A dynamically evolving human resource task matching method system for design crowdsourcing

Abstract

The present invention discloses a dynamically evolving human resource task matching method and system for design crowdsourcing, which belongs to the field of resource task planning. It includes: S1. Obtain all tasks that need to allocate human resources as a set of tasks to be allocated, and obtain all available human resources as a set of available human resources; S2. Perform resource task matching on the current set of tasks to be allocated and the set of available human resources; S3. Monitor whether there are tasks that have been completed, if so, update the set of tasks to be allocated, receive the evaluation value of the execution of the tasks, and enter step S4, otherwise, continue monitoring until all tasks are completed; S4. Input the evaluation value of the task into a multi-dimensional evaluation model of human resources, obtain the latest characteristic attributes of human resources, and update the set of available human resources, and enter step S2. The present invention evaluates the performance of resources according to the actual situation of the resources completing the tasks, updates the resource status in the resource set, thereby affecting the next resource task matching and improving the accuracy of resource task matching.

Description

Technical Field

[0001] The present invention belongs to the technical field of resource task planning, and more specifically, relates to a dynamically evolving human resource task matching method and system for design crowdsourcing. Background Art

[0002] With the rapid development of Internet technology, crowdsourcing has become a development trend as an open project task organization model. Crowdsourcing refers to the practice of a company or organization outsourcing work tasks that were previously performed by employees to a non-specific (and usually large) mass network on a voluntary basis. This model reduces the innovation cost of enterprises, meets the needs of both supply and demand, and has good development prospects.

[0003] Crowdsourcing projects for complex product design often consist of a large number of subtasks, requiring the participation of people with different expertise and in different fields. As the functions and structures of products become more and more complex, the number of tasks increases significantly, and the relationship between R&D tasks becomes more complex. At the same time, there are a large number of R&D, design, testing, and operation personnel in the project, and it is common for project personnel to master multiple skills. Due to the high complexity of products, the large number of tasks, and the continuous iteration of the entire product R&D, the project cycle is often very long. Therefore, the scheduling problem of product R&D projects has become more and more complex and there are many feasible solutions. It is of great significance to formulate reasonable scheduling plans for reference by enterprise managers in decision-making.

[0004] A realistic product development scheduling model and an efficient and accurate algorithm have become the key to scheduling enterprise product development tasks. In order to simplify the process, the existing project scheduling methods for multi-skilled personnel in product development are usually based on static evaluation methods, that is, factors such as human resource skill growth, service attitude changes, and cost changes are not considered. This is inconsistent with the actual execution of product development projects and cannot match the best resources for crowdsourcing tasks.

[0005] In view of the problems of resource management and evaluation, combinatorial explosion, etc. caused by the uncertainty (dynamicity) of human resource participants in the crowdsourcing process and the complexity of crowdsourcing tasks, there is an urgent need for a resource / task description model, a dynamically evolving matching and scheduling system and method to improve the matching accuracy of resources and tasks in the crowdsourcing process. Summary of the invention

[0006] In view of the defects and improvement needs of the prior art, the present invention provides a dynamically evolving human resource task matching method and system for design crowdsourcing, which aims to improve the matching accuracy of resource tasks in the crowdsourcing process.

[0007] To achieve the above object, according to a first aspect of the present invention, a dynamically evolving human resource task matching method for design crowdsourcing is provided, the method comprising:

[0008] S1. Obtain all tasks that need to allocate human resources as a set of tasks to be allocated, wherein the task is characterized by the characteristic attributes of the task itself and the demand for human resources for the task, and obtain all available human resources as a set of available human resources, wherein the human resources are characterized by the characteristic attributes of the human resources themselves;

[0009] S2. Matching the current set of tasks to be assigned and the current set of available human resources;

[0010] S3 monitors whether any tasks have been completed. If so, updates the set of tasks to be assigned, receives the evaluation value of the execution of the task, and proceeds to step S4. Otherwise, continues monitoring until all tasks are completed.

[0011] S4. Input the evaluation value of the task into the multi-dimensional evaluation model of human resources to obtain the latest characteristic attributes of human resources, and update the available human resources set, and enter step S2.

[0012] Preferably, the characteristic attributes of the task itself include: task number, estimated workload, budget cost, difficulty coefficient, planned start time and planned end time; the demand for human resources of the task includes: professional field and skill level required for the task; the characteristic attributes of the human resources themselves include: human resource number, professional field, skill level, learning ability, service attitude, unit time cost and timely completion rate.

[0013] Preferably, in step S2, the following matching priority principle is adopted:

[0014] 1) The skill level of human resources must not be lower than the skill level required for the task;

[0015] 2) Under the same circumstances, the one with the highest comprehensive score will be given priority;

[0016] 3) Only consider human resources with free time;

[0017] The calculation formula of the comprehensive score is as follows:

[0018]

[0019] Among them, Z n is the comprehensive evaluation value, T n , S n , P n 、F n , C n They are the skill level evaluation value, learning ability evaluation value, service attitude evaluation value, completion timeliness evaluation value, and unit time cost value of human resources, all of which are dimensionless; are the average skill level evaluation values, learning ability evaluation values, service attitude evaluation values, completion timeliness evaluation values, and unit time cost values ​​of all optional human resources in the task matching process, all dimensionless; δ1, δ2, δ3, δ4, and δ5 are weight coefficients, all dimensionless; subscript n is the evaluation value obtained after the nth task is completed.

