Multi-robot collaborative scheduling forest fire extinguishing method based on genetic programming

Abstract

The invention discloses a multi-robot collaborative scheduling forest fire extinguishing method based on genetic programming, and the method comprises the steps: obtaining task data through analog simulation, and dividing a training set and a test set; calculating phenotypic features of the individuals, and constructing an agent auxiliary model; in the evolution process, a large number of filial generations are generated through crossover and variation, then the filial generations are divided into two sub-populations, firstly, niche division is conducted on the first sub-population through phenotypic characteristics, an individual with the minimum scale is selected from each niche, then the fitness of individuals in the second sub-population is estimated through an agent model, and the individuals with the top ranking are selected; and after iterative evolution is completed, selecting an individual with the optimal fitness to test to obtain an optimal scheduling rule. Under the condition of complex constraints, multiple robots can be reasonably dispatched, the fire extinguishing and rescue time is shortened, the system performance is improved to the maximum extent, and good stability and interpretability are achieved.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a multi-robot collaborative scheduling forest fire extinguishing method based on genetic programming. Background Art

[0002] Effective firefighting and rescue efforts in the face of large-scale fires are currently receiving significant attention from fire departments both domestically and internationally. After a fire breaks out, rapidly implementing rescue and response measures to improve the survival rate of trapped personnel and minimize property losses has become a key and challenging issue in emergency management. Traditional forest firefighting methods rely primarily on manpower and limited mechanical equipment. Firefighters, armed with firefighting tools, conduct firefighting operations deep within forested areas. This approach is inefficient and poses significant risks to firefighters' lives. For example, in complex terrain and rapidly spreading forest fires, firefighters struggle to quickly reach the center of the fire and effectively control its spread. With the continuous advancement of robotics technology, the use of multi-robot collaboration for forest firefighting has become a promising solution. Multi-robot systems can achieve collaborative division of labor in complex forest environments, enabling rapid response to fires. However, current multi-robot collaborative scheduling in forest firefighting applications still faces numerous challenges. For one thing, existing multi-robot scheduling algorithms lack the ability to rapidly make decisions based on dynamic environmental factors (such as sudden changes in wind direction and varying fire spread) in complex and changing forest fire scenarios. This leads to irrational task allocation and fails to fully utilize the collaborative advantages of multi-robot systems. On the other hand, the existing algorithms have high computational complexity. Under limited computing resources and time constraints, it is difficult to generate efficient scheduling plans in a short period of time, which delays the opportunity to extinguish the fire.

[0003] Currently, forest fire fighting methods primarily include manual firefighting, aerial firefighting, and ground firefighting. Ground firefighting is direct and efficient, but due to factors such as terrain and fire intensity, a single robot struggles to complete complex firefighting tasks. Therefore, multi-robot collaborative firefighting technology has gradually become a research hotspot. This technology improves firefighting efficiency through the rational allocation of tasks and coordinated cooperation. However, existing methods, primarily based on fixed rules or heuristic algorithms, suffer from the following shortcomings: First, they lack adaptability. Fixed rules are difficult to handle complex and changing forest fire scenarios, and firefighting strategies cannot be adjusted according to actual conditions. Second, they suffer from weak global optimization capabilities. Heuristic algorithms improve firefighting efficiency to a certain extent, but are prone to falling into local optimality, making global optimization difficult to achieve.

[0004] Genetic programming is an optimization algorithm that simulates the processes of natural selection and genetic variation and is suitable for solving complex optimization problems. Compared to traditional optimization algorithms, genetic programming has the following advantages: First, it is highly adaptable, automatically adjusting its search strategy based on the characteristics of the problem and adapting to complex environments. Second, it has excellent global search capabilities. Based on population evolution, it iteratively optimizes numerous candidate solutions (i.e., different multi-robot scheduling schemes), enabling it to search within a vast solution space. Third, it is highly capable of handling complex problems, integrating complex factors and comprehensively evaluating the pros and cons of different scheduling schemes by constructing appropriate fitness functions. However, research on genetic programming for multi-robot coordinated scheduling of forest firefighting remains limited, and standard genetic programming suffers from issues such as premature loss of population diversity and premature convergence. In forest firefighting scenarios, this means that the algorithm may cease evolution before finding the true global optimal solution, resulting in suboptimal multi-robot scheduling solutions. Therefore, research is needed to improve genetic programming performance and optimize multi-robot scheduling schemes. Summary of the Invention

[0005] The purpose of this invention is to overcome the above defects and propose an improved genetic programming method, which improves the diversity and convergence of the population and realizes the optimization of multi-robot scheduling scheme.

