Multi-target dynamic resource scheduling method for underwater robot

Abstract

The invention discloses an underwater robot multi-target dynamic resource scheduling method, and relates to the technical field of robot resource scheduling. The method comprises the following steps: constructing an underwater robot resource scheduling problem model, and establishing a target optimization model which comprises the steps of minimizing task maximum execution time, minimizing total power consumption and balancing load; generating reference points and distributing the reference points in a target space; a population is coded and initialized, a scheduling scheme is represented by adopting multi-segment chromosome coding, the scheduling scheme comprises a task allocation sequence and tasks allocated by a robot, and an initial parent population is generated in a random mode; a non-dominated solution set is obtained through evolutionary iteration, and the non-dominated solution set is obtained through reference point niche selection and combined evolutionary iteration of crossover and mutation operation based on an improved NSGA-III algorithm; and obtaining an optimal scheduling scheme, and screening an optimal compromise solution from the non-dominated solution set as the optimal scheduling scheme. According to the invention, a data-driven intelligent scheduling scheme is provided for collaborative operation of underwater robots in a complex marine environment.

Description

Technical Field

[0001] This invention belongs to the field of robot resource scheduling technology, and particularly relates to a multi-objective dynamic resource scheduling method for underwater robots. Background Technology

[0002] With the increasing demand for ocean exploration and underwater operations, the application of underwater robots in fields such as marine scientific research, deep-sea exploration, environmental monitoring, and emergency rescue is gradually expanding. These robots typically perform tasks in harsh marine environments, and their task efficiency is closely related to the scheduling of robot resources. Underwater robots need to coordinate and optimize the use of multiple resources, including computing resources (such as CPU and memory), energy resources (such as batteries and fuel), and communication resources (such as data bandwidth and signal processing capabilities). Therefore, underwater robot resource scheduling technology is of great significance for improving task execution efficiency, extending robot operation time, and ensuring the safety and accuracy of tasks.

[0003] A major challenge in resource scheduling is the efficient allocation of limited computing and energy resources. Underwater robots typically perform computationally intensive tasks such as data processing, image recognition, and environmental modeling, which place high demands on computing resources. Simultaneously, the power resources required for robot operation are limited, and battery life directly impacts task completion time. In this resource-constrained environment, a rational scheduling strategy can effectively improve robot operational efficiency and avoid resource waste.

[0004] Furthermore, underwater robots often perform tasks in swarms, making the coordination of resource sharing and task allocation among multiple robots another important research direction. When multiple robots perform tasks in the same area, resource conflicts or task duplication may occur. How to rationally allocate resources based on factors such as task priority, remaining robot energy, and computing power to achieve collaborative operation among multiple robots remains a current research hotspot.

[0005] Currently, most traditional resource scheduling methods are based on static task allocation strategies, neglecting the impact of dynamic environmental changes on resource scheduling. Underwater operating environments are complex and variable; factors such as ocean currents, seabed obstacles, and communication signal interference can all affect robot task execution. Each robot has fixed computing resources (such as the number of CPU cores, GPU processing units, and memory capacity) and limited battery power. The system needs to schedule a series of dynamically arriving tasks, which may have dependencies, and each task can only be scheduled and executed once. Therefore, scheduling algorithms need to possess good adaptability and dynamic adjustment capabilities to cope with the task scheduling requirements in complex environments.

[0006] The application provides a multi-objective optimization-based underwater robot task allocation method, which combines a non-dominated solution set screening technology of an entropy weight method and is based on an improved NSGA-III algorithm to solve underwater robot resource scheduling, so as to realize comprehensive optimization of task completion time, total energy consumption and load balancing. SUMMARY

[0007] The application aims to provide a multi-objective dynamic resource scheduling method for underwater robots, so as to solve the problems of the conventional resource scheduling method based on a static task allocation strategy, lack of adaptability and dynamic adjustment capability and inability to cope with task scheduling requirements in a complex environment and the like in the background art.

