Project planning method under variable project network and construction period random environment

By building the MDPs model in the variable project network and using approximate decomposition and improved GA algorithm, the problem of task duration uncertainty is solved, effective project planning in a random environment is realized, and solution efficiency is significantly improved.

CN119090426APending Publication Date: 2024-12-06XIANGNAN UNIV

Patent Information

Application Number
CN202411120467.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-15
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

In a variable project network, task duration uncertainty and potential priority constraint relationships make it difficult to carry out project planning effectively, and the existing technology fails to effectively solve project scheduling problems in random environments.

Method used

Build the MDPs model and explore its approximate solution algorithm, decompose the random problem approximately into deterministic sub-problems, and use improved GA algorithm and priority rules for solving, compress the solution space and improve the solution efficiency.

Benefits of technology

Through approximate decomposition and improved GA algorithm, the project planning efficiency under large-scale random problems is significantly improved, providing a better solution, which is better than the traditional priority rule method.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119090426A_ABST
    Figure CN119090426A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of production task construction period calculation, in particular to a project planning method under a variable project network and a construction period random environment, which comprises the following steps of: 1, constructing an MDPs model of the variable project network under the construction period random environment; 2, solving a shortage solution of the MDPs model by using random dynamic programming; 3, proposing an approximate decomposition framework for processing a large-scale random problem; 4, establishing a resource flow model for the deterministic sub-problem; 5, solving the deterministic sub-problem by using a meta-heuristic improved GA (Genetic Algorithm); 6, solving the deterministic sub-problem by using a priority rule scheduling algorithm; 7, comparing the effects of the approximate decomposition algorithm based on the improved GA algorithm; by establishing an MDPs model and exploring an approximate solution algorithm of the MDPs model, a random problem is approximately decomposed into deterministic sub-problems in various states, and a solution space is compressed.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of production task duration calculation, and in particular to a project planning method under a variable project network and random duration environment. Background Art

[0002] During the execution of a project, tasks need to be selected because the selection may lead to other tasks that need to be selected or executed. In addition to the priority constraints between tasks, there are also potential priority constraints related to optional tasks. This feature of changes in the project execution network and uncertain task duration due to selection is common in project management. For example, in a customized equipment project, when designing a new equipment unit, you can choose to produce it yourself or outsource it. After the selection, new tasks will be generated. If it is produced by yourself, you need to purchase raw materials, process and assemble products, etc. If it is outsourced, you need to go through the procurement process. New equipment needs to be verified and tested many times. This process has a potential correlation with the selected tasks. If it is processed by yourself, you can test while producing. If it is outsourced, you must wait until the outsourced product comes back before testing. In addition, because it is a new product, the relevant tasks are not proficient, and the duration of the task is also very uncertain. The above example is defined as a variable project network scheduling problem.

[0003] Variable project network is a variant or extension of the traditional standard resource-constrained project scheduling problem (RCPSP). Compared with the standard RCPSP, the biggest difference is that tasks no longer have to be executed in full. In recent years, this problem has also attracted more and more attention from scholars. There are multiple groups of activity chains with a sequential relationship in the project that can replace each other and are defined as the resource-constrained alternative activity chain project scheduling problem (RCPSP-AC). During the project execution, when one chain starts to execute, the tasks in other chains will not be executed, and the simulated return algorithm is used to solve it. At present, there is no research on the RCPSP-PS problem in a random environment. Summary of the invention

[0004] The purpose of the present invention is to propose a project planning method under a variable project network and a random environment of construction period. By establishing an MDPs model and exploring its approximate solution algorithm, the random problem is approximately decomposed into deterministic sub-problems under each state, and the solution space is compressed; for the deterministic sub-problems, a resource flow model under a variable network environment is constructed, and the performance of the improved GA algorithm and the priority rule solution algorithm are compared and analyzed, which proves that the improved GA algorithm is better, so compared with the priority rule, the improved GA algorithm has better performance.

[0005] To achieve this object, the present invention adopts the following technical solutions:

[0006] A project planning method under a variable project network and random duration environment includes the following steps:

[0007] Step 1: Construct the MDPs model of variable project network under random duration environment;

[0008] Step 2: Use stochastic dynamic programming to find the tight solution of the MDPs model;

[0009] Step 3: The MDPs model is used to deal with small-scale problems, so an approximate decomposition framework is proposed to deal with large-scale random problems and decompose random problems into solutions to deterministic sub-problems;

[0010] Step 4: Establish a resource flow model for the deterministic sub-problem;

[0011] Step 5: Solve the deterministic sub-problem using the GA algorithm improved by meta-heuristics;

[0012] Step 6: Solve the deterministic sub-problem using the priority rule scheduling algorithm;

[0013] Step 7: Compare the effects of the approximate decomposition algorithm based on the improved GA algorithm.

[0014] Preferably, an MDPs model of a variable project network under a random duration environment is constructed, wherein the MDPs model includes a decision time set T, a state set X, an action set A, a transition probability function p(x′|x,a), and a cost function c(x′x,a);

[0015] The decision time set T includes multiple decision times t, t∈T;

[0016] The state set X is the project state and needs to contain all the project information needed for decision making; x = [t x ,W x ,E x ,G x ] T Indicates that x∈X; X is the set of all possible states of the project; where t x represents the decision moment in the current state; W x Indicates the set of tasks waiting to be started at the current moment; E x represents the set of choices that are waiting at the current moment, e∈E x ,e=1,2,...,E, when the task a is triggered e After completion, e is added to the selection set E x Middle; G x represents the set of tasks in progress at the current moment, then G x ={(j1,s1),(j2,s2),...,(j n ,s n)), where jn represents the nth in-progress task of the project at the current moment, s n It represents the remaining completion time of the nth task in progress at the current decision time. At the initial time of the project, all tasks have not started, and the initial state of the system is When all tasks are completed, the system will stay in a final state permanently. Where f is the state number;

[0017] The action set A is the set of tasks waiting to be started based on the obtained task set W in any given state x∈X. x , waiting for the selection set E x , the set of tasks in progress G x , we get that if there are several pending actions a, each action a represents a task start combination that satisfies the task precedence constraint and resource constraint, multiple optional actions constitute an optional action set a∈A x ;

[0018] If there is no task that meets the start condition in state x∈X, then but It will become a special action;

[0019] The transition probability function p(x′|x,a) represents the state x taking a∈A x After that, the probability that the system transfers to state x′(x′≠x,x′∈X);

[0020] The cost function c(x′|x,a) represents taking action a∈A in state x. x After that, the system transfers to the state x′(x′≠x,x,x′∈x) with immediate cost.