[0006] The technical solution adopted by the present invention is: a multi-robot collaborative scheduling forest fire extinguishing method based on genetic programming, comprising the following steps:

[0007] S1, establish a task simulation to obtain task data; the task data includes the location of the fire point, the number of fire-fighting robots, the fire point spread rate, the robot fire-fighting efficiency, the distance between each fire point, the time required to detect the fire point, the initial position of the robot, and the initial distance between the robot and each fire point;

[0008] S2, processing and dividing the task data;

[0009] S3, uses the genetic programming algorithm to iteratively update the population and find the optimal scheduling rule through search iteration;

[0010] S4, the individual with the shortest completion time in the test population, evaluates its performance on the test set and obtains the scheduling rule for the task.

[0011] Furthermore, the step S2 includes the following steps:

[0012] S21, read the data in S1, divide the fire extinguishing points into known points and unknown points. The unknown points need to be monitored by the robot during the fire extinguishing process, and the other data are used as the initial data of the fire model;

[0013] S22, dividing the processed data set into a training set and a test set in a ratio of 4:1; the training set adopts a non-overlapping mini-batch strategy, using five different training data in each generation to train the optimal scheduling rule; the test set is used to evaluate the performance of the optimal scheduling rule obtained from the genetic programming search. 10. The method of multi-robot collaborative scheduling forest firefighting based on genetic programming according to claim 1 is characterized in that S3 includes the following steps:

[0014] S31, initialization and generation of GP tree individuals;

[0015] S32, evaluate the fitness of the population;

[0016] S33, calculate individual phenotypic characteristics and build proxy models;

[0017] S34, generates intermediate individuals of twice the initial population size through selection, crossover, and mutation;

[0018] S35, calculate the phenotypic characteristics of the intermediate individuals and randomly divide the intermediate individuals into two subpopulations;

[0019] S36, divide the individuals in subpopulation 1 into microhabitats based on phenotypic characteristics, and select the smallest individual in each microhabitat;

[0020] S37, estimates the fitness value in subpopulation 2 and selects the top-ranked individuals;

[0021] S38, determine whether the maximum number of iterations of the population has been reached; if so, stop the evolution; otherwise, go to S32 to continue the search process.

[0022] Furthermore, the step S31 includes the following steps:

[0023] For S311, the GP population size is set to 500, the number of iterations is set to 50, the individual generation method is “Rampedhalf-and-half”, and the tree depth range is [2,8];

[0024] S312, the function set in the GP individual is "+", "-", "*", and " / "; among them, " / " is a protected division, which returns 1 when the divisor is 0, instead of an error or infinity;

[0025] S313, the terminal set in the GP individual is some real-time key feature data of the task simulation process; when a decision needs to be made, these real-time key feature data need to be extracted, that is, the input of the GP individual, which are: robot fire extinguishing efficiency, fire point spread rate, fire point intensity, robot travel cost, candidate fire point spread rate, fire point intensity when the machine arrives at the fire point, unextinguished fire points, the number of robots assigned to candidate fire points and the total fire extinguishing efficiency of the robot, predicted fire point spread rate, the time when the robot performs fire extinguishing for the first time and the time when the robot performs fire extinguishing for the last time, the reduction in completion time when the robot group executes the candidate fire extinguishing point, the monitoring time of the candidate fire extinguishing point and the number of unassigned fire extinguishing points.

[0026] Further, the S32 includes the following steps:

[0027] All GP individuals are trained on five simulation tasks respectively. Then, the maximum completion time of each simulation task is obtained, and their average is taken. Finally, the minimum average completion time is calculated to obtain the fitness. The calculation formula is as follows:

[0028]

[0029] Among them, makespan i Indicates the maximum completion time of the model for the i-th simulation task.