[0008] To achieve the above object, the application adopts the following technical scheme:

[0009] The application provides a multi-objective dynamic resource scheduling method for underwater robots, comprising the following steps:

[0010] A resource scheduling problem model for underwater robots is constructed; a target optimization model is obtained based on the resource scheduling problem model for underwater robots, and the target optimization model comprises minimization of maximum execution time of a task, minimization of total power consumption and load balancing;

[0011] Reference points are generated and distributed in a target space; time, energy consumption and load balancing are taken as three targets, and the reference points are ensured to be evenly covered in each region of the target space;

[0012] A population is coded and initialized; a multi-segment chromosome coding is adopted to represent a scheduling scheme, and the scheduling scheme comprises a task allocation sequence and a task allocated to a robot; and an initial parent population is generated in a random manner;

[0013] A non-dominated solution set is obtained through evolution iteration; based on the improved NSGA-III algorithm, the non-dominated solution set is obtained through evolution iteration of reference point niche selection and combination of crossover and mutation operations;

[0014] An optimal scheduling scheme is obtained; the non-dominated solutions are objectively weighted by using an entropy weight method, and an optimal compromise solution is selected from the non-dominated solution set as the optimal scheduling scheme.

[0015] Preferably, the resource scheduling problem model for underwater robots is constructed as follows:

[0016] The resource scheduling problem model for underwater robots comprises task scheduling description, construction of a target function and modeling of a constraint condition, and a target optimization model is obtained;

[0017] Let a task set be Let a robot set be For each task The following definitions are given: d i,ris the predicted execution duration of task i on robot r; x i,r denotes whether task i is executed by robot r; e i,r is the consumed electricity of task i on robot r; e 0,r is the minimum required consumed electricity of robot r;

[0018] The objective model is constructed as follows:

[0019] f1 = min(max(d i,r )

[0020]

[0021] wherein f1, f2, f3 represent minimizing the maximum execution time of tasks, minimizing the total electricity consumption, and load balancing, respectively, and the load balancing is the variance of CPU, GPU, and memory usage rates in the machine cluster .

[0022] Further, the consumed electricity of task i on robot r is specifically as follows:

[0023]

[0024] wherein α r (t), β r (t), σ r (t) are the CPU, GPU, and memory energy consumption functions about time on robot r; c i is the number of CPUs occupied by task i during execution; g i is the number of GPUs occupied by task i during execution; m i is the memory capacity occupied by task i during execution;

[0025] Further, the calculation of the load balancing is specifically as follows:

[0026]

[0027]

[0028] wherein w j is the variance weight of CPU, GPU, and memory usage rates in the machine cluster, satisfying

[0029] Preferably, the reference points are generated and distributed in the target space, specifically as follows:

[0030] The reference points are distributed on an M-1 dimensional hyperplane of the target space, where M is the dimension of the target space, and in a target space with M = 3, the distribution of the reference points needs to satisfy a weight normalization constraint, which defines a closed two-dimensional hyperplane in the three-dimensional space, and all the reference points need to be located in this plane;

[0031]

[0032] wherein λ1, λ2, λ3 respectively correspond to the weight coefficients of the three objectives of time, energy consumption, and load balancing;

[0033] The number of reference points is determined by a parameter H, which is the equal division number of each objective, and the calculation formula is:

[0034]

[0035] All the reference points form a uniform grid on the two-dimensional hyperplane, and the uniformity is achieved by traversing non-negative integer combinations (n1, n2, n3):

[0036]

[0037] wherein n1+n2+n3=H, and each combination corresponds to the coordinates of a reference point.

[0038] Preferably, the encoding and initialization of the population are as follows:

[0039] The chromosome is represented as two lists: task priority encoding and robot assignment encoding; the task priority encoding uses a natural number sequence [o1, o2,..., o n ] to encode, wherein represents the execution order of the tasks; the robot assignment encoding uses a natural number sequence [b1, b2,..., b n ] to encode, wherein represents the robot assigned to each task;

[0040] A random sequence satisfying the task topological order is generated at initialization, and an initial parent population P1 with a size of N is generated; a penalty mechanism is introduced to constrain the total energy consumption of the robots from exceeding the robot energy; an expansion and inheritance mechanism is introduced, and when a new task is added, the task size is changed to N+ΔN, ΔN is the number of new tasks, and the new task is inserted at the end of the task priority encoding and randomly assigned to a robot, and the executed tasks remain unchanged and the unexecuted tasks are optimized and adjusted;

[0041] Set the algorithm parameters: the current iteration number t, the maximum iteration number t max rate P N , the mutation probability P M , and the equal division number H of each objective.

