. Below, the example of the constraint term, the cost term, and the matching term corresponding to Example 1 is described.

(Example of constraint term) _{l, 1} E _cnst (constraint regarding designation of delivery time):

[0033]

(Equation 3)

If a certain solution satisfies all delivery time designation conditions, N _no = 0, so N _no / N _time = 0.
If all are not satisfied, N _no = N _time, so N _no / N _time = 1. In addition, the value of N _no / N _time increases in proportion to the ratio of the number of unsatisfied houses. Therefore, the expression of the degree of violation of the delivery time designation condition is more detailed than that of the related art 2.

In the present invention, the expression described in the right column of Table 1 should be used originally, but the expression described in the left column is used for the convenience of electronic application processing. In Table 1, the expressions in the left column and the expressions in the right column mean the same thing.

[0036]

[Table 1]

_{L, 2} E _cnst (restriction on loading weight): The sum of excess amounts with respect to the upper limit of loading weight is used as the size of the penalty. That is, in the delivery plan x _l , if the loading weight of the vehicle i is w _i (x _l ) and the maximum loading weight of the vehicle _i is _max w _i ,

[0038]

(Equation 4)

It is defined by However, M _l is the delivery plan x _l
It is the number of vehicles used in.

_{L, 3} E _cnst (restrictions on the number of vehicles used):
The value of M ₁ is used as the number of vehicles used in the delivery plan.

_{L, 4} E _cnst (constraint _regarding deviation of delivery time): The value of standard deviation of delivery time used in the delivery plan is used.

(Example of cost term) _{l, 1} E _cost (term of total delivery time): The total time from the depot departure to the return of the vehicle used in the delivery plan x _l is used.

(Example of matching item) _l E _match (item showing how much the assigned delivery destination matches the driver's specialty area):

[0044]

(Equation 5)

Is as follows. The following discussion is based on the proficiency function h (x
_l ) is a "normalized function" such that the value integrated in a certain area is a constant value, even if it is not a normal distribution function, and it holds if it is a two-dimensional function with a chevron shape. .

Learning the proficiency function is learning the optimum values of the parameters {о _{m, n} } and {б _{m, n} } of the equation (5) using the learning delivery case. A two-layer random field model is also used for this learning. The parameter learning mechanism using the two-layer random field model will be described below. The two-layer random field model defines two layers. In this case, the lower layer state ω _low is the total delivery plan X _total = {x _l } itself, and the upper layer state ω _up is the request from the system operator (delivery planner) to the lower layer total delivery plan X _total. A vector expressing the degree of satisfaction ω _up ＝
(Y ₁ , y ₂ , y ₃ , y ₄ , y ₅ ) (0 ≦ y ₁ ≦ 1).

Energy function E _l of each delivery case for learning
For each constraint term and matching term included in, the demand of the delivery planner will be expressed in detail with an analog function. A function expressing requirements for the constraint term and the matching term included in the energy function E _l is referred to as “constraint target value function” { _(i) R _l } (i = 1, ..., 5). The vector element y _i of the upper layer state ω _up is defined by _(i) R _l (X _total ).

The constraint target value function { _(i) R _l } is the delivery plan x
_From 0 to 1 how much the _l satisfies the constraint conditions described in the constraint term and the matching term
This is a function expressed using the analog function of. The closer the value is to 1, the more satisfied the requirements of the delivery planner are
The closer to, the less the demand is met.
In addition, when the constraint condition expressed by the constraint term is different for each delivery vehicle, delivery destination, or learning delivery case,
This can be dealt with by defining a constraint target value function for each delivery vehicle, delivery destination, or learning delivery case.

An example of the constraint target value function is shown below. (Example of Constraint Target Value Function) ₍₁₎ R _{l, k} (constraint target value function relating to time zone designation of delivery time): Defined for each delivery destination k (k = 1, ..., N _l ).
The constraint target value function is a function that takes an analog value of 0 to 1. In the example shown in FIG. 5, for the delivery plan x _l ,
The time t _k (x _l ) at the delivery destination k is the specified time zone [
_{low er} t _k , _upper t _k ], Δt = t
_{k (x l) -upper t k} , if you arrive early, Δt = t _k
It is defined as (x _l ) _{-low er} t _k . Δt represents the amount of violation regarding the arrival time (deviation from the specified time zone), but in the example of FIG. 5, if | Δt | ≦ Δt ₁ , the deviation from the specified time zone is allowed, and | Δt | If> Δt _2, it means that the delivery planner completely refuses. The constraint target value function relating to the arrival time constraint is not limited to that shown in FIG. 5, and can be freely defined by the delivery planner according to the targeted delivery problem. The method described in the present invention can be applied in exactly the same manner as long as it is an arbitrary function that takes an analog value of 0 to 1.