[0030] Further, the S33 includes the following steps:

[0031] S331, selecting the individual with the best fitness value as the reference rule, extracting its characteristic data during the simulation process, and sampling the robot's decision-making scenario;

[0032] S332, giving the characteristic data to the GP individual whose phenotypic characteristics need to be calculated;

[0033] S333, select the task with the highest priority among the candidate tasks in each decision scenario and record it as the phenotypic feature;

[0034] S334, finally a phenotypic feature vector with 20 dimensions is obtained;

[0035] S335, the individual fitness values obtained in S32 are matched to individual phenotypic characteristics, a database is maintained, and an agent model is constructed.

[0036] Further, the S34 includes the following steps:

[0037] The selection operation selects individuals for genetic operation through a tournament selection strategy; randomly selects seven individuals from the current population as contestants; compares the fitness of the contestants and selects the individual with the best fitness as the winner; repeats the above steps until a sufficient number of winners are selected;

[0038] The crossover operation is to randomly select two points in the parent tree and exchange the subtrees according to the selected points to generate two new trees;

[0039] The mutation operation is to randomly select a point in the tree, replace the branch with the root at the point with a randomly generated subtree, and generate a new tree.

[0040] Furthermore, the S36 includes the following steps:

[0041] S361, placing individuals with the same phenotypic characteristics in population 2 in the same niche;

[0042] S362, select the smallest individual from each microhabitat.

[0043] The S37 includes the following steps:

[0044] S371, remove duplicate individuals, and remove duplicates from the individuals in population 2 according to phenotypic characteristics;

[0045] S372, supplement individuals, randomly generate individuals to replace the individuals removed in S371, until the phenotypic characteristics of individuals in population 2 are all unique;

[0046] S373, uses the surrogate model constructed in S33 to estimate the fitness of individuals in population 2;

[0047] S374, select the individual with the highest fitness ranking.

[0048] Compared with the prior art, the present invention has the following advantages:

[0049] 1. Reduce the size of the model and improve convergence performance; divide a large number of intermediate individuals into two sub-populations, and adopt different pre-selection strategies for each sub-population; the niche strategy can reduce the complexity of the GP model, reduce training time, and improve model interpretability; the agent-assisted pre-selection strategy can select promising individuals and improve convergence performance;

[0050] 2. Population diversity increases; the niche strategy of subpopulation one and the operation of removing duplicate individuals in subpopulation two both increase population diversity and prevent premature convergence of the population. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0052] The techniques described below are susceptible to various modifications and embodiments, and are described in detail herein with reference to specific embodiments in conjunction with the accompanying drawings. However, this is not intended to limit the techniques described below to the specific embodiments. It should be understood that the present invention includes all similar modifications, equivalents, and alternatives that do not depart from the spirit and technical scope of the techniques described below.

[0053] The technical solution adopted by the present invention is: a multi-robot collaborative scheduling forest fire extinguishing method based on genetic programming, comprising the following steps:

[0054] S1, obtain the task data through the simulation module. The task data that can be obtained through simulation include the location of the fire point, the number of fire-fighting robots, the fire point spread rate, the robot fire-fighting efficiency, the distance between each fire point, the time required to monitor the fire point, the initial position of the robot, and the initial distance between the robot and each fire point, etc.

[0055] S2, data processing and data division are performed through the preprocessing module. The specific steps are as follows:

[0056] S21, read the data in S1, divide the fire extinguishing points into known points and unknown points. The unknown points need to be monitored by the robot during the fire extinguishing process, and the other data are used as the initial data of the fire model;

[0057] S22, the processed data set is divided into a training set and a test set in a ratio of 4:1; the divided training set adopts a non-overlapping mini-batch strategy, using 5 different training data in each generation to train the optimal scheduling rule, and the test set is used to evaluate the performance of the optimal scheduling rule obtained from the genetic programming search.