[0042] Preferably, the evolutionary iteration obtains a non-dominated solution set, specifically as follows:

[0043] The individuals in the parent population P1 are subjected to crossover and mutation operations to generate a cross offspring population Q1; R1=P1∪Q1 is subjected to non-dominated sorting, and the solutions are divided into multiple frontiers F1, F2, …, F i , to obtain a non-dominated solution set F;

[0044] The reference point niche number ρ j is calculated, and elite selection is performed to generate the next generation of parents P2, which still has a size of N;

[0045] If the preset termination condition is not met, the crossover and mutation operations are returned to continue iteration, otherwise a suitable solution scheme is selected from the non-dominated solution set for scheduling.

[0046] Further, the crossover and mutation operations are specifically as follows:

[0047] For the mutation operation, a greedy replacement strategy is adopted in the machine selection part, all available robots are traversed for each task, and the expected processing time is calculated, and the robot with the shortest processing time is selected to replace the current allocation; a three-state mutation operator is designed in the task sorting part, one of the three operations of exchange, insertion, and reverse order is randomly executed with uniform probability, and the population diversity is maintained through diversified mutation;

[0048] For the crossover operation, a multi-point crossover strategy is adopted in the machine selection part, multiple crossover points are randomly set and the robot allocation fragments of the parents are alternately inherited; an improved order reservation crossover is adopted in the task sorting part, the excellent task order fragments in the parents are preferentially identified and reserved, and the remaining tasks are supplemented by a greedy replacement strategy to inherit the efficient scheduling mode.

[0049] Further, the calculation process of the reference point niche number ρ j first normalizes the objective function, then matches the individual with the reference point through the correlation operation, and finally counts the niche number; specifically as follows:

[0050] First, the objective function is normalized:

[0051] The minimum value of each objective function is taken as the ideal point , wherein M is the target dimension, M=3, corresponding to time, energy consumption, and load balancing;

[0052] For each target m, a direction vector ω m =(ω m1 ,ω m2 ,…,ω mM ) is defined, wherein ω mm =1, ω mk =10 -6(k≠m); find extreme points by minimizing adaptive scalar function ASF Extreme points set is

[0053] Map objective function value to [0,1] interval:

[0054] Second, associate individual with reference point:

[0055] Reference point λ j = (λ j1 , λ j2 , λ j3 ) is distributed on M-1 = 2 dimensional hyperplane, satisfying 1, total number

[0056] For each individual x, calculate its normalized objective vector to the vertical distance of the corresponding reference line of each reference point: Where · is the vector dot product, ||·|| is the Euclidean distance;

[0057] Associate individual x to the reference point λ j with the smallest distance, i.e. j (x) = argmin j d(x, λ j );

[0058] For each reference point λ j , count the total number of individuals associated with it, i.e. niche number ρ j = count{x∈R1|λ j (x) = λ j}.

[0059] Preferably, the elite selection generates the next generation of parents P2, as follows:

[0060] First, initialize P2 as an empty set, and sequentially include individuals in the order of frontiers (F1, F2,...); if F k is included, |P2| + |F k | ≤ N, then F k is included in its entirety; if it exceeds, only select some individuals from F k ;

[0061] Second, let F τ be the first frontier that causes |P2| + |F τ | > N, and N-|P2| individuals need to be selected from F τ : for all individuals in F τ , group them according to the associated reference points, and count the current niche number ρ j; preferentially select the individual associated with the smallest number of niches of the reference point: if the reference point λ j ρ j is the smallest, select the individual associated with λ τ and the smallest distance d(x, λ j ) from F j into P2, and update ρ j = ρ j + 1; repeat the above process until |P2| = N.

[0062] Preferably, the optimal compromise solution is screened from the non-dominated solution set, specifically as follows:

[0063] First, construct a decision matrix;

[0064] Let the non-dominated solution set be F = {X1, X2, …, X p}, and for three minimum type objectives, normalize the original target value to the [0, 1] interval; define the dimensionless value of solution X i under target f j :

[0065]

[0066] where a i,j is the original value of solution X i under target \(f j \); the max and min operations ensure that the dimensions are uniform and b i,j is larger, the target performance is better; combine the dimensionless results into a p × 3 dimensional matrix B = (b i,j ) p×3 , where the rows correspond to the non-dominated solutions and the columns correspond to the optimization targets;

[0067] Second, calculate the target weight using the entropy weight method:

[0068] For the jth target, the proportion of the dimensionless value of solution X i under this target is: If b i,j ≡ 0, add a small perturbation 10 -6 to avoid the singularity of the logarithmic operation;

[0069] Calculate the target entropy value

[0070] Calculate the difference coefficient d j = 1 – E j ; normalize the weight

[0071] For each solution X i in the non-dominated solution set, the comprehensive satisfaction is the weighted sum of the dimensionless value and the corresponding weight:

[0072]

[0073] Traverse the non-dominated solution set F, screen the solution with the maximum satisfaction degree, Corresponding X k That is the optimal compromise solution.