[0020] Preferably, step S2 comprises:

[0021] S21. Initialize two resource task matching solution populations with a population size of NS, wherein the population is a feasible solution set, and the feasible solution is a feasible matching scheme of the task set to be assigned and the available human resource set;

[0022] S22. The Q values ​​of the two populations are initialized to Q1 and Q2 respectively;

[0023] S23. The Q-learning agent senses the states s1 and s2 of the two populations respectively, where the states are [ED, HV], where ED is the average distance between the Pareto solution and the ideal point, and HV is the volume of the target space dominated by the non-dominated Pareto frontier;

[0024] S24. Determine whether the critical point is reached. If the critical point is reached, select the best execution actions a1 and a2 for the two populations according to the action selection strategy, calculate the rewards r1 and r2 of actions a1 and a2, and update the Q value of the state action value function according to the rewards of the action execution. Otherwise, apply the random strategy to select actions to execute actions a1 and a2;

[0025] S25. Execute actions a1 and a2 in the two populations respectively to generate new populations, and add the Pareto solutions of the two new populations to the other population to accelerate population evolution;

[0026] S26. Then repeat S23-S25 until the termination condition is met.

[0027] Beneficial effects: The present invention preferably uses the above method (QLCA) for human resource task matching. Compared with other existing matching algorithms: (1) QLCA contains multiple efficient search strategies. QLCA can capture the current environmental characteristics to detect the current population state. Search is a function from state to action (search strategy). The performance feedback of the action search process is used to guide the algorithm to select iterative actions under different states to efficiently solve better solutions. Compared with the blind iteration of search strategies without state and knowledge in intelligent algorithms, the search is more efficient. (2) QLCA has the co-evolution of two populations. The two populations can exchange knowledge. Individuals with good performance on the Pareto frontier are added to each other's population to help faster evolution.

[0028] Preferably, the calculation formula of ED is as follows:

[0029]

[0030]

[0031] Among them, PF c represents the Pareto solution of the population, |PF c |Representing PF c The number of solutions, d(s ideal ,s i ) is PF c The i-th solution s in i and the ideal point s ideal The Euler distance of The solutions are s i The first and second target values ​​are divided by the worst values ​​of the corresponding targets, i.e. the normalized target values. For cost, the worst value is equal to the most expensive person assigned to each task. For the worst duration, assuming that all tasks are completed at the required skill level, they are completed serially and the duration is accumulated directly. and They are the normalized values ​​of the ideal point target values, and the normalized calculation is the same as the solution s i same.

[0032] Preferably, HV is calculated in the following manner:

[0033] (1) Invert the optimization target to obtain the inverted target space. The target inversion calculation is as follows:

[0034]

[0035] Among them, f j,worst are the worst values ​​of the jth target, f j is the jth objective value of the solution obtained, f j,revert is the target value after the jth target is reversed;

[0036] (2) Calculate HV in the new target space as follows:

[0037]

[0038] Where X is the non-dominant solution set obtained by the algorithm, P is the reference point corresponding to the true Pareto frontier, which is usually a vector formed by the maximum values ​​on each target, and v(x,P) is the hypervolume of the space formed between the solution x and the reference point P in the non-dominant solution set X, that is, the volume of the hypercube constructed with the line between the solution x and the reference point P as the diagonal.

[0039] Preferably, the calculation formula of the reward is as follows:

[0040]

[0041] Among them, HV offspring is the HV value of the offspring population, HV parent is the HV value of the parent population; the Q value update formula of the state action value function is as follows:

[0042]

[0043] Where g is the number of iterations, It is the maximum Q value in the state-action pair value corresponding to the new state s(g) after executing action a(g-1) in state s(g-1); r(g-1) is the reward for executing action a(g-1) in state s(g-1); 0≤α≤1 is the learning rate of reinforcement learning; 0≤γ≤1 is the discount coefficient, which is the impact of future rewards on the current situation.

[0044] Preferably, in step S4, the human resources multi-dimensional evaluation model includes:

[0045] Skill level evaluation model:

[0046] T n =T n-1 *q+t n *(1-q)

[0047]

[0048] Learning ability evaluation model:

[0049] Unit time cost evaluation model:

[0050] Service attitude evaluation model: P n =P n-1 *k+p n (1-k)

[0051] Completion timeliness evaluation model:

[0052]

[0053] D n =D n-1 +d n

[0054] Among them, T n 、T n-1 are the resource skill evaluation values ​​after the nth and n-1th tasks, dimensionless; q is a coefficient between [0,1]; t n is the skill evaluation value of the nth task; gn is the minimum skill level required for the nth task; h n is the difficulty coefficient of the nth task; v n is the comprehensive evaluation value after the nth task is completed, dimensionless; d n is the planned duration of the nth task, d′ n is the actual duration of the nth task, in man-days; c n is the budget cost of the nth task, c′ n is the actual cost of the nth task, in yuan; S n , S n-1 are the resource learning ability evaluation values ​​after the nth and n-1th tasks, dimensionless; w is a coefficient between [0,1]; β is the standard ability improvement coefficient, unit is 1 / man-day; C n , C n-1 are the resource unit time cost evaluation values ​​after the nth and n-1th tasks, dimensionless; m is a coefficient between [0,1]; α is the standard cost coefficient, which is the industry average unit time cost, in yuan / person-day; P n , P n-1 are the service attitude evaluation index after the nth and n-1th tasks, dimensionless; p n is the service attitude evaluation value of the nth task; k is a coefficient between [0,1]; F n 、F n-1 are the timely completion rate indexes after the nth and n-1th tasks respectively; D n , D n-1 They are the cumulative planned duration after the nth and n-1th tasks respectively, in man-days.