[0021] Preferably, the state evolution process of the MDPs model is: for a given state x, obtain the next possible state set Y, select the state x′ in the set Y as the current state, and then calculate the next new state set Y′, and traverse all the states in the set Y in turn until no new state is generated.

[0022] Preferably, a task j∈W x When , j needs to meet the following conditions:

[0023] All task nodes in the must-execute task set V included in the immediate predecessor task set of task j have been completed;

[0024] If task j is an associated execution task of task i, that is, j∈β i , task i must have been completed;

[0025] If task j belongs to a certain choice, that is, j∈C e , triggering task ae Completed.

[0026] Preferably, the stochastic dynamic programming method for finding the tight solution of the MDPs model is:

[0027] Let V(x) denote the value function of state x∈X, For the possible state set after taking n actions, the value iteration algorithm mainly solves the Bellman equation for each state x iteratively, and the equation is expressed as:

[0028]

[0029] During the iteration process, the optimal function value will continue to transfer from the next state to the previous state. When the value function values ​​of all states no longer change, it means that the value iteration algorithm converges and the optimal function value is V*(x0), where x0 is the initial state.

[0030] The corresponding optimal action formula for each state is:

[0031] Where m represents the number of iterations.

[0032] Preferably, decomposing the random problem into solutions to deterministic sub-problems specifically includes:

[0033] Obtain the currently unfinished task network, including the unscheduled task network and the in-progress task network, and use it as the network of the deterministic sub-problem;

[0034] Get the remaining resources in the current state as the resource input for the deterministic sub-problem;

[0035] For a task node with a random duration in a deterministic network, its remaining expected duration is taken as the duration of the task node.

[0036] Preferably, establishing a resource flow model for the deterministic sub-problem specifically includes:

[0037] Define the set of schedulable tasks under a certain state as K x Equal to the remaining unfinished nodes and in-progress nodes in state x, introducing a virtual start task and virtual termination tasks The completion time of the virtual task is equal to the current time of state x, that is, f s =t x ; The resource transfer quantity of the virtual start task is the remaining available resources of the current state x, m kx Represents, where k is the k-th resource; define task j∈K x ∪{s,e}The maximum resource transfer amount for the k-th resource is The maximum amount of resource transfer The formula is:

[0038]

[0039] The objective function that minimizes the total completion time of the project is expressed as:

[0040] minf e , where f e Indicates the last completed task;

[0041] The expression formula for the tasks in the set of tasks that must be executed is:

[0042]

[0043] The trigger task of the optional set is executed. Only one task can be selected in the set for execution. If the trigger task is not executed, all optional tasks will not be executed. The expression formula is:

[0044]

[0045] The relationship between the associated main task and all associated tasks is expressed as follows:

[0046] v i -v j =0,j∈β i ;

[0047] If a task is not executed, the expression formula for canceling the arc associated with the task is:

[0048] y ij -v j =0, i≠j,

[0049] In the associated tasks, the expression formula for the subtask to start only after the main task is completed is:

[0050] j∈β i ;

[0051] The expression formula that the resource arc between two tasks cannot be circulated is:

[0052] y ij +y ji ≤1, i<j,

[0053] The expression formula that the resource arcs between the three tasks cannot be circulated is:

[0054] y ij +y jk -y ik ≤1, i≠j≠k,

[0055] The relationship between the amount of resource transfer and the maximum number of tasks required is expressed as:

[0056]

[0057] All resources of the task node will be transferred out and equal to the resource demand expression formula:

[0058] i∈K x ∪{s},

[0059] The expression formula for all resource inputs of a node equal to resource demand is:

[0060] j∈K x ∪{e},

[0061] If resources flow from i to j, the formula for completing task j later than task i is:

[0062] j∈S j , i∈K x ∪{s},i≠j;

[0063] The expression formula that the amount of resource transfer between tasks must be greater than zero is:

[0064] j∈Sj , i∈K x ∪{s}, i≠j:

[0065] The expression formula for whether there is resource flow between two tasks is:

[0066] j∈K x ∪{e},i≠j,

[0067] y ij ∈{0,1}i∈K x ∪{s}, j∪Kx∪{e}, i≠j.

[0068] Preferably, solving the deterministic sub-problem using a meta-heuristic improved GA algorithm includes encoding mode, crossover mutation operation and fitness calculation;

[0069] Encoding methods include:

[0070] Individuals are represented by:

[0071]

[0072] where a E ; C E is the fixed item auxiliary information, indicating the Eth choice, a E is the selected trigger task, C E is a set of optional tasks, and the corresponding α E From C E When the task node selected in is not triggered and not executed, α E =-1; task list J = (j1, j2, ..., j n ) is a task sequence, the earlier the task is, the higher its priority is; Can be 0 or 1. Indicates that the task is not executed, otherwise it is executed;

[0073] New sorting rules are introduced to describe the situation of task trigger selection. The specific sorting rules are:

[0074] Trigger task a e Ranking among all optional task sets C e In front of the task in C e The tasks in a e The first task j∈C e For the selected task, C e The remaining tasks in the program are never executed;

[0075] If the task a is triggered e There is a choice C m In the optional task set, task a m Must be placed in task a e in front of;

[0076] The set of associated tasks β i The tasks in must be placed after the main task i;

[0077] If the main task i of the associated task has a C e In the set, this main task i must also be ranked before task a e after;

[0078] The specific process of crossover mutation operation is as follows:

[0079] Multiple crossover points are randomly selected, and the two parents exchange the fragments between these crossover points to form new individuals;

[0080] For the offspring individual mutation, there are two operations: insertion and exchange, and the probability of insertion mutation is m α ,0≤m α ≤1, the probability of exchange mutation is m β ,0≤m β≤1. First, perform the insertion operation, and take a random number m from U(0,1) for each position in the task queue. When m<m α , randomly select a task after this position and insert it into this position, and then perform the exchange operation. For each position in the task queue, take a random number m in U(0,1), and when m<m β When , a task is randomly selected after the position and exchanged with the task at the position;

[0081] The fitness calculation is as follows:

[0082] The fitness expression formula of the improved GA algorithm GAII is:

[0083] Where T is the completion time of the project, LF j is the latest completion time of task j, and γ is a large number.