[0074] Compared with the prior art, the present application has the beneficial effects that:

[0075] (1) The method in the present application generates a Pareto front solution set by constructing a multi-objective optimization model, and objectively weights and evaluates the satisfaction degree of non-dominated solutions based on an entropy weight method, which is suitable for dynamic task scheduling requirements in scenarios such as ocean exploration, environmental monitoring, underwater operation, etc., and solves the problem of decision bias caused by relying on expert experience or single objective optimization in traditional methods, providing a data-driven intelligent scheduling scheme for underwater robot collaborative operation in complex marine environments.

[0076] (2) The method in the present application adopts an improved NSGA-III algorithm, which optimizes the maximization of resource utilization and minimization of energy consumption in underwater robot scheduling problems through the combination of reference point niche selection and crossover and mutation operations. BRIEF DESCRIPTION OF DRAWINGS

[0077] Figure 1 The flowchart of the underwater robot multi-objective dynamic resource scheduling method in the present application. DETAILED DESCRIPTION

[0078] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0079] Embodiment 1:

[0080] The underwater robot multi-objective dynamic resource scheduling method mainly includes the following steps:

[0081] Step 1, construct an underwater robot resource scheduling problem model, which includes task scheduling description, construction of objective function, modeling of constraint conditions, etc., mainly considering CPU, GPU, memory and battery capacity as resource constraints.

[0082] Consider a task scheduling problem consisting of multiple underwater robots. Each robot has fixed computing resources (CPU cores, GPU processing units, memory) and limited battery power. The system needs to schedule a series of dynamically arriving tasks, each of which may be added during the scheduling process. At the same time, it is desirable to balance the resource utilization of the robots as much as possible to avoid some robots being overloaded while others are idle. For this problem, the invention establishes an objective optimization model, including: minimizing the maximum execution time of tasks, minimizing total power consumption, and load balancing.

[0083] Let the task set be The robot set is For each task Define its resource requirements and attributes: c i is the number of CPUs occupied when task i is executed; g i is the number of GPUs occupied when task i is executed; m i is the memory capacity occupied when task i is executed; d i,r is the estimated execution time of task i on machine r; decision variable x i,r ∈0,1 indicates whether task i is executed by robot r; e i,r is the power consumed by task i on machine r; e 0,r is the minimum required power consumption on machine r; α r (t), β r (t), σ r (t) are the CPU, GPU, and memory energy consumption functions with respect to time on machine r; there are:

[0084]

[0085] Resource utilization calculation:

[0086]

[0087] Resource average utilization calculation:

[0088]

[0089] Resource variance calculation:

[0090]

[0091] Resource variance weighting:

[0092]

[0093] where w j is the variance weight of the three resource utilization rates in the machine cluster, satisfying

[0094] The target constraints are as follows:

[0095]

[0096] where C r is the number of available CPU cores on machine r; G r is the number of available GPU units on machine r; M r is the available memory capacity on machine r; E r is the remaining energy on machine r.

[0097] The target model is constructed as follows:

[0098] f1 = min(max(d i,r ))

[0099]

[0100]

[0101] where f1, f2, and f3 represent minimizing the maximum execution time of tasks, minimizing total power consumption, and load balancing, respectively.

[0102] Step 2: Generate reference points and distribute them in the target space: the reference points are distributed on the M-1 dimensional hyperplane in the target space, where M is the dimension of the target space.

