CN106447121A - Intelligent optimization scheduling method based on city delivery - Google Patents

Abstract

The invention relates to an intelligent optimization scheduling method based on city delivery. The method comprises the following steps of S1) using an OVRPTW scheduling model, and through different vehicles, carrying out vehicle dispatching and order arrangement on multi-time delivery of a same commodity receiving point; S2) setting an optimization target of vehicle dispatching and order arrangement settlement: based on a minimized delivery vehicle number, realizing shortest total vehicle work hours; S3) using an EHPBIL algorithm to solve the OVRPTW scheduling model; S4) through a 3-dimensional probability amplitude learning model, realizing vehicle information independence inheritance; S5) through an observation rule based on the learning model, realizing decimal encoding; and S6) through high-quality solution information accumulation, realizing algorithm search convergence, and through maximum and minimum learning models, avoiding algorithm precocity. In the invention, the vehicle dispatching and order arrangement can be performed for assisting scheduling personnel; manual scheduling vehicle dispatching and order arrangement efficiency and accuracy are increased; high robustness and stability are possessed; and the method can be used by city logistics delivery enterprises with a medium or small scale.

Description

A kind of intelligent optimization dispatching method based on city distribution

Technical field

The present invention relates to a kind of Logistic Scheduling method, more particularly, to a kind of intelligent optimization dispatching party based on city distribution Method, belongs to the optimizing and scheduling vehicle technical field in logistics management category.

Background technology

Vehicle Routing Problems (Vehicle Routing Problem, VRP) from nineteen fifty-nine by Dantzig etc. proposition since, It is always the study hotspot of the subject such as applied mathematics, operational research, computer, management science.A kind of discrete variable of VRP problem is Optimization problem (i.e. combinatorial optimization problem, combinatorial optimization problem also include Job Shop Scheduling, 0-1 knapsack problem, bin packing, Clustering problem etc.), have been demonstrated with NP-hard attribute.In solving VRP problem, with the increase of problem scale, it is right to combine The quantity growth of elephant is exceedingly fast, even medium scale example, the scale of a combination thereof also can reach the mysterious order of magnitude and (produce Raw multiple shot array).At present, there is no a kind of efficient algorithm in acceptable time frame, accurately solve extensive VRP problem, Therefore the research to VRP problem and its method for solving has high learning value；With the fast development of modern logistics, logistics is made For the important step of lean supply chain, therefore the research to VRP problem has high economic worth simultaneously.

Based on population population incremental learning (Population-based Increased Learning, PBIL) initially by Baluja.S. proposed in 1994, and be successfully applied to solve the multiple combination optimization such as traveling salesman problem and solving job shop scheduling problem Problem, PBIL algorithm belongs in Distributed fusion algorithm (estimation of distribution algorithm, EDA) A simple class (i.e. the EDA algorithm of each variable independence), is to solve for one of effective ways of combinatorial optimization problem such as VRP.

PBIL is integrated with the genetic search and two kinds of strategies of competition learning based on function optimization, and evolutionary process is considered as learning Process, to instruct the product of offspring by the knowledge learning probability (Learning Probability) acquired in competition learning Raw.This probability is the information accumulation of whole evolutionary process, compared with parents' genetic recombination of genetic algorithm, is instructed with it and produces Offspring will be more eugenic, therefore can obtain convergence rate and preferable result of calculation faster, obtain in practical problem Application.

Based on classical PBIL algorithm, the present invention devises one kind effectively decimal-coded mixed population incremental learning (EHPBIL) algorithm, for solving the city distribution problem of the put forward type of the present invention.Mainly propose a kind of of 3-dimensional probability amplitude Practise model, a kind of observation rule based on learning model of design builds for circuit, improves the self study mode of learning model, The inventive method is made to have stronger robustness and stability, the urban logistics distribution enterprise being available for middle-size and small-size scale uses.

Content of the invention

It is contemplated that assisting dispatcher to send a car, row is single, and raising manual dispatching sends a car to arrange single efficiency and accuracy, carries Go out a kind of intelligent optimization dispatching method of city distribution.

For realizing above-mentioned target, the present invention employed technical scheme comprise that the intelligent optimization providing a kind of city distribution is dispatched Method, comprises the steps：S1) adopt OVRPTW scheduling model, by different vehicle to same being repeatedly dispensed into a little of receiving The row row of sending a car is single；S2) set the optimization aim that the row's of sending a car unijunction is calculated：Realize gross vehicle on the basis of minimizing distribution vehicle number Operating time is the shortest；S3) adopt EHPBIL Algorithm for Solving OVRPTW scheduling model；S4) pass through 3-dimensional probability amplitude learning model, real Existing information of vehicles independent inheritance；S5) by the observation rule based on learning model, realize decimal coded；S6) pass through high-quality solution Information accumulation realizes algorithm search convergence, and avoids algorithm precocious by maximum-minimize learning model.