[0055] Beneficial effects: The present invention preferably uses the above-mentioned multi-dimensional evaluation model for human resources, because this model comprehensively considers data in both vertical and horizontal aspects: 1) Vertical data: It fully considers the growth data in the dynamic evolution of human resources, and does not simply use the completion of a single task to evaluate resource characteristics. 2) Horizontal data: When evaluating a single indicator of a resource, it considers the impact of multiple characteristics of the task on the indicator. This improves the objectivity and accuracy of human resource evaluation, thereby improving the accuracy of task resource matching.

[0056] Preferably, the pre-trained human resources multi-dimensional evaluation model is obtained by:

[0057] (1) Collect human resource task matching case collection;

[0058] (2) Experts evaluate the tasks and human resources in the case set respectively to obtain task evaluation values ​​and human resource evaluation values;

[0059] (3) Sampling from the task case set to generate a sample case set;

[0060] (4) Calculate the system parameters of the human resources multidimensional evaluation model through the least squares method to generate the human resources multidimensional evaluation model.

[0061] To achieve the above-mentioned object, according to a second aspect of the present invention, there is provided a dynamically evolving design crowdsourcing human resource task matching system, comprising: a computer-readable storage medium and a processor;

[0062] The computer-readable storage medium is used to store executable instructions;

[0063] The processor is used to read the executable instructions stored in the computer-readable storage medium to execute the above-mentioned dynamically evolving human resource task matching method for design crowdsourcing.

[0064] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects:

[0065] The present invention evaluates the performance of resources according to the actual situation of completing tasks by resources, and updates the resource status in the resource set, thereby affecting the matching of resources and tasks next time, and can better support the accuracy of resource and task matching in the design crowdsourcing process. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 A block diagram of a human resource task matching system for dynamically evolving design crowdsourcing provided by the present invention;

[0067] Figure 2 The pseudo code diagram of resource task matching provided by the present invention;

[0068] Figure 3 A flow chart of a dynamically evolving human resource task matching method for design crowdsourcing is provided for the present invention. DETAILED DESCRIPTION

[0069] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0070] like Figure 1 As shown, the present invention provides a dynamically evolving human resource task matching system for design crowdsourcing, including: a set of tasks to be assigned, a set of available resources, a resource-task matching planner, and a resource-task performance evaluator.

[0071] The set of tasks to be assigned includes all tasks that need to be assigned resources, and describes the characteristic attributes of the tasks and the demand constraints on resources. The task characteristic attribute information includes but is not limited to: the skill field and skill level required for the task, the planned completion time of the task, the task budget workload, the task budget cost, etc.

[0072] Available resource set: includes all available resources, especially human resources, and describes the multi-dimensional characteristics of resources. The resource characteristic attribute information includes but is not limited to: resource rate (hourly price), resource service attitude index, skill field and level, learning ability index, completion timeliness rate, etc. The human resources can be designers or design companies.

[0073] Resource task matching planner: Use heuristic optimization algorithm to achieve resource and task matching scheduling. The resource task matching planner can use genetic algorithm and reinforcement learning algorithm for matching planning.

[0074] Resource task performance evaluator: Based on the task performance evaluation model and resource performance evaluation model, the performance of resources is evaluated according to the actual situation of the resources completing the task. The evaluation indicators of the task performance evaluation model include service attitude, timeliness, completion quality, labor rate, skill level, etc.; the evaluation indicators of the resource performance evaluation model include resource rate (man-hour unit price), resource service attitude index, skill field and level, learning ability index, completion timeliness, etc.

[0075] Resource task matching planner

[0076] The present invention adopts a dual population co-evolution algorithm based on Q-learning to construct a resource task matching planner. Figure 2As shown, firstly, a population initialization strategy is adopted to generate a certain size of two resource task matching solution populations, see row 1; here, "population" refers to a set of feasible solutions generated at one time, each feasible solution is an individual, and multiple feasible solutions constitute a population. The resource task matching of the present invention refers to a feasible matching scheme between the task set and the resource set. "Feasible" means that this matching scheme satisfies all preset constraints. However, a "feasible" solution is not necessarily the optimal solution, and it is necessary to find a solution with a lower total cost and less total time through an optimization algorithm. After that, the Q values ​​in the state action table of the two populations are initialized respectively. At this time, the selection probability of each action in each state is equal, see row 2. In row 4, the Q-learning agent perceives the state of the two populations respectively; because the quality of the population solution is relatively low at the beginning, the population state changes greatly during the search process, and cannot correctly reflect the quality of the action strategy, so a critical point is set (set to a certain proportion of the defined state indicator value), and before that, a random selection is made between the actions (see row 8), and row 5 is used to determine whether the critical point is reached. If the critical point is reached, the best execution action is selected for each of the two populations according to the action selection strategy, as shown in line 6. The action is executed to generate a new population (line 10), and then the strategy (executing the action in the previous state) is evaluated, and the Q value of the state action value function is updated according to the reward of the action execution, as shown in lines 12-13. The Pareto solutions of the two populations are added to each other's population to speed up population evolution, as shown in lines 15-17, and then the above operations are repeated until the termination condition is met (line 3).