[0103] In this embodiment, M = 3; in the target space with M = 3, the distribution of reference points needs to satisfy the weight normalization constraint:

[0104]

[0105] where λ1, λ2, and λ3 correspond to the weight coefficients of the time, energy consumption, and load balancing targets, respectively. This constraint defines a closed two-dimensional hyperplane in a three-dimensional space, and all reference points must be located within this plane to ensure that the trade-off relationship between the three targets is comparable. The number of reference points is determined by the parameter H (the equal division number of each target), and the calculation formula is:

[0106]

[0107] For example, when M = 3 and H = 5, 5 reference points can be generated, which form a uniform grid on the two-dimensional hyperplane. Uniformity is achieved by traversing non-negative integer combinations (n1, n2, n3), where n1 + n2 + n3 = H, and each combination corresponds to the coordinates of a reference point:

[0108]

[0109] Through the method, the reference points are in equilateral triangle grid distribution on the hyperplane, distances of adjacent reference points are equal, and balanced coverage of each area of the target space is ensured, and local optimization of the algorithm is avoided.

[0110] Step 3, encoding and initialization of population: for the scheduling problem, the application adopts multi-section chromosome coding to represent a scheduling scheme, including task allocation sequence and task-machine allocation part.

[0111] Specifically, the chromosome can be represented as two string lists: task priority coding and robot allocation coding; an initial parent population P1 of size N is generated in a random manner; algorithm parameters are set: current iteration number t, maximum iteration number t max Rate P N , mutation probability P M , and equalization number H of each target.

[0112] The task priority coding is encoded by a sequence, representing the execution order of the tasks, and uses a natural number sequence [o1, o2,..., o n ] to directly determine the scheduling order of the tasks on the underwater robot, and to affect the total completion time and load balancing, wherein It is ensured that each task is scheduled only once. The robot allocation coding represents the execution device (robot) allocated to each task, and is encoded by using a natural number sequence [b1, b2,..., b n ] to determine the mapping relationship between the tasks and the resources, and directly affect the resource utilization rate and load balancing, wherein It is ensured that the tasks are allocated to valid robots.

[0113] In generating the initial population, heuristics are used to ensure that the individuals meet the basic constraints. In randomly allocating the task sequence and the machines, if it is found that the total power consumption of the tasks allocated to a machine exceeds its battery capacity, the allocation of the task to other robots can be adjusted in initialization, so as to reduce the number of obviously infeasible solutions. For the dependency relationship, a random sequence that meets the task topological sequence is generated in initialization, so that most of the initial population is a feasible solution. For the constraints that cannot be corrected by decoding (such as battery capacity overrun), the application introduces a penalty mechanism. If it is found that the total power consumption of a robot exceeds E_r after decoding of an individual, it is considered to violate the constraint. A large penalty value can be added to the objective function of the individual, so that it is inferior to any feasible solution in selection.

[0114] To adapt the algorithm to the dynamic task joining, the following encoding expansion and solution inheritance mechanism is introduced: when a new task arrives, the problem size changes from N to N+ΔN, ΔN is the number of new tasks. The chromosome length is expanded, and the position of each new task is added in the priority encoding and assignment encoding. The initial processing can temporarily insert each new task at the end of the priority sequence and randomly or heuristically assign a robot. In this way, the previous solution encoding is still valid, and the scheduling order corresponding to the old task subset remains unchanged, and the new task is attached as an "unscheduled" state at the end. If the task continues to arrive dynamically, set a fixed scheduling period or threshold (such as a certain number of tasks or resource changes reach a certain degree) for rescheduling. Between two rescheduling, the previously calculated scheduling scheme can be partially executed or locked, the tasks that have started execution remain the original plan until completion, and the unstarted or newly added tasks are partially optimized and adjusted.

[0115] Step 4, crossover and mutation operation: the individuals in the parent population P1 are subjected to crossover and mutation operation to generate the cross offspring population Q1.

[0116] The genetic operation is performed on the individuals of the parent population P1 to generate offspring individuals:

[0117] For the mutation operation, in the machine selection part, a greedy replacement strategy is adopted, all available machines are traversed for each task and the predicted processing time is calculated, and the machine with the shortest processing time is selected to replace the current assignment to optimize the overall completion time; in the process sorting part, a three-state mutation operator is designed, one of the three operations of exchange, insertion and reverse order is randomly executed with uniform probability, and the population diversity is maintained through diversified mutation

[0118] For the crossover operation, in the machine selection part, a multi-point crossover strategy is adopted, multiple crossover points are randomly set and the machine assignment fragments of the parents are alternately inherited; in the process sorting part, an improved sequential preservation crossover (SPX) is adopted, the excellent process order fragments in the parents are preferentially identified and preserved, and the remaining processes are supplemented through a greedy strategy to inherit an efficient scheduling mode.