[0084] Preferably, the deterministic sub-problem is solved using a priority rule scheduling algorithm as follows:

[0085] Define n to represent the stage, t n is the scheduling time, and the unstarted set W n , that is, the set of tasks that meet the immediate precedence constraint and the selection-related execution order constraint but have not yet been scheduled, and the set of tasks that can be started Q n , that is, the set of tasks that satisfy the immediate predecessor constraint, resource constraint, and selection-related execution order constraint but have not yet been scheduled, and the completion set F n , in the set P n , discard the set as N n , the parallel scheduling steps are as follows:

[0086] Step A: Determine the decision time of the current stage n. If P n There are tasks in t n If completed, the completed task is transferred to F n , and at the same time, according to the immediate preceding constraint, select the associated constraint update set W n ;

[0087] Step B: Let Q n is an empty set, in W n Find the tasks that meet the resource constraints and add them to Q n middle;

[0088] Step C: If Q n If Q is an empty set, jump to step A and make n+1 stage decisions; n If it is not empty, use the given priority rule to n Sort the tasks in Q nSelect the task with the highest priority and check whether the selected task exists in the selection set C. e Or the association set β i If it exists, the selection association constraint check is performed. When the task a is triggered e , the main task of the associated set is to make i in the discarded set N n Medium, or C e There are already tasks in F n or P n In the process, the selected task needs to be abandoned, and the selected task will be removed from Q n Transfer to N n Otherwise, the selected task will be n Start working at any time and take it from Q n Transfer to P n , jump to step B;

[0089] Step D: Repeat the above steps until all tasks have been scheduled.

[0090] The technical solution provided by the present invention may include the following beneficial effects:

[0091] The present invention mainly studies and establishes an MDPs model and explores its approximate solution algorithm by taking a variable project network and a random environment with uncertain duration of some tasks as the background, including constructing an MDps model of a variable project network under a random duration environment; suggesting that random problems can be approximately decomposed into deterministic sub-problems in each state, thereby compressing the solution space; for deterministic sub-problems, a resource flow model under a variable network environment is constructed, and the performance of an improved GA algorithm and a priority rule solution algorithm are compared and analyzed, proving that the improved GA algorithm is better. Since there is no optimal solution in the case of large-scale random problems, the improved GA algorithm has better performance than the priority rule. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] Figure 1 This is a schematic diagram of an example of a variable project network and a random duration project according to an embodiment of the present invention;

[0093] Figure 2 It is a schematic diagram of the project status evolved from an example project of an embodiment of the present invention;

[0094] Figure 3 is a schematic diagram of a small-scale calculation example of an embodiment of the present invention;

[0095] Figure 4 is a schematic diagram of average RDI under all calculation examples of an embodiment of the present invention;

[0096] Figure 5 It is a schematic diagram of RDI for placing orders with different random intensities according to an embodiment of the present invention;

[0097] Figure 6 It is a schematic diagram of RD under different variable intensities according to an embodiment of the present invention. DETAILED DESCRIPTION

[0098] The embodiments of the present invention are described in detail below, and examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0099] Combine the following Figures 1 to 6 , describing a project planning method in a variable project network and random duration environment according to an embodiment of the present invention.

[0100] A project planning method under a variable project network and random duration environment includes the following steps:

[0101] Step 1: Construct the MDPs model of variable project network under random duration environment;

[0102] Step 2: Use stochastic dynamic programming to find the tight solution of the MDPs model;

[0103] Step 3: The MDPs model is used to deal with small-scale problems, so an approximate decomposition framework is proposed to deal with large-scale random problems and decompose random problems into solutions to deterministic sub-problems;

[0104] Step 4: Establish a resource flow model for the deterministic sub-problem;

[0105] Step 5: Solve the deterministic sub-problem using the GA algorithm improved by meta-heuristics;

[0106] Step 6: Solve the deterministic sub-problem using the priority rule scheduling algorithm;

[0107] Step 7: Compare the effects of the approximate decomposition algorithm based on the improved GA algorithm.

[0108] The present invention mainly studies and establishes an MDPs model and explores its approximate solution algorithm by taking a variable project network and a random environment with uncertain duration of some tasks as the background, including constructing an MDps model of a variable project network under a random duration environment; suggesting that random problems can be approximately decomposed into deterministic sub-problems in each state, thereby compressing the solution space; for deterministic sub-problems, a resource flow model under a variable network environment is constructed, and the performance of an improved GA algorithm and a priority rule solution algorithm are compared and analyzed, proving that the improved GA algorithm is better. Since there is no optimal solution in the case of large-scale random problems, the improved GA algorithm has better performance than the priority rule.

[0109] In view of the situation that the project network structure is variable and the duration of some tasks is discrete and random, the present invention proposes to construct an MDPs model and use random dynamic programming value iteration to accurately solve it. Because the state space of MDPs will increase exponentially with the increase of the problem scale, it will lead to the curse of dimensionality. The present invention proposes to approximately decompose the original random problem into deterministic scheduling sub-problems in each state, so as to obtain the suboptimal scheduling strategy of the original problem. In order to solve the deterministic sub-problems in each state, the present invention constructs a resource flow model and solves it through GA and priority rules.

[0110] Constructing an MDPs model of a variable project network under a random duration environment, wherein the MDPs model includes a decision time set T, a state set X, an action set A, a transition probability function p(x′|x,a), and a cost function c(x′x,a);

[0111] The decision time set T includes multiple decision times t, t∈T; the project manager makes a decision to start a new task when a task is completed, that is, at the decision time, which is also the total time consumed by the project.

[0112] The state set X is the project state and needs to contain all the project information needed for decision making; x = [t x ,W x ,E x ,G x ] T Indicates that x∈X; X is the set of all possible states of the project; where t x represents the decision moment in the current state; W x Indicates the set of tasks waiting to be started at the current moment; E x represents the set of choices that are waiting at the current moment, e∈E x ,e=1,2,...,E, when the task a is triggered e After completion, e is added to the selection set Ex; Gx represents the set of tasks in progress at the current moment, then G x ={(j1,s1),(j2,s2),...,(j n ,s n )), where jn represents the nth in-progress task of the project at the current moment, s n It represents the remaining completion time of the nth task in progress at the current decision time. At the initial time of the project, all tasks have not started, and the initial state of the system is When all tasks are completed, the system will stay in a final state permanently. Where f is the state number; when different choices and different scheduling strategies are currently made, the entire process from project start to project completion contains different states and the final time t f It may also be different.