Further：Described step S1) dispensing workload to complete its last task of receiving of receiving a little for knot Beam identification, described OVRPTW scheduling model has maximum carrying capacity, each receive a time window and dispensing need to each car The amount of asking is constrained restriction as follows：

a_i≤s_i≤b_i(8)

Wherein, N represents cluster result set, N ∈ { 1,2, L, n }；S represents order set, S={ S₁,S₂,…,S_n, S_n Represent the order set in the n of region；K represents vehicle set, makes Len (K) represent the element number in set K, formula (1), (2) Represent vehicle restraint,(m=1,2, L ,+∞) represents the vehicle of the m time use in the n of region；V (V={ 1,2, L, v_max, v_maxRepresent one to receive a little) represent point set of receiving, V={ 0 } represents home-delivery center；t_ij(t_ij>0, t_ii=∞, i, j ∈ V) table Show vehicle from the time overhead to the point j that receives for the point i of receiving；Represent vehicleComplete to receive whether to provide and deliver after point i and receive a little J,Represent and receive a little whether there is vehicleDispensing；q_i(i ∈ V) represents point i goods demand of receiving,Represent vehicle's Maximum carrying capacity limits；Formula (6) represents that the vehicle directly service that completes to receive after point i dispensing is received point j；Formula (7) represents Vehicle only serviced the point i that receives before service receiving note j；[a_i,b_i] represent the time window of point i of receiving, s_iRepresent vehicle Start service to receive the point i moment, formula (8) represents the time windows constraints of the point i that receives.

Further：Described step S2) include：

Wherein, f₁For realizing minimizing distribution vehicle number, f₂It is to realize gross vehicle on the basis of minimizing distribution vehicle number Operating time is the shortest,Represent and complete the moment of point j of receiving, σ from the point i that receives_jWhen representing the unloading of the point j that receives or freighting Between.

Further：Described step S3) include：01 canonical matrix is generated by sampled probability width learning matrix, then adopts The regular generation standard decimal coded disaggregation of observation；At the initial stage of algorithm evolution, maintain the search width of algorithm based on roulette wheel model, Thereafter, with the iteration of algorithm, after individual similarity in population reaches pre-set threshold value, maximum-minimize learning model is set Mutation probability width, expands 01 canonical matrix observed result codomain, realizes the effective search to global issue solution space.

Further：Described step S4) in 3-dimensional probability amplitude learning model β (m × v_max×v_max) be：

By above-mentioned learning model β, realize the independence transmission of different vehicle difference train number information of vehicles, strengthen EHPBIL and calculate Method is directed to the solution performance of multi-vehicle-type many train numbers problem, accelerates convergence of algorithm.

Further：Described step S5) pass through learning model β (m × v_max×v_max), obtain standard 01 observing matrix α (m ×v_max×v_max), and definition α meets equation below (12) and the constraint of (13), vehicleTurning from the point i to the point j that receives that receives Move shown in probability such as formula (14),Represent vehicleDo not access point set of receiving, randomly choose a transition probability position Put j', and make it correspondingAgain metric solution form is decoded into by standard 01 observing matrix α；

α_xgα_x'=0 (wherein, x, x' ∈ [1,2, L, m]) (12)

Further：Described step S6) adopt equation below (15)～(17) self refresh learning model：

Wherein ρ represents Forgetting coefficient.

Further：The population scale also including arranging in 3-dimensional probability amplitude learning model β is n_popsize=20, maximum Iterations is n_dim=200, and during setting self refresh study, forgetting rate is ρ=0.2.

The present invention compared with prior art has following advantage：It is effectively decimal-coded mixed that the present invention devises one kind Close population incremental learning (EHPBIL) algorithm；Design a kind of 3-dimensional probability amplitude learning model β, realize different vehicle difference train number car The independence transmission of information, strengthens the solution performance that EHPBIL algorithm is directed to multi-vehicle-type many train numbers problem, accelerates convergence of algorithm； Realize algorithm search convergence through high-quality solution information accumulation, and avoid algorithm precocious by maximum-minimize learning model, realize calculating Method searches for the balance of breadth and depth；It is ensured that the enforceability of the row's of sending a car unijunction fruit automatically in the form of man-machine interaction.