[0119] Step 5, evolution iteration: R1=P v ∪Q1 is non-dominantly sorted, and the solutions are divided into multiple frontiers F1, F2, …, F i ; the reference point niche number ρ j is calculated and elite selection is performed to generate the next generation of parents P2, the size is still N.

[0120] 5.1, non-dominant sorting:

[0121] (1) For each individual x∈R1, calculate the number of individuals dominated by x n xi.e. the number of y's that satisfy "all objective values of x are better than individual y"; the set of individuals dominated by x S x i.e. the set of all y's dominated by x;

[0122] (2) include all individuals with n x = 0 into F1, which means these individuals are not dominated by any other individual;

[0123] (3) for each individual x in F k , traverse its dominating set S x , decrease n x of each y ∈ S y by 1; when n y = 0, include y into the next front F k+1 . Repeat this process until all individuals are assigned to the corresponding front.

[0124] 5.2 Reference point niche number calculation:

[0125] Reference point niche number ρ j is used to measure the number of individuals associated with each reference point, which is a key indicator of elite selection. The calculation process needs to normalize the objective function first, then match individuals with reference points through association operation, and finally count the niche number.

[0126] 5.2.1 Objective function normalization:

[0127] (1) take the minimum value of each objective function to form the ideal point where M is the objective dimension, M = 3 in this invention, corresponding to time, energy consumption, and load balancing;

[0128] (2) for each objective m, define the direction vector ω m = (ω m1 , ω m2 , …, ω mM ), where ω mm = 1, ω mk = 10 -6 (k ≠ m). Find the extreme point by minimizing the adaptive scalar function (ASF) The extreme point set is

[0129] (3) map the objective function value to the [0, 1] interval to eliminate the dimension effect:

[0130] 5.2.2 Individual and reference point association:

[0131] (1) reference point λ j = (λ j1,λ j2 ,λ j3 )distributed on M-1 = 2-dimensional hyperplanes, satisfying total number (H is the target equalization number);

[0132] (2) For each individual x, calculate its normalized target vector to the perpendicular distance of the corresponding reference line of each reference point: where • is the vector dot product and ||•|| is the Euclidean distance;

[0133] (3) Associate individual x to the reference point λ j with the minimum distance, i.e. j (x) = argmin j d(x, λ j );

[0134] (4) For each reference point λ j , count the total number of associated individuals, i.e. the niche number ρ j = count{x∈R1|λ j (x) = λ j}.

[0135] 5.3 Elite selection to generate the next generation population P2:

[0136] (1) Initialize P2 as an empty set and sequentially incorporate individuals in the order of the frontiers (F1, F2, …). If incorporating F k does not exceed |P2| + |F k | ≤ N, incorporate all individuals in F k ; if it exceeds, only select some individuals from F k ;

[0137] (2) Let F τ be the first frontier that causes |P2| + |F τ | > N, and N-|P2| individuals need to be selected from F τ : for all individuals in F τ , group them according to the associated reference points and count the current niche number ρ j of each corresponding reference point; preferentially select individuals associated to the reference point with the minimum niche number: if the niche number ρ j of the reference point λ j is the minimum, select the individual in F τ associated to λ j with the minimum distance d(x, λ j ) and incorporate it into P2, and update ρ j = ρ j + 1; repeat the above process until |P2| = N.

[0138] Step 6, if the preset termination condition is not met, return to step 4 to continue iteration, otherwise schedule according to the selected solution scheme from the front solution F1.

[0139] Step 6, if the preset termination condition is not met, return to step 4 to continue iteration, otherwise schedule according to the selected solution scheme from the front solution F1.

[0140] 6.1, construct the decision matrix

[0141] Let the non-dominated solution set be F = {X1, X2, …, X p}, for three minimum objectives, normalize the original target value to the interval [0, 1]. Define the solution X i The dimensionless value under the target f j :

[0142]

[0143] Where a i,j is the original value of the solution X i in the target \(f j \); the max and min operations are for all solutions in the non-dominated solution set F, to ensure that the dimension is unified and b i,j is larger, the target performance is better; combine the dimensionless results into a p x 3 matrix B = (b i,j ) p×3 , the matrix row corresponds to the non-dominated solution, and the column corresponds to the optimization target.

