CN101231720A - Enterprise process model multi-target parameter optimizing method based on genetic algorithm - Google Patents

Abstract

The invention relates to the enterprise engineering and the information technology field, and discloses an optimization method for the multi-objective enterprise process model parameters, which is based on the genetic algorithm. In the optimization field of the enterprise process model parameters, most optimization methods use the countermeasure weighing principle to combine each sub-goal into a single objective, so as to process an optimization objective, thus, some defects exist in the comprehensive evaluation process. Aiming mainly at various optimal parameters, such as the production rate of the product, the quantities of various resources, the persistent time of each activity and the configuration schemes for the required resources, as well as the scheduling strategies for the selectable resources, the invention comprehensively evaluates a plurality of indexes, such as the running time, the running cost, the final product quality, the utilization rate of the product, the queue length, etc. The invention adopts the method which divides the problem space in a multi-layer way to process the combinatorial constraint relationship among the various optimal parameters, thereby increasing the flexibility of the optimization parameter selection, and avoiding the analysis calculation to the ineffective parameter combination scheme, furthermore facilitating the maintenance of the group diversity.

Description

Enterprise process model multi-target parameter optimizing method based on genetic algorithm

Technical field

The present invention relates to enterprise engineering and areas of information technology, particularly, relate to a kind of multiple goal enterprise process model parameter optimization method based on genetic algorithm.

Background technology

Enterprise process model is to the simplification of this complex object of enterprise and abstract, not only comprises the activity of anabolic process and the logical relation between the activity, also comprises the product as movable input, output, and the resource object of support activities execution.By simulation analysis and optimization, can find the traffic bottleneck problem that enterprise exists, for the transformation of enterprise with optimize the foundation that operation provides science to enterprise process model.Enterprise process optimization is divided into structure optimization and two work of parameter optimization, wherein parameter optimization is meant the enterprise process model of determining relatively at structure, by the controllable parameter in the adjustment model process is optimized, it can provide the analysis data for further structure optimization.

Parameter optimization is the work of a system, need take into full account in the model all kinds of parameters to the combined influence of process operation, parameter in the process model is of a great variety, wherein have only those to be provided with by the modeling personnel in the modelling phase, can be in the simulation analysis stage by analyst's direct regulation and control, and the parameter that the process performance index is exerted a decisive influence could be as the decision variable of optimizing, i.e. parameters optimization.The process of parameter optimization as shown in Figure 1.

In general, the process model parameters optimization mainly comprises object properties parameter and resource dispatching strategy two big classes, wherein the former comprises configuration of resource sum and movable three kinds of resource requirement configuration (general designation resource allocation proposal), the product generating rate schemes of carrying out, to its optimize the rationality problem that enterprise is the business procedure Resources allocation of mainly solving, reasonable resources is utilized problem under the limited resources condition, and product waiting list control problem; The latter's research is mainly concentrated at present how at process simulation pool allocate resource in service, to improve movable degree of parallelism and resource utilization, promptly be devoted to seek one or more resource dispatching strategies the research of aspect and rarely seen combination for multiple resource dispatching strategy is selected the superior.Modern enterprise is a complicated social technological system, its parameter type is many, quantity is big, often there are various constraints between the dissimilar parameters optimization, change arbitrary parameter, all may influence choosing of other parameters, therefore not only increased the complicacy of handling problems, and its assembled scheme is difficult to represent with unified method.

The target of enterprise process model optimization is for adapting to the needs of market competition, obviously improve the key element of reflection enterprise competitiveness, sum up in the point that the evaluation content of simulation analysis, relate generally to multinomial targets such as working time, operating cost, final products output, resource utilization, queue length.Very big relevance is arranged between these targets, for example Huo Dong execution time is long more under the certain situation of resource quantity, movable cost will be high more, but the cost that the shortening activity execution time not necessarily can the reduction activity, because the shortening activity execution time is a cost to increase the related resource number, may be that the operating cost of final process rises.Therefore when taking all factors into consideration multinomial evaluation index, enterprise process model optimization belongs to multi-objective optimization question, and it optimizes the result is not single separating, but one group of equilibrium solution, promptly so-called Pareto optimum solution.