[0058] S3, iteratively updates the population through the genetic programming module and searches for the optimal scheduling rule through iteration. The specific steps are as follows:

[0059] S31, initialize and generate GP tree individuals, the specific steps are:

[0060] S311, the GP population size is set to 500, the number of iterations is set to 50, the individual generation method is "Ramped half-and-half", and the tree depth range is [2,8]

[0061] S312, the function set in the GP individual is "+", "-", "*", and " / "; among them, " / " is a protected division, and when the divisor is 0, it returns 1 instead of an error or returning infinity.

[0062] S313, the terminal set in the GP individual is some real-time key feature data of the task simulation process; when a decision needs to be made, these real-time key feature data need to be extracted, that is, the input of the GP individual, which are: the efficiency of the robot's fire extinguishing, the rate of spread of the fire point, the intensity of the fire point, the robot's travel cost, the rate of spread of the candidate fire point, the intensity of the fire point when the machine arrives at the fire point, the unextinguished fire point, the number of robots assigned to the candidate fire point and the total fire extinguishing efficiency of the robot, the predicted spread rate of the fire point, the time when the robot performs fire extinguishing for the first time and the time when the robot performs fire extinguishing for the last time, the reduction in completion time when the robot group executes the candidate fire extinguishing point, the monitoring time of the candidate fire extinguishing point, and the number of unassigned fire extinguishing points.

[0063] S32, the fitness of the population is evaluated through the simulated fitness evaluation module. All GP individuals are trained on 5 simulation tasks respectively. Then, the maximum completion time of each simulation task is obtained, and their average is taken. Finally, the minimum average completion time is calculated to obtain the fitness. The calculation formula is as follows:

[0064]

[0065] Among them, mokespan i Indicates the maximum completion time of the model for the i-th simulation task

[0066] S33, calculate the individual's phenotypic characteristics and build a proxy model. The specific steps are:

[0067] S331, select the individual with the best fitness value as the reference rule, extract its characteristic data during the simulation process, and sample the robot's decision scenarios (each decision scenario corresponds to a candidate task for the robot), with a total of 20 decision scenarios;

[0068] S332, giving the characteristic data (input of the GP tree) to the GP individual for which the phenotypic characteristics need to be calculated;

[0069] S333, select the task with the highest candidate task priority (output of GP tree) in each decision scenario and record it as the phenotypic feature;

[0070] S334, finally a phenotypic feature vector with 20 dimensions is obtained;

[0071] S335, based on the individual fitness values obtained in S32, and then corresponding to the individual phenotypic characteristics, maintain a database and build our agent model.

[0072] S34, generate intermediate individuals of twice the initial population size through selection, crossover and mutation. The specific steps are:

[0073] S41, select operation;

[0074] Individuals are selected for genetic manipulation using a tournament selection strategy. Seven individuals are randomly selected from the current population as contestants. The fitness of the contestants is compared, and the individual with the best fitness is selected as the winner. Repeat the above steps until a sufficient number of winners are selected.

[0075] S42, crossover operation;

[0076] The crossover operation generates two new trees (offspring trees) based on two selected trees (parent trees). The crossover operation first randomly selects two points in the parent tree, and then exchanges the subtrees according to the selected points to generate two new trees.

[0077] S44, mutation operation;

[0078] Randomly select a point in the tree and then replace the branch rooted at that point with a randomly generated subtree to generate a new tree.

[0079] S35, calculate the phenotypic characteristics of the intermediate individuals and randomly divide the intermediate individuals into two subpopulations;

[0080] S36, using the niche module to divide individuals in subpopulation 1 into niches according to their phenotypic characteristics, and selecting the smallest individual in each niche. The specific steps are:

[0081] S361, placing individuals with the same phenotypic characteristics in population 2 in the same niche;

[0082] S362, select an individual with the smallest size (the least number of tree nodes) from each microhabitat.

[0083] S37, through the auxiliary agent module, estimates the fitness value of subpopulation 2 and selects the top-ranked individuals. The specific steps are as follows:

[0084] S371, remove duplicate individuals, and remove duplicates from the individuals in population 2 according to phenotypic characteristics;

[0085] S372, supplement individuals; randomly generate individuals to replace the individuals removed in S371 until the phenotypic characteristics of individuals in population 2 are all unique;

[0086] S38, determine whether the maximum number of iterations of the population has been reached; if so, stop the evolution; otherwise, go to S32 to continue the search process.