[0144] 6.2, entropy weight method to calculate the target weight

[0145] For the jth target, the proportion of the dimensionless value of the solution X i under the target is: If b i,j ≡ 0, add a small perturbation 10 -6 to avoid the singularity of logarithmic operation;

[0146] Calculate the target entropy value

[0147] Calculate the difference coefficient d j = 1 - E j ; normalize the weight

[0148] For each solution X i in the non-dominated solution set, the comprehensive satisfaction is the weighted sum of the dimensionless value and the corresponding weight:

[0149]

[0150] Iterate through the non-dominated solution set F, select the solution with the maximum satisfaction, and the corresponding X k is the optimal compromise solution.

[0151] The above description is only used to help understand the method of the present application and its core idea, but the protection scope of the present application is not limited thereto. For those skilled in the art, according to the technical scheme and the inventive concept of the present application, equivalent replacement or change within the technical scope disclosed by the present application should be covered within the protection scope of the present application. In summary, the content of the present specification should not be understood as a limitation of the present application.

Claims

1. A method for multi-objective dynamic resource scheduling of an underwater robot, the method comprising: The method comprises the following steps: A model of underwater robot resource scheduling problem is constructed; a target optimization model is obtained based on the model of underwater robot resource scheduling problem, and the target optimization model comprises minimizing the maximum execution time of a task, minimizing total power consumption, and load balancing; Reference points are generated and distributed in the target space; time, energy consumption, and load balancing are taken as three targets to ensure that the reference points evenly cover each region of the target space; A population is coded and initialized; a multi-segment chromosome coding is used to represent a scheduling scheme, and the scheduling scheme comprises task allocation sequence and robot allocation tasks; and an initial parent population is generated in a random manner; Non-dominated solution sets are obtained through evolutionary iteration; the non-dominated solution sets are obtained through reference point niche selection and the combination of crossover and mutation operations based on the improved NSGA-III algorithm; An optimal scheduling scheme is obtained; the non-dominated solutions are objectively weighted by using an entropy weight method, and an optimal compromise solution is selected from the non-dominated solution set as the optimal scheduling scheme.

2. The method of claim 1, wherein, The model of underwater robot resource scheduling problem is constructed in the following manner: The model of underwater robot resource scheduling problem comprises task scheduling description, construction of a target function, and modeling of a constraint condition to obtain a target optimization model; Let the set of tasks be Let the set of robots be For each task is defined as follows: d i,r is the estimated execution time of task i on robot r; x i,r ∈0,1 indicates whether task i is executed by robot r; e i,r is the consumed energy of task i on robot r; e 0,r is the minimum required consumed energy of robot r; The target model is constructed in the following manner: f1 = min(max(d i,r )) Wherein, f1, f2, f3 respectively represent minimizing the maximum execution time of tasks, minimizing the total power consumption and load balancing, and the load balancing is the variance of the usage rates of CPU, GPU and memory in the machine cluster minimizing.

3. The method of claim 1, wherein, The reference points are generated and distributed in the target space in the following manner: The reference points are distributed on an M-1-dimensional hyperplane in the target space, where M is the dimension of the target space; in a target space with M=3, time, energy consumption, and load balancing are taken as three targets, and the distribution of the reference points needs to satisfy a weight normalization constraint; the constraint defines a closed two-dimensional hyperplane in a three-dimensional space, and all the reference points need to be located in the plane; the number of the reference points is determined by a parameter H, which is the equal division number of each target; all the reference points form a uniform grid on the two-dimensional hyperplane.

4. The method of claim 1, wherein, The population is coded and initialized in the following manner: The chromosome is represented as two lists: task priority coding and robot allocation coding; the task priority coding is coded by using a natural number sequence to represent the execution sequence of the tasks; the robot allocation coding is coded by using a natural number sequence to represent the robot allocated to each task; A random sequence satisfying the task topology sequence is generated during initialization, and an initial parent population P1 of size N is generated; a penalty mechanism is introduced to constrain the total power consumption of the robot to exceed the energy of the robot; an extension and inheritance mechanism is introduced to optimize and adjust the number of new tasks; algorithm parameters are set: the current iteration number t, the maximum iteration number t max Rate P N , mutation probability P M , and the equalization number H of each target.