[0113] The action set A is the set of tasks waiting to be started based on the obtained task set W in any given state x∈X. x , waiting for the selection set E x , the set of tasks in progress G x , we get that if there are several pending actions a, each action a represents a task start combination that satisfies the task precedence constraint and resource constraint, multiple optional actions constitute an optional action set a∈A x ;

[0114] If there is no task that meets the start condition in state x∈X, then but It will become a special action; in the context of the problem studied in this paper, the decision-making process made by the project manager actually includes two steps, first task selection and then start-up decision; the details are as follows:

[0115] Task selection: The project manager first looks at the task selection set E x Is there a choice waiting for decision? For the choice e∈E x Execute the selection operation, a node j∈C e After being selected, the task node will enter the task set W waiting to be started x At the same time, other task nodes i∈C e It will be abandoned and never executed, because there are cases of selection and associated execution, and there will also be cases of associated abandonment. If the abandoned task is a trigger task of a certain choice e, that is, there is a e =i, for C e If there is no node in the immediate preceding task that needs to be executed except the triggering task, the task also needs to be abandoned. If the main task that is associated with execution is abandoned, the associated task will definitely be abandoned.

[0116] Start-up decision: From the set of in-progress tasks G x In the task set W, the information of currently available resources can be obtained. Considering resource constraints and the relationship between the top and the bottom, the project manager selects the task set W from the task set W waiting to be started. x Select tasks and combine them to form a start action a.

[0117] The transition probability function p(x′|x,a) represents the state x taking a∈A xAfter that, the system transfers to the state x′ (x′≠x, x′∈X) with probability. The system state transfer is driven by task completion or start. When a is selected for start, the system will immediately transfer to a new intermediate state with probability 1. Only when a task is completed will the system transfer to the next state x′. We call this intermediate state the transition state. Because the transition state is only an intermediate state given to reduce the state space and facilitate calculation, the transition state does not belong to the state set X. The set of tasks in progress in the transition state is G x and the start action a, the next state x′ of the system depends on the combination of completed tasks γ of the transition state WIP at the next moment, and the corresponding next state transition probability is the combination probability of γ. Let all possible next state sets be Y, with x′òY and p(x′|x,a)=p(γ). Since the present invention considers that the task duration follows a discrete probability distribution, the conditional probability can be used to obtain task j∈G' x The probability of completing it alone at the next moment is p(γ j ). Due to the independence between tasks, we know that the completion combination of multiple tasks at the next moment is γ = [γ1,γ2,…,γ n The probability of And satisfy

[0118] The cost function c(x′|x,a) represents taking action a∈A in state x. x After that, the system transfers to the state x′ (x′≠x, x, x′∈x) with immediate cost. In this study, c(x′|x·a) is defined as the project completion period. When the project is not completed, all immediate costs are 0; when the project is completed in the x′ state, the immediate cost is the decision moment in the current state, that is, the project completion time, c(x′|x·a)=t x′ =t f .

[0119] The state evolution process of the MDPs model is: for a given state x, get the next possible state set Y, select state x' in set Y as the current state, and then calculate the next new state set Y', and traverse all states in set Y in turn until no new state is generated. The traversal process for each state x, that is, the state evolution process, will eventually get all possible states.

[0120] See also Figure 1As shown, the present invention uses 6 generalized nodes to describe the state evolution process of the MDPs model. The example project contains a selection, and the selected trigger task is task number 1, that is, a1=1, and the optional task range C1=[2,3]. Task 2 is the main task of the associated task, and the associated task set β2=[5]. Task 2 has a random duration, which means that the probability of completing the task in duration 1 is 0.4, and the probability of completing the task in duration 3 is 0.6. Task 2 also has a random duration, and the probability of duration 1 and 2 is 0.5 respectively. Assume that the maximum resource amount is 3. Modeling with MDPs and performing state evolution can be obtained as follows Figure 2 Evolutionary state shown.

[0121] exist Figure 2 In the state x1, there are 3 optional actions. When the start action a=(2,3) is selected, there is a 0.4 probability of switching to state x2 and a 0.6 probability of switching to state x3. When the start action a=(3) is selected, there is a 0.5 probability of switching to state x4 or x5. From the project evolution state diagram, it can be seen that the optimal state evolution queue may be [x0,x1,x4,x5]. 13 ,x 14 ] or [x0,x1,x6,x 13 ,x 14 ].

[0122] A task j∈W x When , j needs to meet the following conditions:

[0123] All task nodes in the must-execute task set V included in the immediate predecessor task set of task j have been completed;

[0124] If task j is an associated execution task of task i, that is, j∈β i , task i must have been completed;

[0125] If task j belongs to a certain choice, that is, j∈C e , triggering task a e Completed.

[0126] Dynamic programming median iteration and policy iteration are two basic algorithms for finding the optimal solution of the MDPs model. Policy iteration requires solving a system of linear equations in a single iteration, which has more computational cost than value iteration. Therefore, the present invention adopts the dynamic programming median iteration algorithm to find the most optimal strategy. The specific solution of the MDPs model in random dynamic programming is:

[0127] Let V(x) denote the value function of state x∈X, For the possible state set after taking action a, the value iteration algorithm mainly solves the Bellman equation for each state x iteratively, and the equation is expressed as:

[0128]

[0129] During the iteration process, the optimal function value will continue to transfer from the next state to the previous state. When the value function values ​​of all states no longer change, it means that the value iteration algorithm converges and the optimal function value is V*(x0), where x0 is the initial state.

[0130] The corresponding optimal action formula for each state is:

[0131] Where m represents the number of iterations.

[0132] The evolution of the MDPs model will obtain all possible states, while the value iteration solution needs to traverse all states and actions to find the optimal solution. As the number of tasks increases, the state space will grow exponentially, which will cause the solution to face the curse of dimensionality and lead to a long solution time. In order to minimize the size of the state space, consider only obtaining the set of actions that can be started in the current state A. x The optimal start action in the project is to enumerate the project status only at the optimal start action, so that the state space is much smaller, which is conducive to rapid solution. The final solution may not be the optimal solution, but a suboptimal solution. However, compared with the cost of solving time, a better suboptimal solution is more desired by project managers.

[0133] The solution to decompose the random problem into deterministic sub-problems specifically includes: In order to obtain the optimal action in the current state, the project situation in the current state is transformed into a deterministic sub-problem. The main tasks to be completed in transforming the scheduling problem in the current state into a deterministic sub-problem are:

[0134] Obtain the currently unfinished task network, including the unscheduled task network and the in-progress task network, and use it as the network of the deterministic sub-problem;

[0135] Get the remaining resources in the current state as the resource input for the deterministic sub-problem;

[0136] For a task node with a random duration in a deterministic network, its remaining expected duration is taken as the duration of the task node.