Brief description

Fig. 1 is a kind of algorithm flow chart of the intelligent optimization dispatching method of city distribution of the present invention；

Fig. 2 is the vehicle dispatch system flow chart of the present invention；

Fig. 3 is 2-opt operation chart of the present invention；

Fig. 4 is or-opt operation chart of the present invention.

Specific embodiment

With reference to examples Example and accompanying drawing, the present invention is further elaborated, but embodiments of the present invention do not limit In this.

Refer to Fig. 1 and Fig. 2, a kind of intelligent optimization dispatching method based on city distribution that the present invention provides, bag Include following steps：

(1) sequence information, the data of receive an information, information of vehicles and home-delivery center's information are read in；

(2) according to the information of receiving being related in order, calculated using manhatton distance, and then the trip between obtaining receiving a little Row time matrix, and be loaded in a kind of intelligent optimization dispatching method of city distribution；

(3) the intelligent optimization dispatching method parameter initialization of a kind of city distribution, described parameter includes：Algorithm greatest iteration Number of times (n_dim=200), population scale number (n_popsize=20), Forgetting coefficient ρ (ρ=0.2), set up 3-dimensional probability amplitude Learning model β₀(m×v_max×v_max)；

(4) end condition judges：If n_gene>N_dim, exports final operation planAnd exit algorithm calculating；If No, then go to step (5)；

(5) observe β in the following way₀Obtain α₀, and then it is individual to obtain decimal coded

α_xgα_x'=0 (wherein, x, x' ∈ [1,2, L, m])

(6) it is directed toIn (pop ∈ { 1,2, L, n_popsize }), all individual execution 2-opt and or-opt operations, real Now rightEffective search of solution space nearby, and updateInformation, as shown in Figure 3 and Figure 4；(7) in the following way Renewal learning model, is completed historical search information is learnt, and is operated by maximum-minimize learning model, it is to avoid algorithm search Premature Convergence；

(8) update n_gene=n_gene+1, update global search optimal solutionAnd go to step (5)；

(9) limited to diligent, road and other situation with reference to the actual vehicle vehicle condition on the same day, personnel by dispatcher, suitably Adjustment completes operation plan and exports finally executable operation plan.

Below taking Shanghai reasonable logistics today August in 2016 dispensing task on the 1st as a example：Accumulation dispensing order volume (160 is single), Working specification is as follows：

(1) sequence information (amounting to 160 single), an information of receiving, information of vehicles (fixed vehicle 33) and home-delivery center's letter The data of breath is read in；

(2) according to the information of receiving being related in order, it is calculated dot spacing of receiving by manhatton distance from information (acquired sampling, the average speed per hour obtaining all logistics vehicles is about speed=50 (km/ hour)), refills to change into and receives a little Between time of vehicle operation matrix, and be loaded in intelligent optimization dispatching method；

(3) maximum number of run (n_dim=200), the population scale number (n_popsize=of EHPBIL algorithm are set 20), Forgetting coefficient ρ (ρ=0.2), sets up the following 3-dimensional probability amplitude learning model β showing₀(m×v_max×v_max), n_gene table Show the current evolutionary generation of EHPBIL algorithm；

(4) set algorithm end condition judges for maximum iteration time.Wherein, if n_gene>200, output is final to dispatch PlanAnd exit algorithm calculating；If it is not, then going to step (5)；

(5) observe β in the following way_{n_dim}Obtain α_{n_dim}, and then it is individual to obtain decimal coded

α_xgα_x'=0 (wherein, x, x' ∈ [1,2, L, m])

(6) it is directed toAll individual execution 2-opt and or-opt operations in (pop ∈ { 1,2, L, n_popsize }), It is right to realizeEffective search of solution space nearby, and updateInformation；

(7) renewal learning model in the following way, completes historical search information is learnt, and chemical by minimax Practise model manipulation, it is to avoid algorithm search Premature Convergence；

(8) update n_gene=n_gene+1, update global search optimal solutionAnd go to step (5)；

(9) limited to diligent, road and other situation with reference to the actual vehicle vehicle condition on the same day, personnel by dispatcher, suitably adjust Whole complete operation plan and export finally can perform operation plan.

Machine of the present invention August in 2016 row's of sending a car 27 seconds (systems of single part total time-consuming on the 1st：Win7 64bit, CPU：i5- 5200U, dominant frequency：2.20GHz；Internal memory：12GB；Translation and compiling environment：Python27).