₍₂₎ R _{l, m} (constraint target value function regarding upper limit of loaded weight): Defined for each vehicle m (m = 1, ..., M _l ). FIG. 6 shows an example. In FIG. 6, the loaded weight is 3.
It expresses the request of the delivery planner that the ideal is 8 tons or less and that the delivery plan that is 4.0 tons or more is never accepted.

₍₃₎ R ₁ (constraint target value function relating to the number of vehicles used): FIG. 7 shows an example thereof. In the example of FIG. 7, the minimum required vehicle platform is estimated to be M _min vehicles from the total weight of delivered luggage and the upper limit of the loaded weight per vehicle, and the upper limit of the number of vehicles that can be used is M _max vehicles in advance. If you know. The smaller the number of vehicles used, the better the delivery plan. ₍₄₎ R _l (constraint target value function regarding deviation of delivery time (target deviation) for each vehicle): FIG. 8 shows an example thereof. If the delivery planner requests a plan with a very small deviation in delivery time, the width of SD ₁ and SD ₂ may be narrowed and the value of SD ₁ may be reduced. For example, SD ₁ =
If you set 10, SD ₂ = 30 (min), the standard deviation is 1
It is desirable that the delivery plan be 0 minutes or less, and that the delivery plan that is 30 minutes or more is absolutely unacceptable. On the contrary, if the deviation of the delivery time does not become a problem and other items such as the total delivery time, the number of vehicles used, and the delivery time designation are to be prioritized, the values of SD ₁ and SD ₂ may be increased. According to the situation of the delivery problem, the delivery planner can determine the desired value (SD) of the deviation (standard deviation) of delivery times.
₁ ) and the upper limit (SD ₂ ) allowed are freely expressed.

₍₅₎ R _{l, m} (each driver m in the learning delivery case l
For each delivery destination k (εΩ (x _l , m)) for each of the positions P _l located within a circle within a radius γσ _n from the center of the proficiency function (n) of the driver m Function expressing the request of the delivery planner): However, γ is a certain positive constant, for example, 2.0.

In the present invention, as described above, the delivery planner can express the requirement (desired value and allowable upper limit value) for each constraint term in the form of a constraint target value function. Upper-layer state vector ω _up = (y ₁ , y ₂ , y ₃ , y ₄ , y ₅ )
Is obtained from equations (7) to (11) using these constraint objective functions { _(i) R _l }.

[0054]

(Equation 6)

(Parameter updating rule) The two-layer random field model and the parameter updating rule have already been disclosed in Japanese Patent Laid-Open No. 5-12.
No. 0252, “Control Device” and “Determination of Parameter Value in Energy Function in Simulated Annealing Method” (IEICE Transactions D-II, Vol.
J75D-II, No. 7, pp1232-1240 (1
992)). Although the traveling literature is dealt with in the above-mentioned publication, the present invention is applied to the learning of a proficiency function.

First, the satisfaction degree of the constraint condition П ({θ _i }; T)
Is defined as follows. “Satisfaction with constraints” means S
It is a time-average value of the values that represent the degree of satisfaction of the constraint condition (delivery plan) of the system appearing in the thermal equilibrium state at a certain temperature of the method A with an analog value of 0 to 1. An example of a function that satisfies this definition is given below.