[0137] The resource flow model for the deterministic sub-problem includes:

[0138] For the deterministic subproblem, since the expected duration may be a decimal, it is difficult to establish a traditional integer-based TI model. Therefore, the present invention adopts a resource flow model, i.e., a resource flow model, to model the problem. The model uses the number of direct resource transfers between tasks as a decision variable, so it is suitable for continuous-time project scheduling problems. The difference is that the model established by the present invention is oriented to a variable project network, and the relationship between task selection and task-associated execution is considered. The set of schedulable tasks under a certain state is defined as K x Equal to the remaining unfinished nodes and in-progress nodes in state x, introducing a virtual start task and virtual termination tasks The completion time of the virtual task is equal to the current time of state x, that is, f s =t x ; The resource transfer quantity of the virtual start task is the remaining available resources of the current state x, m kx Represents, where k is the k-th resource; define task j∈K x ∪{s,e}The maximum resource transfer amount for the k-th resource is The maximum amount of resource transfer The formula is:

[0139]

[0140] The objective function that minimizes the total completion time of the project is expressed as:

[0141] minf e , where f e Indicates the last completed task;

[0142] The expression formula for the tasks in the set of tasks that must be executed is:

[0143]

[0144] The trigger task of the optional set is executed. Only one task can be selected in the set for execution. If the trigger task is not executed, all optional tasks will not be executed. The expression formula is:

[0145]

[0146] The relationship between the associated main task and all associated tasks is expressed as follows:

[0147] v i -v j =0,j∈β i ;

[0148] If the main task row is associated, all the associated tasks will be executed. If the main task of the associated task is not executed, all the associated tasks will not be executed either.

[0149] If a task is not executed, the expression formula for canceling the arc associated with the task is:

[0150] y ij -v j =0, i≠j,

[0151] In the associated tasks, the expression formula for the subtask to start only after the main task is completed is:

[0152] j∈β i ;

[0153] The expression formula that the resource arc between two tasks cannot be circulated is:

[0154] y ij +y ji ≤1, i<j,

[0155] The expression formula that the resource arcs between the three tasks cannot be circulated is:

[0156] y ij +y jk -y ik ≤1, i≠j≠k,

[0157] The relationship between the amount of resource transfer and the maximum number of tasks required is expressed as:

[0158]

[0159] All resources of the task node will be transferred out and equal to the resource demand expression formula:

[0160]

[0161] The expression formula for all resource inputs of a node equal to resource demand is:

[0162]

[0163] If resources flow from i to j, the formula for completing task j later than task i is:

[0164] j∈S j , i∈K x ∪{s},i≠j;

[0165] The expression formula that the amount of resource transfer between tasks must be greater than zero is:

[0166] j∈S j, i∈K x ∪(S), i≠j;

[0167] The expression formula for whether there is resource flow between two tasks is:

[0168] i∈K x ∪{s}, j∈K x ∪{e},i≠j,

[0169] y ij ∈{0, 1}, i∈K x ∪{s}, j∈K x ∪{e}, i≠j.

[0170] Other relevant parameters and sets of the resource flow model are shown in the following table:

[0171]

[0172]

[0173] Table 1 RF model symbols and their meanings

[0174] Specifically, for the deterministic RCPSP problem, the present invention will decompose the GA algorithm and the improved GA algorithm GAII to solve the resource flow model, and describe the scheduling method based on priority rules, and use it as the evaluation benchmark for the resource flow model and the corresponding metaheuristic algorithm. The improved GA algorithm GAII proposed in the present invention still adopts the classic genetic algorithm process, so it will not be repeated.

[0175] The meta-heuristic improved GA algorithm is used to solve the deterministic sub-problems, including encoding, crossover and mutation operations, and fitness calculation;

[0176] Encoding methods include:

[0177] For variable project network structures, a feasible solution should contain two parts. The first part contains the selected structure of the realized activity network, and the second part contains the task execution queue. The individual can be represented as:

[0178]

[0179] where a E ; C E is the fixed item auxiliary information, indicating the Eth choice, a E is the selected trigger task, C E is a set of optional tasks, and the corresponding α E From C E When the task node selected in is not triggered and not executed, αE =-1; task list J = (j1, j2, ..., j n ) is a task sequence, the earlier the task is, the higher its priority is; Can be 0 or 1. Indicates that the task is not executed, otherwise it is executed;

[0180] The improved GAII algorithm simplifies the individual representation and only uses a task sequence to describe the individual I = [j1, j2, ...j n ]. Similarly, the closer the task is to the front, the higher its priority is and the earlier it is executed.

[0181] New sorting rules are introduced to describe the situation of task trigger selection. The specific sorting rules are:

[0182] Trigger task a e Ranking among all optional task sets C e In front of the task in C e The tasks in a e The first task j∈C e For the selected task, C e The remaining tasks in the program are never executed;

[0183] If the task a is triggered e There is a choice C m In the optional task set, task a m Must be placed in task a e in front of;

[0184] The set of associated tasks β i The tasks in must be placed after the main task i;

[0185] If the main task i of the associated task has a C e In the set, this main task i must also be ranked before task a e after.

[0186] Based on the above rules, when we obtain the individual encodings, we can analyze the execution of variable network task selection.

[0187] The specific process of crossover mutation operation is as follows:

[0188] The improved GA algorithm GAII uses multi-point crossover as the parent crossover to improve the diversity of the population; multiple crossover points are randomly selected, and the two parents exchange the fragments between these crossover points to form new individuals.

[0189] For the offspring individual mutation, there are two operations: insertion and exchange, and the probability of insertion mutation is m α,0≤m α ≤1, the probability of exchange mutation is m β ,0≤m β ≤1. First, perform the insertion operation, and take a random number m from U(0,1) for each position in the task queue. When m<m α , randomly select a task after this position and insert it into this position, and then perform the exchange operation. For each position in the task queue, take a random number m in U(0,1), and when m<m β When , a task is randomly selected after the position and exchanged with the task at the position;

[0190] Since the crossover mutation actions are all performed by randomly selecting positions, the new individuals may not meet the constraint rules and need to be repaired. When repairing, check whether the task at each current position meets the constraints. If not, find the task at the next position and check the constraints until a task that meets the constraints is found and put it into the current position, and so on, until the entire sequence meets the specifications.

[0191] The fitness calculation is as follows:

[0192] The fitness expression formula of the improved GA algorithm GAII is:

[0193] Where T is the completion time of the project, LF j is the latest completion time of task j, and γ is a large number.

[0194] The smaller the fitness f, the better the fitness. From the calculation method of fitness, it can be seen that the project corresponding to the individual not only needs the shortest total completion time, but also needs to complete each task as early as possible, so as to reduce the impact of the randomness of the construction period on the project completion time.

[0195] Preferably, the deterministic sub-problem is solved using a priority rule scheduling algorithm as follows:

[0196] There are many classic priority rules that are widely used in deterministic RCPSP. The priority rule method is simple and efficient, but different priority solutions are different in different problem contexts. This paper selects 8 better priority rules to approximately solve the resource-constrained scheduling problem of variable project network random duration.