In sum, the present invention is based on city distribution logistics distribution and disperses, a dispersion of receiving, and dispensing is ageing strong, logistics The problems such as dispensing number of packages is many, improve a kind of vehicle routing optimization scheduling model (Open with time windows constraints of opening Vehicle Routing Problem with Time Windows, OVRPTW) it has therefore proved that having NP-hard attribute；Root again According to scheduling market settlement, realize the shortest (the definition operating time of gross vehicle operating time on the basis of minimizing distribution vehicle number： Vehicle is from home-delivery center, until completing last total time-consuming a little received of receiving).So, the present invention designs one kind to be had Decimal-coded mixed population incremental learning (the Efficient Hybrid Population-based Increased of effect Learning, EHPBIL) algorithm, for solving the type city distribution problem.First, the present invention using based on History Order with Receive and be a little accessed for Density Clustering method, mark off the cluster (being that each cluster represents a classification) of some arbitrary shapes, thereafter Realize the classification to new order according to cluster result, thus effectively reducing Solve problems scale；Then, design a kind of being based on to learn The observation rule of model, realizes circuit on the premise of meet the constraint condition and builds；Secondly, using 2-opt and or-opt method Realize circuit improvement, and realize the transmission of parent information by learning model；Finally, complete to send a car by way of man-computer cooperation Row is single.Test checking by a large amount of creation datas, the present invention can assist in dispatcher and sends a car to arrange list, improve manual dispatching and send a car The single efficiency of row and accuracy, have stronger robustness and stability, are available for the urban logistics distribution enterprise of middle-size and small-size scale Use.

Examples detailed above is one embodiment of the present invention, but embodiments of the present invention are not subject to above-mentioned embodiment Limit, other any Spirit Essences without departing from the present invention and the change made under principle, modification, replacement, combine and simplify, Equivalent conversion regime, is included within protection scope of the present invention.

Claims (8)

1. a kind of intelligent optimization dispatching method of city distribution is it is characterised in that comprise the steps：

S1) adopt OVRPTW scheduling model, the row of sending a car is carried out by different vehicle to the same multiple dispensing received a little single；

S2) set the optimization aim that the row's of sending a car unijunction is calculated：When realizing vehicle total working on the basis of minimizing distribution vehicle number Length is the shortest；

S3) adopt EHPBIL Algorithm for Solving OVRPTW scheduling model；

S4) pass through 3-dimensional probability amplitude learning model, realize information of vehicles independent inheritance；

S5) by the observation rule based on learning model, realize decimal coded；

S6) algorithm search convergence is realized by high-quality solution information accumulation, and avoid algorithm early by maximum-minimize learning model Ripe.

2. city distribution according to claim 1 intelligent optimization dispatching method it is characterised in that：Described step S1) , to complete its last task of receiving of receiving a little as end of identification, described OVRPTW scheduling model is to each for dispensing workload Car have maximum carrying capacity, each receive a time window and dispensing demand constrained restriction as follows：

K=K₁∪K₂∪L K_n,

$\∀Len\left(K_n\right)=1,n\∈N---\left(2\right)$

$x_{ij}^{K_n^m}\∈\{0,1\},\∀i,j\∈V,K_n^m\∈K---\left(3\right)$

$y_i^{K_n^m}=\{0,1\},\∀i\∈V,K_n^m\∈K---\left(4\right)$

$\underset{i\∈V}{\Σ}{q}_i\×y_i^{K_n^m}\≤Q_{K_n^m},K_n^m\∈K---\left(5\right)$

$\underset{j\∈V_{K_n^m}\backslash \left\{i\right\}}{\Σ}{x}_{ij}^{K_n^m}=y_i^{K_n^m},i\∈V,K_n^m\∈K---\left(6\right)$

$\underset{i\∈V_{K_n^m}\backslash \left\{j\right\}}{\Σ}{x}_{ij}^{K_n^m}=y_j^{K_n^m},j\∈V,K_n^m\∈K---\left(7\right)$

a_i≤s_i≤b_i(8)

Wherein, N represents cluster result set, N ∈ { 1,2, L, n }；S represents order set, S={ S₁,S₂,…,S_n, S_nRepresent area Order set in the n of domain；K represents vehicle set, makes Len (K) represent the element number in set K, formula (1), (2) represent car Constraint,Represent the vehicle of the m time use in the n of region；V (V={ 1,2, L, v_max, v_maxTable Show that one is received a little) represent point set of receiving, V={ 0 } represents home-delivery center；t_ij(t_ij>0, t_ii=∞, i, j ∈ V) represent car From the time overhead to the point j that receives for the point i of receiving；Represent vehicleComplete to receive the point j that receives that whether provides and delivers after point i, Represent and receive a little whether there is vehicleDispensing；

q_i(i ∈ V) represents point i goods demand of receiving,Represent vehicleMaximum carrying capacity limit；Formula (6) represents The vehicle directly service that completes to receive after point i dispensing is received point j；Formula (7) represents that vehicle only serviced before service receiving note j One point i that receives；[a_i,b_i] represent the time window of point i of receiving, s_iRepresent that vehicle starts service and receives the point i moment, formula (8) table Show the time windows constraints of the point i that receives.