[0057]

(Equation 7)

Here, П ({θ _i }; T) is the satisfaction degree of the constraint condition, and the temperature T of the system and parameters {θ _i } = {a _lj }, {such as weighting coefficient in the total energy function. c _l },
It is a function with {ο _{m, n} }, {σ _{m, n} }). Function f on the right side
(U, v) is a function representing the degree of similarity between the two vectors u and v. As a function representing the similarity, for example, an inner product between vectors is used. Further, <> represents a time average operation. ₀ ω _{up in} Expression (12) is a vector (∀i, y _i = 1) in which all the elements are 1. Formula (7)-
As can be seen from the definitions of (11) and the constraint target value function { _(i) R _l }, ₀ ω _up completely satisfies all the requirements of the system operator represented by { _(i) R _l }. It represents the state. The upper layer state ₀ ω _up is the lower layer state ω _low = X
It is calculated from _total = {x _l } using the constrained target value function { _(i) R _l } and equations (7) to (11), but the state of the lower layer fluctuates with time during the annealing process. Therefore, the state of the upper layer also changes with time. At this time,
The time averaging operation is represented by the symbol <> on the right side of Expression (12).

Next, a function П ₀ (T) that gives a target value at each temperature is defined for the satisfaction degree of the constraint conditions defined above. An example of the function П ₀ (T) is shown in FIG. П _{init in} FIG. 10 is the initial value of the target value of constraint satisfaction.
In the present invention, the annealing start temperature T in the SA method is used.
_{At start} (= T ₀ ), the state of the system is observed for a certain period of time, the constraint satisfaction is calculated, and the value is set as П _init . In addition, as a method of setting the values of T _max and T _min , for example, the magnitude of the change in the value of Σ _ll E _cost due to the state change is used as a guide, and the maximum value and the minimum value of the travel time between the delivery destinations are set, respectively. Use the value of the value. Constraint satisfaction П defined above
The degree of difference U between ({θ _i }; T) and its target value П ₀ (T)
To

[0060]

(Equation 8)

It is defined by For each temperature, {a _lj },
_{{C l}, {ο m} , n}, and updates the values of {σ _{m, n}.} That is,

[0062]

[Equation 9]

The parameter θ _i is updated by. However, {θ _i } is {a _lj }, {cl _l }, {ο _{m, n} },
It represents {σ _{m, n} }. The function E _total is a total energy function, and in the above embodiment, the function of Expression (1) may be used. <> Represents a time average operation. Equation (14) is
It is a parameter update rule in a two-layer random field model. In the present invention, the parameter is updated at each temperature according to this update rule. Since the time averaging operation at the time of thermal equilibrium is required on the right side of Expression (14), the operation of creating a thermal equilibrium state and the operation of slightly changing the parameters by Expression (14) are alternately repeated. The parameter updating ends when the value on the right side of Expression (14) converges to almost 0 or the number of times of repetition is cut off at a fixed number of times. The metropolis method may be used to reach the thermal equilibrium state of the system. Further, since the strength of the constraint term {a _lj } is preferably the minimum necessary size for satisfying the constraint condition, the initial value of the constraint term weighting coefficient {a _lj } is the cost term weighting coefficient {b _l } value (fixed to = 1) is set as small as possible.

FIG. 11 shows a block diagram of the minimization unit 24 that minimizes the total energy function shown in FIG. Further, FIG. 12 shows a block diagram of the parameter adjusting unit (also shown in FIG. 11) shown in FIG. After learning the proficiency function (ie, parameters {ο _{m, n} }, {σ _{m, n} }
In order to formulate the optimal delivery plan for the delivery cases that are the target of the planning after the learning, the delivery plan determining unit 3 shown in FIG. 1 does not automatically adjust the weighting factor in the energy function. It can be obtained (detailed in Prior Art 2). However, {ο
The values of _{m, n} } and {σ _{m, n} } are fixed to the learned values. The energy function at this time considers the driver's specialty area by multiplying the cost term and the constraint term used in the prior art 2 and the matching term represented by the equation (5) and adding it. The optimal delivery plan can be planned and determined. As the constraint target value function and constraint condition satisfaction required at this time, the constraint target value function and constraint condition satisfaction used for learning can be used as they are. However, the state of the lower layer is ω _low = x _l , where l represents the targeted delivery case (usually a single delivery case). The constraint target value function and the delivery cases used in the equations (7) to (11) are not for learning, and only the delivery cases targeted for planning are used.