[0087] S4, selects the individual with the shortest completion time in the population through the test module, evaluates its performance on the test set, and obtains the task scheduling rules.

[0088] To further illustrate the superiority of the present invention in solving the problem of multi-robot collaborative scheduling of forest fire fighting, Table 1 shows the performance of the present invention, the standard genetic method, and the manually designed strategy on the test set.

[0089] The data for this experiment was obtained from simulations. Simulated tasks are named in the order of the number of robots, the number of fire points, the robot's efficiency, and the task's growth rate. For example, 1020SS represents 10 robots and 20 fire points. The first "S" indicates low robot firefighting efficiency, and the second "S" indicates a low task growth rate. Comparison methods include NNT, which prioritizes the closest task, MaxR, which prioritizes the task with the highest growth rate, MinD, which prioritizes the task with the lowest current demand, DD, which represents a dynamic allocation strategy, and GP, which represents a traditional genetic programming algorithm.

[0090] We conducted 20 independent experiments, with the population size set to 500 and the maximum number of iterations set to 50, and recorded the performance index values obtained each time. Table 1 shows the mean and standard deviation of the maximum completion time, with the best results highlighted in bold.

[0091] Table 1

[0092] NNT MaxR MinD DD GP The present invention 1020SS 562±66 1292±262 1295±163 542±459 321±21.8 305±9.1 1020LM 569±93 1429±231 1404±235 484±417 329±15.7 319±14.8 1520LS 475±49 1124±128 1118±118.7 231±73.7 216±18.4 208±7.8 1520LM 503±46.7 1217±154.9 1211±166.8 303±178.6 242±20.4 236±12 2020SS 486±40.0 1234±104.1 1230±112.5 299±77 226±16.8 211±7.5 2020LM 447±48.3 1126±120.5 1097±125.7 242±86.0 186±11.4 178±9.9

[0093] The comparison shows that the proposed method demonstrates superior performance in multi-robot collaborative scheduling. Specifically, the proposed method outperforms the comparison method in both the mean and standard deviation of firefighting task completion times, demonstrating that the proposed method improves performance on each firefighting task and exhibits good stability.

Claims

1. A multi-robot collaborative dispatching forest fire extinguishing method based on genetic programming, characterized in that: The following steps are involved: S1, establish a task simulation to obtain task data; the task data includes the location of the fire point, the number of fire-fighting robots, the fire point spread rate, the robot fire-fighting efficiency, the distance between each fire point, the time required to detect the fire point, the initial position of the robot, and the initial distance between the robot and each fire point; S2, processing and dividing the task data; S3, uses the genetic programming algorithm to iteratively update the population and find the optimal scheduling rule through search iteration; S4, the individual with the shortest completion time in the test population, evaluates its performance on the test set and obtains the scheduling rule for the task.

2. The method for multi-robot coordinated dispatch of forest fire extinguishing based on genetic programming according to claim 1 is characterized in that: The S2 comprises the following steps: S21, read the data in S1, divide the fire extinguishing points into known points and unknown points. The unknown points need to be monitored by the robot during the fire extinguishing process, and the other data are used as the initial data of the fire model; S22, divide the processed data set into a training set and a test set in a ratio of 4:1; the training set adopts a non-overlapping small batch strategy, using five different training data in each generation to train the optimal scheduling rule; the test set is used to evaluate the performance of the optimal scheduling rule obtained from the genetic programming search.

3. The method for multi-robot coordinated dispatching of forest fire extinguishing based on genetic programming according to claim 1 is characterized in that: The S3 includes the following steps: S31, initialization and generation of GP tree individuals; S32, evaluate the fitness of the population; S33, calculates individual phenotypic characteristics and constructs proxy models; S34, generates intermediate individuals of twice the initial population size through selection, crossover, and mutation; S35, calculate the phenotypic characteristics of the intermediate individuals and randomly divide the intermediate individuals into two subpopulations; S36, divide the individuals in subpopulation 1 into microhabitats based on phenotypic characteristics, and select the smallest individual in each microhabitat; S37, estimates the fitness value in subpopulation 2 and selects the top-ranked individuals; S38, determine whether the maximum number of iterations of the population has been reached; if so, stop the evolution; otherwise, go to S32 to continue the search process.