[0077] 1) Status definition

[0078] For the problem of dynamically evolving resource and task matching scheduling, a population-based solution optimization method is used. For the new population in each iteration, it is described as the state in Q-learning. The Pareto solution set composed of the non-dominated solutions of the population is the final solution to the problem. Convergence and diversity are the main indicators to measure the quality of the Pareto solution. Therefore, the environmental state is described by the following two multi-objective indicators: Euler distance (ED) and hyper volume (HV).

[0079] Euler distance (ED): This indicator is used to measure the average distance between the Pareto solution obtained by the solution algorithm and the ideal point, reflecting the convergence of the obtained Pareto solution - a lower value means that the potential Pareto frontier (the mapping of the Pareto solution in the target space) is closer to the real Pareto. Because the goal of optimizing the dynamically evolving resource and task matching scheduling problem is to minimize cost and duration, the ideal point is considered to be (0,0) in this invention. The value of this indicator is calculated as follows:

[0080]

[0081]

[0082] In the formula, PF c represents the Pareto solution of the population, |PF c |Representing PF c The number of solutions, d(s ideal ,s i ) indicates PF c The i-th solution s in i and the ideal point s ideal The Euler distance of Respectively represent the solution s i The first and second target values ​​are divided by the worst value of the corresponding target, that is, the normalized target value. For cost, the worst value is equal to the most expensive person assigned to each task; for the worst duration, assuming that all tasks are completed at the required skill level, they are completed serially and the duration is accumulated directly. and They represent the normalized values ​​of the ideal point target values, and the normalized calculation is the same as the solution s i same.

[0083] It is worth noting that the ED used to describe the environmental state is normalized by the ED of the initial population. The normalized ED is divided into five intervals [0, 0.6], (0.6, 0.7], (0.7, 0.8], (0.8, 0.9], (0.9, +∞) for use in the state definition table.

[0084]

[0085] In the formula, ED g and ED 0 is the average Euler distance between the g-th iteration and the initial state.

[0086] Hyper Volume (HV): This metric is used to evaluate the convergence and diversity of the obtained Pareto solutions. This metric is used to calculate the area / volume of the target space dominated by the obtained non-dominated Pareto front. It is also called the Size of Space Covered (SSC).

[0087] The present invention needs to invert the obtained target value when calculating HV, so the optimization goal becomes to find the maximum value of the inverted target, taking the (0,0) point as the reference point, so the larger the HV value, the better the convergence and diversity of the obtained Pareto solution. The target inversion calculation is as follows:

[0088]

[0089] In the formula, f j,worst Respectively represent the worst value of the j-th target, f jrepresents the jth objective value of the solution obtained, f j,revert The worst value calculation of the cost and duration targets involved in the present invention is as described in ED above.

[0090] Let X be the non-dominant solution set obtained by the algorithm, and P be the reference point corresponding to the true Pareto frontier, which is usually the vector formed by the maximum values ​​on each target. Then, the hypervolume from the non-dominant solution set to the true Pareto frontier, i.e., the HV index, is calculated as follows:

[0091]

[0092] Among them, v(x,P) represents the hypervolume formed by the space between solution x and reference point P in the non-dominant solution set X, that is, the volume of the hypercube constructed with the line between solution x and reference point P as the diagonal. Therefore, the HV index can also evaluate the convergence and diversity of the dominant solution set at the same time. A larger HV index value of a non-dominant solution set means that the solution set is closer to the true Pareto front in terms of convergence and diversity, and is a better non-dominant solution set.

[0093] Similarly, HV is calculated by the preset maximum hypervolume HV max Normalized. The normalized HV is divided into 5 intervals [0,0.7], (0.7,0.8], (0.8,0.9], (0.9,1], (1, +∞).

[0094]

[0095] Where HV g is the population hypervolume at the g-th iteration.

[0096] Regarding the above-mentioned normalized ED and HV, the present invention defines a total of 25 states, and the state definitions are shown in Table 1.

[0097]

[0098] Table 1

[0099] 2) Action description

[0100] The action operator in the dual-population co-evolution algorithm based on Q-learning consists of the action operator (LSKS) designed by the present invention and three multi-objective intelligent algorithms (NSGA-II, SMS-EMOA, and NTGA).

[0101] The action operator (LSKS) designed by the present invention includes two parts, namely local search (LocalSearch) and knowledge search (Knowledge based Search). Local search mainly searches the field of the current solution to find potential better solutions, and knowledge search uses the knowledge of the current population solution to guide the search for solutions. The present invention adopts a single coding structure. In the search stage, it mainly searches for the assignment of task personnel, and the sorting of tasks is completed in the scheduling generator.

[0102] Local search maintains the number of assignable personnel for each task j in the solution Initially, the task is initialized according to the number of people who can perform the task; afterwards, during the decoding process of the scheduling generator, due to the improvement of the skill level of the personnel, they can be assigned to tasks with higher skill levels that they could not perform before. Then the assignable personnel of the task will join the personnel, and the number of people who can perform the task at this time will be counted. to update.

[0103] The local search process uses the personnel reallocation operator. Referring to the fruit fly algorithm, the local search uses the reallocation operator to generate NS neighboring solutions for each solution, and the best neighboring solution replaces the original solution. The personnel reallocation operator randomly selects a task index linked list TL from each solution. Task j greater than 1, then regenerate a genetic code for the task satisfy Ensure the allocation of new personnel, θ j is the original genetic code for task j. This ensures efficient personnel redistribution in local searches.