3. city distribution according to claim 1 intelligent optimization dispatching method it is characterised in that：Described step S2) bag Include：

Wherein, f₁For realizing minimizing distribution vehicle number, f₂It is to realize vehicle total working on the basis of minimizing distribution vehicle number Duration is the shortest,Represent and complete the moment of point j of receiving, σ from the point i that receives_jRepresent unloading or the loading time of the point j that receives.

4. city distribution according to claim 1 intelligent optimization dispatching method it is characterised in that：Described step S3) bag Include：01 canonical matrix is generated by sampled probability width learning matrix, then using the regular generation standard decimal coded disaggregation of observation； At the initial stage of algorithm evolution, maintain the search width of algorithm based on roulette wheel model, thereafter, with the iteration of algorithm, when in population After the similarity of body reaches pre-set threshold value, maximum-minimize learning model mutation probability width is set, expands 01 canonical matrix observation Result codomain, realizes the effective search to global issue solution space.

5. city distribution according to claim 1 intelligent optimization dispatching method it is characterised in that：Described step S4) in 3 Dimension probability amplitude learning model β (m × v_max×v_max) be：

${\mathrm{\β}}_{m\×v_{max}\×v_{max}}=\left[\begin{array}{ccc}{\left[\begin{array}{ccc}{\mathrm{\β}}_{11}& L& {\mathrm{\β}}_{1,v_{\max }}\\ M& L& M\\ {\mathrm{\β}}_{v_{\max },1}& L& {\mathrm{\β}}_{v_{\max },v_{\max }}\end{array}\right]}_1& L& {\left[\begin{array}{ccc}{\mathrm{\β}}_{11}& L& {\mathrm{\β}}_{1,v_{\max }}\\ M& L& M\\ {\mathrm{\β}}_{v_{\max },1}& L& {\mathrm{\β}}_{v_{\max },v_{\max }}\end{array}\right]}_m\end{array}\right]---\left(11\right)$

By above-mentioned learning model β, realize the independence transmission of different vehicle difference train number information of vehicles, strengthen EHPBIL algorithm pin Solution performance to multi-vehicle-type many train numbers problem, accelerates convergence of algorithm.

6. city distribution according to claim 5 intelligent optimization dispatching method it is characterised in that：Described step S5) lead to Cross learning model β (m × v_max×v_max), obtain standard 01 observing matrix α (m × v_max×v_max), and definition α meets equation below (12) and (13) constraint, vehicleIt is shown from the transition probability such as formula (14) of the point i to the point j that receives that receives,Represent VehicleDo not access point set of receiving, randomly choose a transition probability position j', and make it correspondingAgain by Standard 01 observing matrix α is decoded into metric solution form；

α_xgα_x'=0 (wherein, x, x' ∈ [1,2, L, m]) (12)

And x ≠ x', y, y' ∈ [1,2, L, v_max]) (13)

$P_{j\backslash \left\{i\right\}}^{K_n^m}=\left\{\begin{array}{c}\frac{{\mathrm{\β}}_{ij}^{K_n^m}}{\underset{j\backslash \left\{i\right\}\∈{\mathrm{tabu}}_{K_n^m}}{\Σ}{\mathrm{\β}}_{ij}^{K_n^m}}\\ esle\\ 0\end{array}\right.---\left(14\right).$

7. city distribution according to claim 6 intelligent optimization dispatching method it is characterised in that：Described step S6) adopt With equation below (15)～(17) self refresh learning model：

${\mathrm{\β}}_{ij}^{K_n^m}=(1-\mathrm{\ρ})\×{\mathrm{\β}}_{ij}^{K_n^m}+{\mathrm{V\β}}_{ij}^{K_n^m}---\left(16\right)$

${\mathrm{\&beta;}}_{ij}^{K_n^m}=\left\{\begin{array}{cc}0.05& {\mathrm{if\&beta;}}_{ij}^{K_n^m}<=0.05\\ 5& {\mathrm{if\&beta;}}_{ij}^{K_n^m}>=5\end{array}\right.---\left(17\right)$

Wherein ρ represents Forgetting coefficient.

8. city distribution according to claim 7 intelligent optimization dispatching method it is characterised in that：Also include arranging 3-dimensional Population scale in probability amplitude learning model β is n_popsize=20, and maximum iteration time is n_dim=200, and setting is from more During new study, forgetting rate is ρ=0.2.