4. The method for multi-robot coordinated dispatching of forest fire extinguishing based on genetic programming according to claim 3 is characterized in that: The S31 includes the following steps: For S311, the GP population size is set to 500, the number of iterations is set to 50, the individual generation method is "Ramped half-and-half", and the tree depth range is [2,8]; S312, the function set in the GP individual is "+", "-", "*", and " / "; among them, " / " is a protected division, which returns 1 when the divisor is 0, instead of an error or infinity; S313, the terminal set in the GP individual is some real-time key feature data of the task simulation process; when a decision needs to be made, these real-time key feature data need to be extracted, that is, the input of the GP individual, which are: robot fire extinguishing efficiency, fire point spread rate, fire point intensity, robot travel cost, candidate fire point spread rate, fire point intensity when the machine arrives at the fire point, unextinguished fire points, the number of robots assigned to candidate fire points and the total fire extinguishing efficiency of the robot, predicted fire point spread rate, the time when the robot performs fire extinguishing for the first time and the time when the robot performs fire extinguishing for the last time, the reduction in completion time when the robot group executes the candidate fire extinguishing point, the monitoring time of the candidate fire extinguishing point and the number of unassigned fire extinguishing points.

5. The method for multi-robot coordinated dispatch of forest fire extinguishing based on genetic programming according to claim 3 is characterized in that: The S32 includes the following steps: All GP individuals are trained on five simulation tasks respectively. Then, the maximum completion time of each simulation task is obtained, and their average is taken. Finally, the minimum average completion time is calculated to obtain the fitness. The calculation formula is as follows: ； in, Indicates the model The maximum completion time of a simulation task.

6. The method for multi-robot coordinated dispatch of forest fire extinguishing based on genetic programming according to claim 3 is characterized in that: The S33 includes the following steps: S331, selecting the individual with the best fitness value as the reference rule, extracting its characteristic data during the simulation process, and sampling the robot's decision-making scenario; S332, giving the characteristic data to the GP individual whose phenotypic characteristics need to be calculated; S333, select the task with the highest priority among the candidate tasks in each decision scenario and record it as the phenotypic feature; S334, finally a phenotypic feature vector with 20 dimensions is obtained; S335, the individual fitness values obtained in S32 are matched to individual phenotypic characteristics, a database is maintained, and an agent model is constructed.

7. The method for multi-robot coordinated dispatch of forest fire extinguishing based on genetic programming according to claim 3 is characterized in that: The S34 includes the following steps: The selection operation selects individuals for genetic operation through a tournament selection strategy; randomly selects seven individuals from the current population as contestants; compares the fitness of the contestants and selects the individual with the best fitness as the winner; repeats the above steps until a sufficient number of winners are selected; The crossover operation is to randomly select two points in the parent tree and exchange the subtrees according to the selected points to generate two new trees; The mutation operation is to randomly select a point in the tree, replace the branch with the root at the point with a randomly generated subtree, and generate a new tree.

8. The method for multi-robot coordinated dispatch of forest fire extinguishing based on genetic programming according to claim 3 is characterized in that: The S36 includes the following steps: S361, placing individuals with the same phenotypic characteristics in population 2 in the same niche; S362, select the smallest individual from each microhabitat.

9. The method for multi-robot coordinated dispatching of forest fire extinguishing based on genetic programming according to claim 3 is characterized in that: The S37 includes the following steps: S371, remove duplicate individuals, and remove duplicates from the individuals in population 2 according to phenotypic characteristics; S372, supplement individuals, randomly generate individuals to replace the individuals removed in S371, until the phenotypic characteristics of individuals in population 2 are all unique; S373, uses the surrogate model constructed in S33 to estimate the fitness of individuals in population 2; S374, select the individual with the highest fitness ranking.

Images