[0104] Knowledge search is to use the personnel allocation knowledge of previous elite solutions to guide the search. The search process needs to maintain a global number of assignable personnel for each task j Before the project is scheduled, it is initialized according to the number of people who can execute the task; if the task corresponds to Greater than When, yes Update. Search uses knowledge to regenerate genes for each solution. Gene is a decimal between 0 and 1, so for each task j, the interval range corresponding to the gene is Divide into equal parts, The probability of a gene being generated in each interval is initialized as follows. Update the probability to re-initialize ( It will stabilize quickly during algorithm iteration).

[0105] Each generation of probability updates is performed using the NE best solutions in the population, as shown in the following equation:

[0106]

[0107]

[0108]

[0109]

[0110] In the formula, g is the number of iterations, ∈ represents the learning rate of knowledge accumulation, NE is the number of elite solutions used to update knowledge, The gene corresponding to the task j that solves the elite s. Using knowledge guidance can not only guide the task selection of personnel who can perform the task originally, but also guide the selection of personnel who are added to the set of executable personnel due to the improvement of their learning skills.

[0111] Many excellent algorithms have emerged for solving multi-objective optimization problems. Therefore, the present invention uses these excellent algorithms as single operators, that is, corresponding to the actions in reinforcement learning, and selects appropriate algorithms for search in different environments. The second generation of fast non-dominated sorting genetic algorithm (NSGA-II) was proposed by Kalyanmoy Deb et al. It is a classic multi-objective algorithm and is widely used in solving multi-objective optimization problems. The evolutionary multi-objective algorithm based on hypervolume index selection (SMS-EMOA) is an algorithm that mixes the hypervolume index with non-dominated sorting as a selection operator to optimize the hypervolume dominated in the iteration process. The present invention compares it with SPEA2, ε-MOEA and NSGA-II through simulation calculation experiments to prove that it is superior in coverage and hypervolume indicators. Therefore, NSGA-II and SMS-EMOA are used as action operators of the present invention. For the multi-skilled personnel project scheduling (MSRCPSP) problem, Laszczyk et al. modified the selection operator of NSGA-II and added a copy prevention method. The experimental results show that the proposed non-dominated tournament genetic algorithm (NTGA) improves the convergence, spread and uniformity of the solution. Therefore, NTGA is also used as the action operator of reinforcement learning.

[0112] 3) State-Action Pair Table

[0113] The state-action pairs are shown in Table 2. The Q-learning method uses this table to decide which action to perform in a given state based on a given selection policy.

[0114]

[0115] Table 2

[0116] 4) Action selection strategy

[0117] For a given environment state s(g), a selection policy π(s(g)) is used by the agent to select actions. For each candidate action a i (i=1,…,NA, where NA is the number of candidate actions), defines a selection probability p(s(g), a i ), the probability is determined by the Q-value in the state-action table, and the formula is as follows:

[0118]

[0119] 5) Action Rewards

[0120] After executing action a(g), a corresponding reward r(g) is used to evaluate the performance of the action executed in state s(g), and to update the corresponding state-action pair value Q(s(g), a(g)). In multi-objective optimization problems, hypervolume (HV) is a general measure of the convergence and diversity of the obtained non-dominated Pareto front solutions. A larger HV indicates better convergence and diversity. Here, the ratio of the increase in the HV value of the offspring population compared to the parent population is taken as the reward r(g) for action a(g) in state s(g):

[0121]

[0122] The goal of reinforcement learning is to maximize the total reward obtained from solving the problem.

[0123] 6) Q value update

[0124] The rules in Q-learning are used to update the Q value Q(s(g-1), a(g-1)) of the state-action pair as follows:

[0125]

[0126] in, It is the maximum Q value in the state-action pair value corresponding to the new state s(g) after executing action a(g-1) in state s(g-1); r(g-1) is the reward for executing action a(g-1) in state s(g-1); 0≤α≤1 represents the learning rate of reinforcement learning; 0≤γ≤1 represents the discount coefficient, which indicates the impact of future rewards on the current situation.

[0127] Resource Task Performance Evaluator

[0128] The role of the resource task performance indicator evaluator is to evaluate the performance of resources based on the task performance evaluation model and the resource performance evaluation model according to the actual situation of the resources completing the task.

[0129] Task performance evaluation model

[0130] The evaluation indicators of the task performance evaluation model are defined as: difficulty coefficient, service attitude, performance evaluation, actual cost, actual workload, etc.

[0131] The difficulty coefficient is an indicator of the technical difficulty reflected after the task is completed. The difficulty coefficient is given by the review experts after the task is completed. The service attitude index ranges from 0.5 to 1.5. A value of 1 means a medium service attitude, a value greater than 1 means a service attitude that exceeds expectations, and a value less than 1 means a service attitude that is lower than expectations. The larger the value, the better the service attitude, and the lower the value, the worse the service attitude. The service attitude is given by the project task entrustor after the task is completed.

[0132] The performance evaluation index range is defined as 0.5 to 1.5, and the larger the value, the better the delivery quality. The performance evaluation index is given by the task entrustor and the review expert after the task is completed.

[0133] The actual cost is the final project settlement expenditure, and the unit of measurement is (yuan).

[0134] The actual workload is the total workload actually incurred by the project calculated at the time of project acceptance, and the unit of measurement is (man-day);

[0135] Resource Performance Evaluation Model

[0136] Through the expert mode, a multi-dimensional resource indicator model with the goal of accurate task-resource matching is established. The present invention establishes a five-dimensional indicator model of resource skill level, learning ability, service cost, service attitude, and task completion timeliness.