5. The method of claim 1-4, wherein, The non-dominated solution sets are obtained through evolutionary iteration in the following manner: The individuals in the parent population P1 are subjected to crossover and mutation operations to generate a cross offspring population Q1; the R1=P1∪Q1 is subjected to non-dominated sorting, and the solutions are divided into a plurality of frontiers F1, F2, …, F i , to obtain a non-dominated solution set F; By calculating the reference point niche number p j And the elite selection to generate the next generation of parents P2, size is still N; If the preset termination condition is not met, the crossover and mutation operations are returned to continue iteration; otherwise, a suitable solution scheme is selected from the non-dominated solution set to perform scheduling.

6. The method of claim 5, wherein, The crossover and mutation operations are performed in the following manner: For the mutation operation, a greedy replacement strategy is adopted in the machine selection part; for each task, all available robots are traversed and the predicted processing time is calculated, and the robot with the shortest processing time is selected to replace the currently allocated robot; a three-state mutation operator is designed in the task sorting part to maintain the diversity of the population; For the crossover operation, a multi-point crossover strategy is adopted in the machine selection part; a plurality of crossover points are randomly set and the robot allocation fragments of the parents are alternately inherited; An improved order reservation crossover is adopted in the task sorting part; the excellent task order fragments in the parents are preferentially identified and reserved, and the remaining tasks are supplemented by using a greedy replacement strategy.

7. The method of claim 5, wherein, The reference niche number p j The calculation process is as follows: first, normalize the objective function; then, match the individual with the reference point through the correlation operation; and finally, count the niche number. First, the target function is normalized: The minimum value of each objective function constitutes an ideal point wherein M is the target dimension, M=3, corresponding to time, energy consumption, load balancing; For each target m, define a direction vector ω m = (ω m1 , ω m2 ,..., ω mM ), where ω mm = 1, ω mk = 10 -6 (k ≠ m); find the extremal point by minimizing the adaptive scalar function ASF The set of extremal points is Mapping the objective function value to the interval [0, 1]: Second, the individual is associated with the reference point: Reference point λ j = (λ j1 , λ j2 , λ j3 ) is distributed on an M-1 = 2-dimensional hyperplane, satisfying Total number For each individual x, compute its normalized target vector The perpendicular distance from each reference point to the corresponding reference line: where • is the vector dot product and ||•|| is the Euclidean distance. associating the individual x to the reference point λ of minimum distance j i.e. λ j (x) = argmin j d(x, λ j ); For each reference point λ j , count the total number of individuals associated with it, i.e. the number of niches ρ j = count{x∈R1|λ j (x) = λ j}.

8. The method of claim 7, wherein, The elite selection generates the next generation of parents P2 in the following manner: First, initialize P2 as an empty set, and then include individuals in the frontier order (F1, F2, ...); if F1 is included... k After |P2|+|F k If |≤N, then F k Include all; if more than that, only include from F. k Select a subset of individuals; Secondly, set F τ as the first leading edge of |P2|+|F τ , select N-|P2| individuals from F τ : for all individuals in F τ , group them according to the associated reference points, and count the current niche number ρ j of each group corresponding to the reference points; preferentially select the individual associated to the reference point with the minimum niche number: if the ρ j of the reference point λ j is the minimum, select the individual associated to λ τ and the minimum distance d(x, λ j ) from F j into P2, and update ρ j = ρ j +1; repeat the above process until |P2| = N.

9. The method of claim 5, wherein, The optimal compromise solution is screened from the non-dominated solution set, and the specific process is as follows: Firstly, the decision matrix is constructed; Let the non-dominated solution set be F = {X1, X2, …, X p}, the original target value is normalized to the interval [0, 1] for 3 small targets; define the solution X i The dimensionless value under the target f j ​ wherein a i,j to solve X i the original value of the target \(f j \) Secondly, the target weight is calculated by the entropy weight method: For the jthtarget, solve X i The proportion of the dimensionless value of X under the target is: Computing target entropy value Computing the difference coefficient d j = 1 - E j ; normalized weight For each solution X in the non-dominated solution set i whose overall satisfaction is the weighted sum of the dimensionless values and the corresponding weights: The non-dominated solution set F is traversed, and the solution with the maximum satisfaction is screened, Corresponding X k That is the optimal compromise solution.

Images