20th century, the evolution algorithm of the mid-80 artificial intelligence began to be applied to find the solution multi-objective optimization question, emerge a variety of multi-target evolution algorithms in recent years, some of them successful Application are studied and application thereby form a nearest hot topic in engineering practice.A large amount of examples and sign show the most suitable multi-objective optimization question of finding the solution of mechanism of evolution algorithm, because they can find a plurality of Pareto optimum solutions in the single-wheel simulation process, by seeking the individuality with some feature by the generation combination.Even have the scholar to think, be better than other blind search methods at multiple-objection optimization field evolution algorithm.Genetic algorithm is as one of three kinds of typical algorithm in the evolution algorithm of research at present, it is good to have versatility and robustness, characteristics such as search capability is strong, particularly have the multiple criteria optimization of a plurality of conflict objectives, and the reluctant difficult typical optimization problem aspect of this two class of search volume extensive and high complexityization has very big advantage in solution.

Summary of the invention

In enterprise process model parameter optimization field, most of optimization methods have all followed a fixed mode technology and have solved route, after promptly using countermeasure balance principle that the relative importance of each sub-goal is traded off, being combined into a single goal handles, different with these traditional optimization methods, the present invention handles combination restriction relation between all kinds of parameters optimization in the method that adopts multi-level partition problem space aspect the optimized Algorithm, increased the dirigibility that parameters optimization is selected, not only improve optimization efficient, and helped to keep the diversity of colony; Aspect optimization aim, be purpose with indexs such as comprehensive evaluation time, cost, output, resource utilization, queue lengths, the preference information that provides in conjunction with the decision-maker, between each sub-goal, coordinate balance and compromise, make each sub-goal function reach optimum as much as possible.The partition structure of problem space as shown in Figure 2.

The inventive method mainly is divided into four parts:

1. extract the parameters optimization in the enterprise process model at first according to demand.In general, parameters optimization comprises that mainly product-derived produces the allocation plan of sum, each movable duration and resource requirement of speed, all kinds of resources and alternative resource dispatching strategy etc.Classification at parameters optimization is divided problem space, makes the corresponding one deck problem space of every class parameters optimization;

2. determine respectively to optimize sub-goal according to optimization demand and enterprise process model classification, relate to working time, operating cost, final products output, resource utilization, queue length etc., by the weights omega of decision-maker according to each index of Preferences _iAnd disaggregation size M so that the later stage to individual choice the time, carry out auxiliary judgment;

3. divide level execution layering genetic algorithm according to problem space and carry out optimizing;

4. optimizing finishes, and output optimization is optimized disaggregation.

Wherein the 3rd part specifically describes as follows:

3.1 set the required initial parameter of genetic algorithm, comprise maximum iteration time MaxG, population scale N, crossover probability, variation probability etc.;

3.2 determine the problem space of first kind parameters optimization correspondence, and select the individual initial population P1 that constitutes at random;

3.3 the individual Ci (1≤i≤N), and determine in view of the above the problem space of the second class parameters optimization correspondence to generate its initial population P2Ci among the P1 is selected in circulation;

3.4 and the like, until the initial population PnCi_ that generates last class parameter correspondence problem space ... _ y_j (1≤i≤N, 1≤j≤N, 1≤y≤N).When having restriction relation between the two class parameters optimization (as for any one activity, when selecting it to carry out required resource allocation proposal, resource quantity all can not surpass the sum for the respective resources configuration), just can carry out cutting to the problem space of back one class parameter, thereby avoid producing invalid parameter combinations scheme according to the problem space of last class parameter.In addition, the inhomogeneity parameters optimization can also adopt the different coding mode according to demand;

3.5 PnCi_ is selected in circulation ... _ y_j (1≤i≤N, the individual Ci_ among the 1≤j≤N, 1≤y≤N) ... _ y_j_m (1≤m≤N), and according to the selected individual corresponding parameters combination of each layer population, generate corresponding process model example, this example of simulation run obtains each sub-desired value;

3.6 based on the winning relation of Pareto to PnCi_ ... individual Ci_ among the _ y_j ... _ y_j_m sorts, and determines its fitness value fx.Specific practice is for being inferior to Ci_ in the contemporary population of statistics ... the number n of the individuality of _ y_j_m makes fx=n;