[0197] The following table lists eight better priority rules:

[0198]

[0199]

[0200] Table 2 8 priority rules

[0201] Where: d j ' is the expected duration of task j; LF j is the latest completion time of task j; EF j is the earliest completion time of task j; AP t is the set of all waiting tasks at time t; E(i,j) is the earliest start time of task i when task j starts at time t; S j are all the subsequent tasks of task j.

[0202] The heuristic algorithm based on the priority rule needs to consider the use of the scheduling generation mechanism. Common scheduling generation mechanisms include the serial scheduling generation mechanism (SSGS) and the parallel scheduling generation mechanism (PSGS). The MDPs model in this invention belongs to a parallel MDP. At a certain moment, multiple tasks can be started. Therefore, the parallel scheduling mechanism is also used when solving the variable project network with the priority rule. Define n to represent the stage, t n is the scheduling time, and the unstarted set W n , that is, the set of tasks that meet the immediate precedence constraint and the selection-related execution order constraint but have not yet been scheduled, and the set of tasks that can be started Q n , that is, the set of tasks that satisfy the immediate predecessor constraint, resource constraint, and selection-related execution order constraint but have not yet been scheduled, and the completion set F n , in the set P n , discard the set as N n , the parallel scheduling steps are as follows:

[0203] Step A: Determine the decision time of the current stage n. If P n There are tasks in t n If completed, the completed task is transferred to F n , and at the same time, according to the immediate preceding constraint, select the associated constraint update set W n ;

[0204] Step B: Let Q n is an empty set, in W n Find the tasks that meet the resource constraints and add them to Q n middle;

[0205] Step C: If Q n If Q is an empty set, jump to step A and make n+1 stage decisions; n If it is not empty, use the given priority rule to n Sort the tasks in Q n Select the task with the highest priority and check whether the selected task exists in the selection set C. e Or the association set β i If it exists, the selection association constraint check is performed. When the task a is triggerede , the main task of the associated set is to make i in the discarded set N n Medium, or C e There are already tasks in F n or P n In the process, the selected task needs to be abandoned, and the selected task will be removed from Q n Transfer to N n Otherwise, the selected task will be n Start working at any time and take it from Q n Transfer to P n , jump to step B;

[0206] Step D: Repeat the above steps until all tasks have been scheduled.

[0207] At present, no priority rules considering variable network structures have been found. During the project execution process, when a certain node is executed, other task nodes may not be executed. Based on this feature, the present invention proposes two new priority rules from the two aspects of execution time and the number of executed nodes. Among them, the SPRT rule considers the duration of all remaining unexecuted nodes, and the MTRS rule considers the number of all remaining unexecuted nodes.

[0208]

[0209] Table 3 Priority rules considering variable network structures

[0210] Where: is the set of unexecuted nodes remaining in the current project. Assuming that task j is selected, for the remaining unexecuted nodes in the current project, considering the selection of associated constraints, RJ can be updated j .

[0211] The algorithm of the present invention is implemented by Python 3.12 programming, and the resource flow model solution is solved by Gurobi9.0. The experimental results are obtained by running the computer program on a desktop computer with CPU 4.3GHx and 32G RAM, and the Pythonu operation is accelerated by the PyPy interpreter.

[0212] Experimental data:

[0213] The present invention constructs two examples, small-scale and large-scale. The small-scale comes from a case of automation transformation of a semi-finished product transfer process in a factory. The main content of the project is to design an automated transfer plan according to the characteristics of the product, produce, install and debug related equipment, and finally realize the automation of the production line. Under the project network, it contains 3 selections and 3 associations. If you start task 2 inventory unit design, you need to select a task from the sensor recognition plan and the visual recognition plan to execute. If you start task 3 robot unit design, you need to select one from the truss robot unit design and the fixed robot unit design to execute. After the hardware production of task 12 is completed, you need to choose whether to ship in pieces or to ship as a set after assembly and testing. If the visual recognition solution task 6 is selected, the visual algorithm development 11 must also be executed, otherwise it does not need to be executed. If it is shipped in pieces 16, it needs to be assembled and tested on site 18. If it is shipped as a set after testing 17, it only needs to be assembled 19. The resources of the project are renewable resources, including 4 R&D staff and 3 implementers. The specific basic information of each task node and its corresponding resources is shown in Table 4.

[0214]

[0215]

[0216] Table 4 Basic information of each task node and its corresponding resources

[0217] For large-scale examples, a variable network example set with a determined construction period is constructed by giving different network parameters, but the example set is not classified and analyzed. The present invention classifies the examples by network variable strength. The present invention defines two levels, low and high. The corresponding network parameters of each level are shown in Table 5. Under each set of parameters, 80 examples are selected from the defined J30 example set.

[0218]

[0219]

[0220] Table 5 Task duration randomness parameter settings

[0221] For small-scale actual cases and large-scale examples in each group, each example is set with two random intensities, low and high, and then a new example set required by the present invention is generated according to the different random intensities. Under different random intensities, 20% and 80% of all task nodes in the network have random durations. When the task nodes have random durations, the original duration of the task is used as the expected duration d j ', and then consider that it obeys the triangular distribution, according to the parameter settings in Table 6, the a, c, b parameter values ​​of the construction period under two different randomness intensities are obtained respectively.

[0222]

[0223] Table 6 Task duration randomness parameter settings

[0224] In summary, the present invention constructs 2 small-scale examples, and a total of 2*2*80=320 large-scale examples to participate in experimental testing. For the improved GA algorithm GAII, the population size is set to 80, the iteration is 50 times, and the insertion mutation probability m α is 0.2, and the probability of exchange mutation is m β The value of 0.2 is 0.2, all examples are simulated 500 times, and the mean of the objective function is taken for analysis. The present invention uses the relative deviation index (RDI) to evaluate the performance of each method under different scheduling environments.

[0225]

[0226] Among them, RDI mI Represents the RDI value of method m for solving example I, Cur mI represents the target mean obtained by solving Example I, Worst I Best I They represent the best and worst target means obtained by all methods for solving Example I.

[0227] MDPS is an exact solution algorithm, RF is an exact solution algorithm for deterministic subproblems in the approximate algorithm, and the improved GA algorithm GAII and GA as well as the priority rule scheduling algorithm are used to solve the proposed RF model and the overall approximate decision algorithm to solve the MDPs model. The results obtained in the small-scale experiment are shown in Table 7.