[0137] (1) Resource skill level evaluation model

[0138] A resource may contain multiple skills, each with a level measurement. Whether the skill level of a resource meets the skill requirements of the task is the most fundamental constraint for task-resource matching.

[0139] T n =T n-1 *q+t n *(1-q)

[0140]

[0141] Where: T n 、T n-1 are the resource skill evaluation values ​​after the nth and n-1th tasks, respectively, dimensionless; q is a coefficient between [0,1], calculated by the least squares method based on the sample set data. n is the skill evaluation value of the nth task; g nis the minimum skill level required for the nth task; h n is the difficulty coefficient of the nth task; v n is the comprehensive evaluation value after the nth task is completed (dimensionless); d n is the planned duration of the nth task, d′ n is the actual duration of the nth task, with the dimension (man-day); c n is the budget cost of the nth task, c′ n is the actual cost of the nth task, with the unit of (yuan).

[0142] (2) Resource learning capability evaluation model

[0143] Resources have learning capabilities and can accumulate experience and improve their skills during task execution. The level of resource learning capability reflects the efficiency of resources improving their skills through task execution.

[0144]

[0145] Among them, S n , S n-1 are the resource learning ability evaluation values ​​after the nth and n-1th tasks, dimensionless; w is a coefficient between [0,1], calculated by the least squares method based on the sample set data. n 、T n-1 are the resource skill evaluation values ​​after the nth and n-1th tasks respectively; d′ n is the actual duration of the nth task, with the dimension of (man-day); β is the standard capacity improvement coefficient, with the dimension of (1 / man-day), which is obtained by machine learning through the sample data set. You can also directly take β = 0.01.

[0146] (3) Resource unit time cost evaluation model

[0147] Resources have a unit time cost. The cost of using resources will be one of the important factors to consider in task matching.

[0148]

[0149] Among them, C n , C n-1 are the resource unit time cost evaluation values ​​after the nth and n-1th tasks, respectively, dimensionless; m is a coefficient between [0,1], calculated by the least squares method based on the sample set data. n 、T n-1 are the resource skill evaluation values ​​after the nth and n-1th tasks respectively; d n is the planned duration of the nth task, with the dimension (man-day); c′ nis the actual cost of the nth task, with the unit of (yuan). α is the standard cost coefficient, which is the average unit time cost of the industry, with the unit of (yuan / person-day). This value is obtained from the sample data set.

[0150] (4) Resource service attitude evaluation model

[0151] Service attitude is the subjective evaluation of resources by resource users. The quality of service attitude is one of the factors affecting task resource matching.

[0152] P n =P n-1 *k+p n (1-k)

[0153] Among them, P n , P n-1 are the service attitude evaluation index after the nth and n-1th tasks, dimensionless; p n is the service attitude evaluation value of the nth task. k is a coefficient between [0,1], which is calculated by the least square method based on the sample set data.

[0154] (5) Resource timely completion rate evaluation model

[0155] Timely completion rate is an important evaluation indicator of resource work efficiency and contract performance quality. It is also one of the key factors affecting task resource matching.

[0156]

[0157] D n =D n-1 +d n

[0158] Among them, F n 、F n-1 are the timely completion rate indexes after the nth and n-1th tasks respectively; d n is the planned duration of the nth task, d′ n is the actual duration of the nth task, with the unit of measurement being (man-day); D n , D n-1 They are the cumulative planned duration after the nth and n-1th tasks respectively, and the unit of measurement is (man-day).

[0159] (6) Comprehensive resource evaluation model

[0160] When matching task resources, if the skill level, unit time cost, learning ability, service attitude, and timely completion rate all meet the requirements of the task, and if there are multiple resources to choose from, you need to select from high to low based on the comprehensive evaluation value of the resources.

[0161]

[0162] Among them, Z n It is a comprehensive evaluation value, which is dynamically calculated during the task resource matching process and is a comprehensive score relative to other competing resources. The result of each calculation will be different due to different competing resources. n , S n , P n 、F n , C n They are the resource's latest skill evaluation value, learning ability evaluation value, service attitude evaluation value, timely completion rate evaluation value, and unit time cost value, all dimensionless; are the average values ​​of skill evaluation, learning ability evaluation, service attitude evaluation, timely completion rate evaluation, and unit time cost of all optional resources in the task matching process, all dimensionless; δ1, δ2, δ3, δ4, and δ5 are weight coefficients, all dimensionless. The values ​​of δ1, δ2, δ3, δ4, and δ5 are obtained by machine learning of the sample set.

[0163] like Figure 3 As shown, the present invention provides a dynamically evolving human resource task matching method for design crowdsourcing, comprising the following steps:

[0164] Step 1: Input the set of tasks to be assigned and the set of available resources. The characteristic description information of the tasks and resources must be complete;

[0165] Step 2: Use the resource-task matching planner to match resources and tasks based on a heuristic algorithm based on priority rules to generate a resource-task matching plan;

[0166] Step 3: After the project task is executed, the project task entrustor evaluates the completion of the project task according to the task evaluation model and generates the evaluation value of the task evaluation index;

[0167] Step 4: According to the task evaluation value completed by the resource, the resource characteristic value is recalculated based on the resource multi-dimensional evaluation model to obtain the latest characteristic attribute value of the resource;

[0168] Step 5: Update the resource feature values ​​in the resource library with the latest feature attribute values ​​of the resource so that they can be used for the next matching task.