3.7 with PnCi_ ... optimum individual among the _ y_j copies to the outside advantage collection RnCi_ of this colony's correspondence ... _ y_j, and with RnCi_ ... inferior solution deletion (first generation population does not comprise inferior solution after duplicating, and need the not carry out deletion action) among _ y_j.If RnCi_ ... quantity individual among the _ y_j surpasses M, then it is carried out clustering processing, and specific practice is:

1) utilize the way of linear weighted function to calculate each individual aggreggate utility value $Q\left(X\right)=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\ω}}_i{f}_i\left(X\right),$ ω wherein _iBe the weight coefficient of each index, satisfy $\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\ω}}_i=1,$ M is the quantity of index, f _i(X) each desired value that obtains by emulation for this individuality;

2) size according to Q (X) sorts to individuality is descending, and gives sequence number successively, and the deletion sequence number is greater than the individuality of M.

If 3.8 do not reach the iterations of optimization setting, then use league matches competition mechanism selective advantage individuality (except the first generation) from the present age and previous generation population, generate new pairing pond, and then intersect, make a variation, generate new population PnCi_ ... _ y_j goes to 3.5 and continues to carry out; Otherwise go to 3.9;

3.9 with RnCi_ ... _ y_j is as individual Ci_ ... the disaggregation of _ y_j;

3.10 calculate last layer population P (n-1) Ci_ ... each individual fitness among the _ y, specific practice is:

1) with (RnCi_ ... _ y_1) ∪ (RnCi_ ... _ y_2) ∪ ... ∪ (RnCi_ ... _ optimum solution in y_N) copies to population P (n-1) Ci_ ... outside advantage collection R (n-1) Ci_ of _ y correspondence ... _ y, and the inferior solution of inciting somebody to action is wherein deleted;

2) statistics R (n-1) Ci_ ... belong to individual Ci_ among the _ y ... the quantity m of the individuality of _ y_j correspondence, and with m as individual Ci_ ... the fitness of _ y_j;

3) if R (n-1) is Ci_ ... quantity individual among the _ y surpasses M, then it is carried out clustering processing, and method is with 3.7;

If 3.11 do not reach population P (n-1) Ci_ ... the iterations of _ y optimization setting, then use the league matches competition mechanism from population P (n-1) Ci_ ... selective advantage individuality (except the first generation) among _ y, generate new pairing pond, and then intersect, make a variation, generate new population P (n-1) Ci_ ... _ y goes to 3.4 and continues to carry out.If two individual fitness of competition are identical in the selection course, then relatively should individuality at R (n-1) Ci_ ... all corresponding individual average aggreggate utility value AvgQ select the then high individuality of AvgQ among the _ y; Otherwise go to 3.12;

3.12 the rest may be inferred, also reaches MaxG until the number of times of population P1 iteration optimization;

3.13 with the optimal solution set of R1C as P1.

Flow process as shown in Figure 3.

Description of drawings

The process of Fig. 1 parameter optimization

The partition structure of Fig. 2 problem space

Fig. 3 enterprise process model multi-target parameter optimizing flow process

Embodiment

Developed prototype system based on the inventive method, this system comprises that the user provides interface, active resource allocation plan analysis and processing module, parameters optimization extraction module, hierarchy optimization module, the process model simulation analysis module of enterprise process model and optimizes display module as a result.

Below concrete enforcement of the present invention is further described:

Step 1: provide the process model that needs optimization process by the user by man-machine interface, extract the parameters optimization in the enterprise process model: product-derived produces allocation plan and alternative resource dispatching strategy of sum, each movable duration and the resource requirement of speed, all kinds of resources;

Step 2: determine respectively to optimize sub-index, comprise that specifically the product of PTU working time, operating cost effectiveness PCU, utilization of resources effectiveness RUU, comprehensive evaluation final products output and queue length is piled up effectiveness PHU:

$1)---\mathrm{PTU}=\frac{\mathrm{ProcessDuration}}{\mathrm{ProcessDuration}+\mathrm{Process}{\mathrm{Duration}}_{\mathrm{expect}}},$ Wherein ProcessDuration and