[0228]

[0229]

[0230] Table 7 Small-scale solution results

[0231] According to the small-scale solution results, the following conclusions can be drawn:

[0232] The random variable project network scheduling problem can be solved accurately using MDPS. The solution space will increase rapidly as the randomness increases, making it difficult to solve. Low random intensity and high random intensity have 4 random nodes and 16 random nodes respectively. High random intensity has 4 times more nodes than low random intensity, 10 times more states, and 176 times more solution time.

[0233] The approximate solution algorithm based on the RF model can effectively reduce the dimension of the state space and solve the problem of dimensionality curse. Although the approximate solution algorithm cannot obtain the optimal value, the deviation between the result and the optimal value is only 2%, and the time taken is much less than that of the precise algorithm. In actual projects, managers tend to make better decisions as quickly as possible rather than spending a lot of time pursuing the optimal solution. Therefore, the approximate solution algorithm has practical application value. At the same time, the improved GA algorithm GAII proposed in the present invention has the same solution results as those of RF and GA, which can prove that it is effective in solving the RF model, and can further solve the difficulty of not being able to solve large-scale problems when RF is accurately solved in the approximate solution algorithm.

[0234] In the large-scale experimental example, each method is used to solve the example and obtain the objective function, Table 8 and Figures 4 to 5 The mean RDI and 95% confidence interval corresponding to each method are shown respectively.

[0235]

[0236]

[0237] Table 8 Comparison of average RDI for large-scale examples

[0238] It can be seen from Table 8 that the RDI of the improved GA algorithm GAII is more than 70% lower than that of the GA algorithm. This shows that in most cases, the improved GA can find better sisters than GA.

[0239] The tables in the test environment are significantly better than the priority rule, and its overall RDI is improved by more than 80% compared with the optimal rule. Different random intensities basically do not affect the GA solution, as shown in Table 6 and Figure 5 As shown in Table 6 and Figure 6 shown.

[0240] Based on actual production conditions and existing test set examples, the present invention constructs multiple groups of experiments to verify the effectiveness of the constructed models and algorithms under different environments. The experimental results show that the approximate solution algorithm based on the resource flow model and GA can complete the calculation faster and can also be used to actually guide project managers to make decisions.

[0241] The present invention studies the project scheduling problem under variable network structure and discrete randomness of part of the task duration, and constructs a random model with the minimum project completion time as the decision-making goal based on MDPs theory. In view of the dimensionality disaster of state and action space faced when VI is accurately solved, it is proposed to approximate the project in a certain state into a deterministic project scheduling sub-problem with a variable network structure. By solving the sub-problem, the best start-up action in the current state is quickly found to start the work, thereby compressing the state space and action space, and significantly reducing the complexity of solution. For the deterministic scheduling sub-problem, the corresponding mixed integer programming model and GA algorithm are constructed, and the scheduling algorithm based on priority rules is not used as the evaluation benchmark. Based on actual cases, the present invention constructs small-scale examples and J30 example sets to construct large-scale computing experiments, and analyzes the performance of different scheduling algorithms under different random intensities and different network structure variable intensities. The experimental results show that the RF model and GA solution algorithm constructed by the present invention can obtain very good results in a small scale compared with the accurate results. In a large-scale environment, the GA algorithm is better than the priority rule algorithm, and has good decision-making guidance for actual decision-making.

[0242] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A project planning method under variable project network and random duration environment, characterized in that: The steps include: Step 1: Construct the MDPs model of variable project network under random duration environment; Step 2: Use stochastic dynamic programming to find the tight solution of the MDPs model; Step 3: The MDPs model is used to deal with small-scale problems, so an approximate decomposition framework is proposed to deal with large-scale random problems and decompose random problems into solutions to deterministic sub-problems; Step 4: Establish a resource flow model for the deterministic sub-problem; Step 5: Solve the deterministic sub-problem using the GA algorithm improved by meta-heuristics; Step 6: Solve the deterministic sub-problem using the priority rule scheduling algorithm; Step 7: Compare the effects of the approximate decomposition algorithm based on the improved GA algorithm.

2. A project planning method under a variable project network and random duration environment according to claim 1, characterized in that: Constructing an MDPs model of a variable project network under a random duration environment, wherein the MDPs model includes a decision time set T, a state set X, an action set A, a transition probability function p(x′|x,a), and a cost function c(x′|x,a); The decision time set T includes multiple decision times t, t∈T; The state set X is the project state and needs to contain all the project information needed for decision making; x = [t x ,W x ,E x ,G x ] T Indicates that x∈X; X is the set of all possible states of the project; where t x represents the decision moment in the current state; W x Indicates the set of tasks waiting to be started at the current moment; E x represents the set of choices that are waiting at the current moment, e∈E x ,e=1,2,...,E, when the task a is triggered e After completion, e is added to the selection set E x Middle; G x represents the set of tasks in progress at the current moment, then G x ={(j1,s1),(j2,s2),...,(j n ,s n )), where j n Indicates the nth in-progress task of the project at the current moment, s n It represents the remaining completion time of the nth task in progress at the current decision time. At the initial time of the project, all tasks have not started, and the initial state of the system is When all tasks are completed, the system will stay in a final state permanently. Where f is the state number; The action set A is the set of tasks waiting to be started based on the obtained task set W in any given state x∈X. x , waiting for the selection set E x , the set of tasks in progress G x , we get that if there are several pending actions a, each action a represents a task start combination that satisfies the task precedence constraint and resource constraint, multiple optional actions constitute an optional action set a∈A x ; If there is no task that meets the start condition in state x∈X, then but It will become a special action; The transition probability function p(x′|x,a) represents the state x taking a∈A x After that, the probability that the system transfers to state x′(x′≠x,x′∈X); The cost function c(x′|x,a) represents taking action a∈A in state x. x After that, the system transfers to the state x′(x′≠x,x,x′∈x) with immediate cost.

3. A project planning method under a variable project network and random duration environment according to claim 2, characterized in that: The state evolution process of the MDPs model is as follows: for a given state x, get the next possible state set Y, select the state x′ in the set Y as the current state, and then calculate the next new state set Y′, and traverse all the states in the set Y in turn until no new state is generated.

4. A project planning method under a variable project network and random duration environment according to claim 2, characterized in that: A task j∈W x When j is , the following conditions must be met: All task nodes in the must-execute task set V included in the immediate predecessor task set of task j have been completed; If task j is an associated execution task of task i, that is, j∈β i , task i must have been completed; If task j belongs to a certain choice, that is, j∈C e , triggering task a e Completed.