(Example of Experimental Results) An example of experimental results will be shown. FIG. 13 shows the proficiency function before learning, and in FIG. 13, the proficiency function before learning of three drivers is shown.
The circle shown in the figure is the radius {γσ from the center of {ο _{m, n} }.
It is a circle of _{m, n} } (γ = 2.0). FIG. 14 shows a proficiency function after learning. From FIG. 14, the center and spread of the proficiency function of each driver are updated by learning, and most of the delivery destinations (marked with X) in the delivery case for learning are updated. You can see that it covers.

[0066]

According to the invention of claim 1, the state of the lower layer is defined by a set of delivery plans in the learning delivery case, the energy function of the lower layer is defined by the constraint term, the cost term, and the matching term, and the constraint condition of the upper layer. Using the two-layer random field model whose degree of satisfaction is defined by the degree to which the target value of each of these three types is reached, the most efficient driver sharing area is determined for the delivery destination data prepared in advance. As a result, it is possible to determine how to set the sharing area of each driver to be most efficient.

According to the second aspect of the present invention, in the first aspect of the invention, the driver's shared area can be determined so that a plurality of delivery cases (for example, a plurality of past delivery cases) can be delivered most efficiently. , When deciding the sharing area of the switch, prepare a plurality of past delivery cases and use these delivery cases as learning data to allocate the most effective division area as a total of multiple cases. It can be carried out.

According to the invention of claim 3, in claim 1, as a matching term, the degree to which the driver is familiar with an area is "normalized" (a value obtained by integrating in a defined area is a normalized value) such as a normal distribution function. Since it is expressed as a mountain-shaped function, it is possible to express how familiar the driver is with regard to the knowledge of the road at a certain point with an analog function in detail and evaluate the sharing area. It can be done more appropriately.

According to the invention of claim 4, in claim 3, the driver's specialty area is represented by a plurality of proficiency functions. Therefore, the shape of the driver's specialty area is not a simple circle. It can be expressed even in the case of a complicated shape or a plurality of discontinuous areas.

According to the invention of claim 5, in claim 4, a skilled driver who has a long delivery experience and a beginner driver who has little delivery experience and who is not familiar with geography are expressed by the number of proficiency functions belonging to each driver. (That is, the skilled driver has many areas where he or she is familiar, and conversely, the number of beginner drivers is small is expressed by the number of proficiency functions). There are a large number of skilled drivers and, conversely, there are beginner drivers with a small number of familiar areas, and by increasing or decreasing the number of learning functions according to the experience of each driver, the skill level of the driver can be increased. Can be reflected.

According to a sixth aspect of the present invention, in the first aspect, the driver who learns the proficiency level function and the driver who does not learn the learning function and can be fixed can be arbitrarily selected. When it is not necessary or necessary to change the sharing area, by removing this driver's proficiency function from the learning target, the calculation time at the time of learning can be shortened.

According to a seventh aspect of the present invention, in the first aspect, a proficiency function learned by each driver and a proficiency function not learned by each driver can be arbitrarily selected, and a certain driver can be selected from the learning target. Since a part of the area of expertise of the driver is excluded, when expressing the area shared by the driver with multiple proficiency functions, some proficiency functions are removed from the learning target and fixed. This can be realized by using the learning function of As a result, with respect to a part of the area where the driver is familiar with the geography, it can be used as it is for planning without changing it at the time of learning as experience know-how.

According to the invention of claim 8, in claim 1, the degree of matching between the driver's shared area (expressed by a proficiency level function) and the delivery destination is arbitrarily set by the system operator with the learning of the proficiency level function as a target value. Since it can be set to
Depending on the driver's ability, for example, the learning target can be freely set specifically by the system operator, for example, learning is performed so that all the delivery destinations are assigned within a circular area having a radius of 200 km.

According to the ninth aspect of the present invention, in the third aspect, the parameters such as the center and spread of the proficiency function are set so that the proficiency function converges to a meaningless state during learning. Can be prevented. For example, when a plurality of proficiency functions are assigned to one driver, the centers of the proficiency functions should not be close to each other within a certain range to prevent the two proficiency functions from overlapping. This is effective in cases where a driver's specialty area is limited to a certain area.