ProcessDuration _ExpectRepresent the working time that emulation obtains and the working time of expectation respectively;

$2)---\mathrm{PCU}=\frac{\underset{i=1}{\overset{m}{\mathrm{\Σ}}}\mathrm{ProcessCost}(t_i,t_i+\mathrm{\Δt})}{\underset{i=1}{\overset{m}{\mathrm{\Σ}}}\mathrm{ProcessCost}(t_i,t_i+\mathrm{\Δt})+\mathrm{Process}{\mathrm{Cost}}_{\mathrm{expect}}},$ Wherein m represents that the whole process cycle is divided into the m equal portions, ProcessCost (t _i, t _i+ Δ t) is illustrated in the time interval (t _i, t _i+ Δ t) Nei process cost obtains ProcessCost by each product-derived cost and all kinds of used resources costs that is consumed constantly in the dynamic statistics simulation process _ExpectBe expressed as this expectation value;

$3)---\mathrm{RUU}={(\frac{1}{k\·m}\underset{i=1}{\overset{k}{\mathrm{\Σ}}}\underset{j=1}{\overset{m}{\mathrm{\Σ}}}(\mathrm{RN}\left(r_i\right)-\mathrm{NumofResUsed}(r_i,t_j))/\mathrm{RN}\left(r_i\right))}^{-1},$ Wherein k represents resource type quantity,

The m implication as above, RN (r _i) and NumofResUsed (r _i, t _j) represent resource r respectively _iThe sum and at t _jActual usage quantity constantly obtains by the dynamic statistics in the simulation process;

$4)---\mathrm{PHU}={(\frac{1}{m\·n}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\underset{j=1}{\overset{m}{\mathrm{\Σ}}}\mathrm{QueueLength}(p_i,t_j)/L\left(p_i\right))}^{-1},$ Wherein n represents product type quantity, the m implication as above,

QueueLength (p _i, t _j) and L (p _i) respectively at t _jSome non-final products p in the moment process _iQuantity and p _iThe queue-limit of expectation.

Step 3: the weights omega of setting each index _i, disaggregation size M and the required initial parameter of genetic algorithm, general M is set between 10～30, maximum iteration time is set in 200 generation～1000 between generation, population scale is set between 20～30, crossover probability is set between 0.6～1.00, and the variation probability is set between 0.005～0.01;

Step 4: determine that product-derived produces the problem space of speed correspondence, and select the individual initial population P1 that constitutes at random;

Step 5: (1≤i≤N), generate the corresponding problem space of sum of all kinds of resources selects individual formation initial population P2Ci to the individual Ci among the circulation selection P1 at random;

Step 5: the individual Ci_x among the P2Ci is selected in circulation, accept or reject the movable allocation plan of carrying out resource requirement by active resource allocation plan analysis and processing module, if exceeding the quantity of respective resources among the Ci_x, resource requirement quantity gives up this scheme, when avoiding optimizing thus to the analytical calculation of invalid resource distribution assembled scheme, and structure one deck population P3Ci_x down;

By the individual a among the active resource allocation plan optimal module circulation selection A _i

Step 6: the individual Ci_y_j among the P3Ci_y is selected in circulation, generates the problem space of resource dispatching strategy correspondence, selects the individual initial population P4Ci_y_j that constitutes at random,

Step 7: circulation selects the individual Ci_y_j_m among the P4Ci_y_j to make up according to the selected individual corresponding parameters of each layer population, and by the sub-desired value of process model simulation analysis module calculating Ci_y_j_m, computing method are as shown in step 2;

Step 8: based on the winning relation of Pareto the individual Ci_y_j_m among the P4Ci_y_j is sorted, add up the number n of the individuality that is inferior to Ci_y_j_m in the contemporary population, make the fitness value fx=n of Ci_y_j_m;

Step 9: the optimum individual among the P4Ci_y_j is copied to the outside advantage collection R4Ci_y_j of this colony's correspondence, and,, then it is carried out clustering processing if individual quantity surpasses M among the R4Ci_y_j with the deletion of the inferior solution among the R4Ci_y_j;