[0169] Furthermore, the steps for generating the resource multi-dimensional evaluation model in step 4 are as follows:

[0170] 1) Collect resource task case collection;

[0171] 2) An expert group evaluates the tasks and resources in the case set and gives characteristic values.

[0172] 3) Generate a sample case set by sampling from the task case set;

[0173] 4) Calculate the system parameters of the resource evaluation model through the least squares method to generate the evaluation model.

[0174] Example

[0175] The task set description is shown in Table 3:

[0176]

[0177] Table 3

[0178] Each task requires a resource with certain skills. For example, task t1 requires professional field A and level 3.0 or above technology to be competent. The estimated workload is 30 man-days, the budget cost is 10,000 yuan, the difficulty coefficient is 0.9, the planned start time is September 3, 2019, and the planned end time is October 4, 2019.

[0179] The resource set description is shown in Table 4:

[0180]

[0181] Table 4

[0182] There are four candidate resources in the resource set, U1, U2, U3, and U4. Each resource has characteristic indicators such as professional skills and learning ability, service attitude, cost index, and timeliness. These characteristic indicators are obtained after historical task evaluation. For newly added resources, their evaluation indicators are evaluated by self-evaluation and expert evaluation to get an initial value. For example, U4 is a newly added resource. Its professional skill level and learning ability are described as "A: 1.0: 1.0, C: 4.0: 1.0", which means that the resource has two professional skills A and C respectively. U4's A skill level is 1.0, and A skill learning ability is 1.0; U4's C skill level is 4.0, and C skill learning ability is 1.0. The initial value of U4's service attitude is 1.0, the cost index is 300 (default labor time fee), and the initial value of task completion timeliness is 1.00.

[0183] The service attitude, cost index, and timeliness rate of U1, U2, and U3 are calculated by the system based on their historical tasks.

[0184] Task set data and resource set data can be stored in the working directory in the form of Excel files.

[0185] Resource task matching

[0186] Call the resource task matching planner and specify the task set data file and resource set data file. The resource task matching planner generates a resource task matching solution based on the built-in optimization matching algorithm.

[0187] Matching priority rules:

[0188] 1) The resource skill level must be higher than the skill level required for the task;

[0189] 2) Under the same circumstances, the one with the higher comprehensive score in terms of cost, service attitude and timeliness will be given priority;

[0190] 3) Only consider resources with idle time.

[0191] The resource task matching solution is shown in Table 5:

[0192]

[0193] Table 5

[0194] Evaluate the task after it is completed

[0195] After the resources are matched with the tasks, the tasks are executed. After the tasks are completed, the task entrustor and the task review experts evaluate the execution of the tasks, and the evaluation results are shown in Table 6.

[0196]

[0197] Table 6

[0198] Calculate and update the resource evaluation value based on the task evaluation value.

[0199] Based on the evaluation of the tasks performed, the various resource indicators are re-evaluated and updated using the resource multi-dimensional evaluation model.

[0200]

[0201]

[0202] Table 7

[0203] After a resource completes each task, its evaluation index will be recalculated based on the resource multi-dimensional evaluation model. The calculation results are shown in Table 7.

[0204] For example, for resource U1, after completing task t1, its professional level of skill A was evaluated as 4.44, its learning ability of skill A was evaluated as 0.44, its service attitude was evaluated as 0.93, its cost rate was evaluated as 382.50, and its timely completion index was evaluated as 0.96.

[0205] When U1 completes task t2, it will make another evaluation adjustment based on the previous evaluation and get a new evaluation value, which will have an impact on the next task allocation.