5. The project planning method under a variable project network and random duration environment according to claim 1, characterized in that: The specific solution of the MDPs model in stochastic dynamic programming is: Let V(x) denote the value function of state x∈X, For the possible state set after taking action a, the value iteration algorithm mainly solves the Bellman equation for each state x iteratively, and the equation is expressed as: During the iteration process, the optimal function value will continue to transfer from the next state to the previous state. When the value function value of all states no longer changes, it means that the value iteration algorithm converges and the optimal function value is V * (x0), x0 is the initial state; The corresponding optimal action formula for each state is: Where m represents the number of iterations.

6. A project planning method under a variable project network and random duration environment according to claim 1, characterized in that: The solution to decompose a random problem into deterministic sub-problems specifically includes: Obtain the currently unfinished task network, including the unscheduled task network and the in-progress task network, and use it as the network of the deterministic sub-problem; Get the remaining resources in the current state as the resource input for the deterministic sub-problem; For a task node with a random duration in a deterministic network, its remaining expected duration is taken as the duration of the task node.

7. The project planning method under a variable project network and random duration environment according to claim 1, characterized in that: The resource flow model for the deterministic sub-problem includes: Define the set of schedulable tasks under a certain state as K x Equal to the remaining unfinished nodes and in-progress nodes in state x, introducing a virtual start task and virtual termination tasks The completion time of the virtual task is equal to the current time of state x, that is, f s =t x ; The resource transfer quantity of the virtual start task is the remaining available resources of the current state x, m kx Represents, where k is the k-th resource; define task j∈K x ∪{s,e}The maximum resource transfer amount for the k-th resource is The maximum amount of resource transfer The formula is: The objective function that minimizes the total completion time of the project is expressed as: min f e , where f e Indicates the last completed task; The expression formula for the tasks in the set of tasks that must be executed is: The trigger task of the optional set is executed. Only one task can be selected in the set for execution. If the trigger task is not executed, all optional tasks will not be executed. The expression formula is: The relationship between the associated main task and all associated tasks is expressed as follows: v i -v j =0,j∈β i ; If a task is not executed, the expression formula for canceling the arc associated with the task is: In the associated tasks, the expression formula for the subtask to start only after the main task is completed is: The expression formula that the resource arc between two tasks cannot be circulated is: The expression formula that the resource arcs between the three tasks cannot be circulated is: The relationship between the amount of resource transfer and the maximum number of tasks required is expressed as: All resources of the task node will be transferred out and equal to the resource demand expression formula: The expression formula for all resource inputs of a node equal to resource demand is: If resources flow from i to j, the expression formula for task j to be completed later than task i is: The expression formula that the amount of resource transfer between tasks must be greater than zero is: The expression formula for whether there is resource flow between two tasks is: and ij ∈{0,1},i∈K x ∪{s},j∈K x ∪{e},i≠j。 8. The project planning method under a variable project network and random duration environment according to claim 1, characterized in that: The meta-heuristic improved GA algorithm is used to solve the deterministic sub-problems, including encoding, crossover and mutation operations, and fitness calculation; Encoding methods include: Individuals are represented by: where a E ; C E is the fixed item auxiliary information, indicating the Eth choice, a E is the selected trigger task, C E is a set of optional tasks, and the corresponding α E From C E When the task node selected in is not triggered and not executed, α E =-1; task list J = (j1, j2, ..., j n ) is a task sequence, the earlier the task is, the higher its priority is; Can be 0 or 1. Indicates that the task is not executed, otherwise it is executed; New sorting rules are introduced to describe the situation of task trigger selection. The specific sorting rules are: Trigger task a e Ranking among all optional task sets C e In front of the task in C e The tasks in a e The first task j∈C e For the selected task, C e The remaining tasks in the program are never executed; If the task a is triggered e There is a choice C m In the optional task set, task a m Must be placed in task a e in front of; The set of associated tasks β i The tasks in must be placed after the main task i; If the main task i of the associated task has a C e In the set, this main task i must also be ranked before task a e after; The specific process of crossover mutation operation is as follows: Multiple crossover points are randomly selected, and the two parents exchange the fragments between these crossover points to form new individuals; For the offspring individual mutation, there are two operations: insertion and exchange, and the probability of insertion mutation is m α ,0≤m α ≤1, the probability of exchange mutation is m β ,0≤m β ≤1. First, perform the insertion operation, and take a random number m from U(0,1) for each position in the task queue. When m<m α , randomly select a task after this position and insert it into this position, and then perform the exchange operation. For each position in the task queue, take a random number m in U(0,1), and when m<m β When , a task is randomly selected after the position and exchanged with the task at the position; The fitness calculation is as follows: The fitness expression formula of the improved GA algorithm GAII is: Where T is the completion time of the project, LF j is the latest completion time of task j, and γ is a large number.

9. The project planning method under a variable project network and random duration environment according to claim 1, characterized in that: The specific solution of the deterministic sub-problem using the priority rule scheduling algorithm is as follows: Define n to represent the stage, t n is the scheduling time, and the unstarted set W n , that is, the set of tasks that meet the immediate precedence constraint and the selection-related execution order constraint but have not yet been scheduled, and the set of tasks that can be started Q n , that is, the set of tasks that satisfy the immediate predecessor constraint, resource constraint, and selection-related execution order constraint but have not yet been scheduled, and the completion set F n , in the set P n , discard the set as N n , the parallel scheduling steps are as follows: Step A: Determine the decision time of the current stage n. If P n There are tasks in t n If completed, the completed task is transferred to F n , and at the same time, according to the immediate preceding constraint, select the associated constraint update set W n ; Step B: Let Q n is an empty set, in W n Find the tasks that meet the resource constraints and add them to Q n middle; Step C: If Q n If Q is an empty set, jump to step A and make n+1 stage decisions; n If it is not empty, use the given priority rule to n Sort the tasks in Q n Select the task with the highest priority and check whether the selected task exists in the selection set C. e Or the association set β i If it exists, the selection association constraint check is performed. When the task a is triggered e , the main task of the associated set is to make i in the discarded set N n Medium, or C e There are already tasks in F n or P n In the process, the selected task needs to be abandoned, and the selected task will be removed from Q n Transfer to N n Otherwise, the selected task will be n Start working at any time and take it from Q n Transfer to P n , jump to step B; Step D: Repeat the above steps until all tasks have been scheduled.

Citation Information

Patent Citations

  • Distributed new energy, energy storage and power distribution network planning method considering flexible investment

    CN112217202A

  • Cement production line construction operation multi-project multi-objective optimization scheduling modeling method

    CN117057584A

  • System and Method of Stochastic Resource-Constrained Project Scheduling

    US20140249882A1

Cited By

  • Scheduling optimization method and system for production project management

    CN120410167A

  • Priority rule method for random flexible project scheduling problem

    CN120672068A