Step 10: if do not reach the iterations of optimization setting, then use league matches competition mechanism selective advantage individuality (except the first generation) from the present age and previous generation population, generate new pairing pond, and then intersect, make a variation, generate new population P4Ci_y_j, go to step 7 and continue to carry out; Otherwise go to step 11;

Step 11:, press preceding method and calculate each individual fitness among the last layer population P3Ci_y with the disaggregation of R4Ci_y_j as individual Ci_y_j;

Step 12: if do not reach the iterations of population P3Ci_y optimization setting, then continue to select, intersect, make a variation, generate new population P3Ci_y, go to step 6 and continue to carry out; Otherwise go to step 13;

Step 13: according to above step, the number of times that generates until population P1 iteration also reaches MaxG;

Step 14: by optimize as a result display module with optimum individual collection R1C as optimizing result's output.

The inventive method is carried out parameter optimization from the angle of multiple-objection optimization to enterprise process model, produce multiclass parameters optimization such as the allocation plan of sum, each movable duration and resource requirement of speed, all kinds of resources and alternative resource dispatching strategy at product-derived, many indexs such as comprehensive evaluation working time, operating cost, final products output, resource utilization, queue length are for the corporate decision maker carries out effective business course management and analysis provides strong support.Wherein at the various characteristics of parameters optimization, handle combination restriction relation between all kinds of parameters optimization based on the method in the multi-level partition problem of genetic algorithms use space, not only increased the dirigibility that parameters optimization is selected, when having avoided optimizing to the analytical calculation of Invalid parameter assembled scheme, and help to keep the diversity of colony, effectively improved the efficient of business event model optimization.

Claims (5)

1. multiple goal enterprise process model parameter optimization method based on genetic algorithm, it is characterized in that it realizes as follows: (01) extracts the parameters optimization in the enterprise process model according to demand, classification at parameters optimization is divided problem space, makes the corresponding one deck problem space of every class parameters optimization; (02) determines respectively to optimize sub-goal according to optimization demand and enterprise process model classification, relate to working time, operating cost, final products output, resource utilization, queue length etc., by weight and the disaggregation size M of decision-maker according to each index of Preferences; (03) divides level execution layering genetic algorithm according to problem space and carry out optimizing; (11) optimizing finishes, and output optimization is optimized disaggregation.

2. the multiple goal enterprise process model parameter optimization method based on genetic algorithm according to claim 1, it is characterized in that dividing level execution layering genetic algorithm according to problem space in step (03) carries out optimizing, generate the population of lower layer problem space correspondence by the population individuality of last layer problem space correspondence, when having restriction relation between the parameter, avoid generating invalid parameter combinations scheme;

3. the multiple goal enterprise process model parameter optimization method based on genetic algorithm according to claim 1 is characterized in that calculating in step (03) and adopts following distinct methods when ideal adaptation is spent:

1) individuality in the population is sorted based on the winning relation of Pareto at the bottom, and determine its fitness according to being inferior to this individual number in the contemporary population;

2) when the non-bottom calculates, at first get in this layer population the union R of each individual corresponding disaggregation, add up the quantity m that belongs to this individuality corresponding parameters assembled scheme among the R then, and with m as this individual fitness.

4. the multiple goal enterprise process model parameter optimization method based on genetic algorithm according to claim 1, it is characterized in that the quantity of separating concentrated individuality when outside auxiliary advantage in step (03) surpasses initial limit value and is, then utilize the way of linear weighted function, preference information by means of the user is provided with calculates each individual aggreggate utility value $Q\left(X\right)=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\ω}}_i{f}_i\left(X\right),$ And in view of the above the advantage disaggregation is carried out clustering processing:

5. the multiple goal enterprise process model parameter optimization method based on genetic algorithm according to claim 1, it is characterized in that in step (03) when selected two the individual fitness values of competition selection operation equate, then utilize the way of linear weighted function, preference information by means of the user is provided with calculates each individual aggreggate utility value $Q\left(X\right)=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\ω}}_i{f}_i\left(X\right),$ And in view of the above the advantage disaggregation is carried out clustering processing:

Images