[0206] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A dynamically evolving human resource task matching method for design crowdsourcing, characterized in that: The method includes: S1. Obtain all tasks that need to allocate human resources as a set of tasks to be allocated, wherein the task is characterized by the characteristic attributes of the task itself and the demand for human resources for the task, and obtain all available human resources as a set of available human resources, wherein the human resources are characterized by the characteristic attributes of the human resources themselves; S2. Matching the current set of tasks to be assigned and the current set of available human resources; S3 monitors whether any tasks have been completed. If so, updates the set of tasks to be assigned, receives the evaluation value of the execution of the task, and proceeds to step S4. Otherwise, continues monitoring until all tasks are completed. S4. Input the evaluation value of the task into the multi-dimensional evaluation model of human resources to obtain the latest characteristic attributes of human resources, and update the available human resources set, and enter step S2; In step S2, the following matching priority principle is adopted: 1) The skill level of human resources must not be lower than the skill level required for the task; 2) Under the same circumstances, the one with the highest comprehensive score will be given priority; 3) Only consider human resources with free time; The calculation formula of the comprehensive score is as follows: Among them, Z n is the comprehensive evaluation value, T n , S n , P n 、F n , C n They are the skill level evaluation value, learning ability evaluation value, service attitude evaluation value, completion timeliness evaluation value, and unit time cost value of human resources, all of which are dimensionless; are the average values ​​of skill level evaluation, learning ability evaluation, service attitude evaluation, completion timeliness evaluation, and unit time cost of all optional human resources in the task matching process, all dimensionless; δ1, δ2, δ3, δ4, and δ5 are weight coefficients, all dimensionless; subscript n is the evaluation value obtained after the nth task is completed; Step S2 includes: S21. Initialize two resource task matching solution populations with a population size of NS, wherein the population is a feasible solution set, and the feasible solution is a feasible matching scheme of the task set to be assigned and the available human resource set; S22. The Q values ​​of the two populations are initialized to Q1 and Q2 respectively; S23. The Q-learning agent senses the states s1 and s2 of the two populations respectively, where the states are [ED, HV], where ED is the average distance between the Pareto solution and the ideal point, and HV is the volume of the target space dominated by the non-dominated Pareto frontier; S24. Determine whether the critical point is reached. If the critical point is reached, select the best execution actions a1 and a2 for the two populations according to the action selection strategy, calculate the rewards r1 and r2 of actions a1 and a2, and update the Q value of the state action value function according to the rewards of the action execution. Otherwise, apply the random strategy to select actions to execute actions a1 and a2; S25. Execute actions a1 and a2 in the two populations respectively to generate new populations, and add the Pareto solutions of the two new populations to the other population to accelerate population evolution; S26. Then repeat S23-S25 until the termination condition is met; In step S4, the human resources multi-dimensional evaluation model includes: Skill level evaluation model: T n =T n-1 *q+t n *(1-q) Learning ability evaluation model: Unit time cost evaluation model: Service attitude evaluation model: P n =P n-1 *k+p n (1-k) Completion timeliness evaluation model: D n =D n-1 +d n Among them, T n 、T n-1 are the resource skill evaluation values ​​after the nth and n-1th tasks, dimensionless; q is a coefficient between [0, 1]; t n is the skill evaluation value of the nth task; g n is the minimum skill level required for the nth task; h n is the difficulty coefficient of the nth task; v n is the comprehensive evaluation value after the nth task is completed, dimensionless; d n is the planned duration of the nth task, d′ n is the actual duration of the nth task, in man-days; c n is the budget cost of the nth task, c′ n is the actual cost of the nth task, in yuan; S n , S n-1 are the resource learning ability evaluation values ​​after the nth and n-1th tasks, dimensionless; w is a coefficient between [0, 1]; β is the standard ability improvement coefficient, unit is 1 / man-day; C n , C n-1 are the resource unit time cost evaluation values ​​after the nth and n-1th tasks, dimensionless; m is a coefficient between [0, 1]; α is the standard cost coefficient, which is the industry average unit time cost, in yuan / person-day; P n , P n-1 are the service attitude evaluation index after the nth and n-1th tasks, dimensionless; p n is the service attitude evaluation value of the nth task; k is a coefficient between [0, 1]; F n 、F n-1 are the timely completion rate indexes after the nth and n-1th tasks respectively; D n , D n-1 They are the cumulative planned duration after the nth and n-1th tasks respectively, in man-days.

2. The method according to claim 1, characterized in that The characteristic attributes of the task itself include: task number, estimated workload, budget cost, difficulty coefficient, planned start time and planned end time. The demand for human resources of the task includes: professional field and skill level required for the task; the characteristic attributes of the human resources themselves include: human resource number, professional field, skill level, learning ability, service attitude, unit time cost and timely completion rate.

3. The method according to claim 1, characterized in that The calculation formula of ED is as follows: Among them, PF c represents the Pareto solution of the population, |PF c |Representing PF c The number of solutions, d(S ideal， s i ) is PF c The i-th solution s in i and the ideal point s ideal The Euler distance of The solutions are s i The first and second target values ​​are divided by the worst values ​​of the corresponding targets, i.e. the normalized target values. For cost, the worst value is equal to the most expensive person assigned to each task. For the worst duration, assuming that all tasks are completed at the required skill level, they are completed serially and the duration is accumulated directly. and They are the normalized values ​​of the ideal point target values, and the normalized calculation is the same as the solution s i same.

4. The method according to claim 1, characterized in that HV is calculated as follows: (1) Invert the optimization target to obtain the inverted target space. The target inversion calculation is as follows: Among them, f j，worst are the worst values ​​of the jth target, f j is the jth objective value of the solution obtained, f j，revert is the target value after the jth target is reversed; (2) Calculate HV in the new target space as follows: Where X is the non-dominant solution set obtained by the algorithm, P is the reference point corresponding to the true Pareto frontier, which is the vector formed by the maximum values ​​on each target, and v(x, P) is the hypervolume of the space formed between the solution x and the reference point P in the non-dominant solution set X, that is, the volume of the hypercube constructed with the line between the solution x and the reference point P as the diagonal.

5. The method according to claim 1, characterized in that The reward calculation formula is as follows: Among them, HV offspring is the HV value of the offspring population, HV parent is the HV value of the parent population; the Q value update formula of the state action value function is as follows: Where g is the number of iterations, It is the maximum Q value in the state-action pair value corresponding to the new state s(g) after executing action a(g-1) in state s(g-1); r(g-1) is the reward for executing action a(g-1) in state s(g-1); 0≤α≤1 is the learning rate of reinforcement learning; 0≤γ≤1 is the discount coefficient, which is the impact of future rewards on the current situation.

6. The method according to claim 1, characterized in that The pre-trained human resources multi-dimensional evaluation model is obtained in the following way: (1) Collect human resource task matching case collection; (2) Experts evaluate the tasks and human resources in the case set respectively to obtain task evaluation values ​​and human resource evaluation values; (3) Sampling from the task case set to generate a sample case set; (4) Calculate the system parameters of the human resources multidimensional evaluation model through the least squares method to generate the human resources multidimensional evaluation model.

7. A dynamically evolving crowdsourcing human resource task matching system, characterized in that: include: A computer readable storage medium and a processor; The computer-readable storage medium is used to store executable instructions; The processor is used to read the executable instructions stored in the computer-readable storage medium to execute the dynamically evolving human resource task matching method for design crowdsourcing as described in any one of claims 1 to 6.

Images