According to a tenth aspect of the present invention, a learning delivery case or a delivery case to be planned is simulated by using a proficiency level function after learning or further modified by a system operator. By comparing and confirming the effect of learning the learning function with the effect of the correction by the system operator, the area of each driver's share is finally decided. With regard to the delivery case used as learning data or the delivery case used as a planning target, the effect can be confirmed with the operator's correction and finally the driver's proficiency function can be determined. The system operator can add his / her experience know-how and interactively modify the proficiency function after learning.

[Brief description of drawings]

FIG. 1 is a schematic block diagram of a delivery planning system having a learning function determining function (learning function) according to the present invention.

FIG. 2 is a block diagram of a delivery plan determination unit 3 shown in FIG.

FIG. 3 is a diagram for explaining an example of delivering three vehicles to nine destinations.

FIG. 4 is a block diagram of a proficiency level function determination unit 4 in FIG.

FIG. 5 is a diagram showing a constraint target value function related to designation of a time zone of delivery time.

FIG. 6 is a diagram showing a constraint target value function relating to an upper limit of a loaded weight.

FIG. 7 is a diagram showing a constraint target value function relating to the number of vehicles used.

FIG. 8 is a diagram showing a constraint target value function relating to deviation of delivery time for each vehicle.

FIG. 9 is a diagram showing a function expressing a request of a delivery planner with respect to a position ratio from a central position of a proficiency level function for each driver.

FIG. 10 is a diagram showing a function that gives a target value at a temperature of a system.

11 is a block diagram of a minimization unit 24 that minimizes the total energy function of FIG.

12 is a block diagram of a parameter adjusting unit 25 of FIG.

FIG. 13 is a diagram showing a proficiency level function before learning.

FIG. 14 is a diagram showing a proficiency level function after learning.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Control part, 2 ... Input part, 3 ... Delivery plan determination part, 4 ... Proficiency level function determination part, 5 ... Parameter determination part, 6
... Calculation result output unit, 11 ... Delivery plan drafting unit, 12 ...
Energy function term, 13 ... Energy value calculation section, 14 ...
Energy minimizing unit, 15 ... Adjusting unit, 21 ... Total delivery plan creating unit, 22 ... Total energy function determining unit, 23 ... Total energy value calculating unit, 24 ... Total energy minimizing unit, 25 ...
Parameter adjuster.

Claims (10)

[Claims] 1. The state of the lower layer is defined by a set of delivery plans in a learning delivery case, the energy function of the lower layer is defined by a constraint term, a cost term, and a matching term. A delivery plan characterized by determining the most efficient driver sharing area for the delivery destination data prepared in advance, using a two-layer random field model defined by the degree of reaching the target value of each term Support system. 2. The delivery plan support system according to claim 1, wherein a driver's shared area can be determined so that a plurality of delivery cases can be delivered most efficiently. 3. In claim 1, the degree of familiarity with a driver's area as a matching term is expressed by a “normalized” mountain-like function (hereinafter referred to as a proficiency function) such as a normal distribution function. A delivery plan support system characterized in that 4. The delivery plan support system according to claim 3, wherein each driver's specialty area is represented by a plurality of proficiency level functions. 5. The method according to claim 4, wherein a skilled driver who has a long delivery experience and a beginner driver who has little delivery experience and is unfamiliar with geography are represented by the number of proficiency functions belonging to each driver. A delivery planning support system. 6. The delivery plan support system according to claim 1, wherein a driver who learns a proficiency level function and a driver who does not learn the proficiency function and can be fixed can be arbitrarily selected. 7. The driver according to claim 1, wherein
A delivery planning support system characterized in that a learning function and a learning function can be arbitrarily selected. 8. The system operator according to claim 1, wherein the degree of matching between the driver's shared area (expressed by a proficiency level function) and the delivery destination can be arbitrarily set by the system operator with the learning of the proficiency level function as a target value. A delivery planning support system characterized by the above. 9. The center of the proficiency function according to claim 3,
A delivery planning support system characterized by setting limits on parameters such as spread. 10. A proficiency function after learning or by using a proficiency function further modified by a system operator to perform a simulation on a learning delivery case or a delivery case targeted for planning. The delivery plan support system is characterized in that the driver's share area is finally decided while comparing and confirming the learning effect of the above and the effect of the correction by the system operator.