Vector building element capping conflict processing method based on deep reinforcement learning

Abstract

The invention belongs to the technical field of data processing, and provides a vector building element capping conflict processing method based on deep reinforcement learning, which comprises the following steps: determining and constructing a constraint Delaunary triangulation network and a skeleton line segment between vector building elements in a target space region, and generating a space conflict detection field model; connecting each vector building element with center points of other surrounding vector building elements, and determining abnormal vector building elements with gland conflicts in combination with a space conflict detection field model; determining the current state information of the abnormal vector building elements, performing data grouping according to the area scale of the abnormal vector building elements so as to set constraint condition parameters and action reward information of each group, and obtaining action prediction information in combination with a DQN model; and according to the action prediction information, controlling the abnormal vector building elements to execute the position movement action until the optimal position. According to the scheme provided by the invention, the processing efficiency, precision and feasibility of the capping conflict are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and particularly relates to a method for processing overlapping conflicts of vector building elements based on deep reinforcement learning. Background Art

[0002] In the fields of cartography and geographic information systems, in order to meet the display requirements of maps with different scales, a variety of cartographic generalization algorithms need to be applied, and diverse cartographic symbols are introduced. However, these operations often lead to spatial overlapping conflicts between geographic elements such as vector buildings, making the content expressed on the map chaotic.

[0003] Regarding the problem of spatial overlapping conflicts, traditional displacement algorithms have many defects in practical applications. On the one hand, traditional displacement algorithms have a large amount of data calculation and low processing efficiency, making it difficult to meet the requirements of rapid cartography of maps with large amounts of data; on the other hand, they cannot effectively solve the secondary conflict problems brought about by displacement operations, and their applicability has obvious limitations, and they cannot achieve reasonable allocation of map elements.

[0004] Therefore, the traditional processing scheme for overlapping conflicts of vector building elements has technical problems such as low processing efficiency, insufficient processing accuracy, and poor feasibility. Summary of the Invention

[0005] The present invention provides a method for processing overlapping conflicts of vector building elements based on deep reinforcement learning to solve the defects of the traditional processing scheme for overlapping conflicts of vector building elements, such as low processing efficiency, insufficient processing accuracy, and poor feasibility.

[0006] The present invention provides a method for processing overlapping conflicts of vector building elements based on deep reinforcement learning, and the method includes: Determine and construct a constrained Delaunary triangulation and a skeleton line segment between each vector building element within the target spatial region, and generate a spatial conflict detection field model; Connect the center points of each vector building element with the center points of other surrounding vector building elements to obtain a connection line segment, and combine the spatial conflict detection field model to determine the abnormal vector building elements with overlapping conflicts; Determine the current state information of the abnormal vector building elements, perform data grouping according to the area scale of the abnormal vector building elements to set the constraint condition parameters and action reward information for each group, and combine the pre-constructed DQN model to obtain action prediction information; According to the action prediction information, control the abnormal vector building elements to perform position movement actions until the optimal position.

[0007] According to the method for handling overlapping conflicts of vector building elements based on deep reinforcement learning provided by the present invention, determining the constrained Delaunay triangulation among various vector building elements within the target spatial region includes: performing node encryption on all existing edges within the target spatial region, and constructing a planar Delaunay triangulation among various vector building elements to obtain the constrained Delaunay triangulation.

[0008] According to the method for handling overlapping conflicts of vector building elements based on deep reinforcement learning provided by the present invention, determining the skeleton line segments among various vector building elements within the target spatial region includes: determining the number of adjacent triangles for each triangle in the constrained Delaunay triangulation, and dividing all the triangles in the constrained Delaunay triangulation into three types according to the number of adjacent triangles; traversing all the triangles, and determining the segmentation curves of the blank regions among various vector building elements according to the type corresponding to each triangle, and taking the segmentation curves as the skeleton line segments within the target spatial region.

[0009] According to the method for handling overlapping conflicts of vector building elements based on deep reinforcement learning provided by the present invention, determining the abnormal vector building elements with overlapping conflicts includes: judging whether the connecting line segment intersects with the skeleton line segments in the spatial conflict detection field model, or completely intersects with any triangle within the constrained Delaunay triangulation in the spatial conflict detection field model, to obtain a first judgment result; and determining the abnormal vector building elements with overlapping conflicts according to the first judgment result.

[0010] According to the method for handling overlapping conflicts of vector building elements based on deep reinforcement learning provided by the present invention, determining the abnormal vector building elements with overlapping conflicts according to the first judgment result includes: if the first judgment result is not completely yes, then determining that there are overlapping conflicts among the vector building elements, and obtaining the abnormal vector building elements with overlapping conflicts.

[0011] According to the method for handling overlapping conflicts of vector building elements based on deep reinforcement learning provided by the present invention, the DQN model includes: an online network, which is used to predict the expected return values corresponding to each alternative action through the original fixed parameters, and output action prediction information according to the expected return values; and a target network, which is used to generate target return values through the current fixed parameters, and the parameters of the target network are copied from the online network at a set period.

[0012] According to the method for handling the capping conflict of vector building elements based on deep reinforcement learning provided by the present invention, constraint condition parameters for each group are set, including: determining the map boundary constraint condition according to the map boundary corresponding to the target spatial region; determining the moving buffer constraint condition according to the distance between the abnormal vector building element and the surrounding building groups and the distribution pattern of the surrounding building groups; determining the moving distance constraint condition according to the standard distance range between the abnormal vector building element and other buildings as specified by the cartographic specification; determining the secondary conflict constraint condition according to the state of capping conflict generated between the abnormal vector building element and other vector building elements during the moving process; and taking the map boundary constraint condition, the moving buffer constraint condition, the moving distance constraint condition, and the secondary conflict constraint condition as the constraint condition parameters for each group.

[0013] According to the method for handling the capping conflict of vector building elements based on deep reinforcement learning provided by the present invention, the moving buffer constraint condition is determined according to the distance between the abnormal vector building element and the surrounding building groups and the distribution pattern of the surrounding building groups, including: if the distribution pattern of the surrounding building groups is a regular pattern, determining the minimum distance between the abnormal vector building element and the adjacent building elements in the surrounding building groups, and constructing a first moving buffer with a side length three times the minimum distance and capable of accommodating twice the size of the abnormal vector building element itself, to obtain the moving buffer constraint condition under regular distribution; if the distribution pattern of the surrounding building groups is an irregular pattern, determining the maximum distance between the abnormal vector building element and the adjacent building elements in the surrounding building groups, and constructing a second moving buffer with a side length four times the maximum distance and capable of accommodating twice the size of the abnormal vector building element itself, to obtain the moving buffer constraint condition under irregular distribution.

[0014] According to the method for handling the capping conflict of vector building elements based on deep reinforcement learning provided by the present invention, the action reward information of each group is set, including: judging whether the abnormal vector building element satisfies the map boundary constraint condition during the movement process to obtain a second judgment result, and calculating the boundary constraint reward value in the current state according to the second judgment result; judging whether the abnormal vector building element satisfies the movement buffer constraint condition during the movement process to obtain a third judgment result, and calculating the buffer constraint reward value in the current state according to the third judgment result; judging whether the abnormal vector building element satisfies the movement distance constraint condition during the movement process to obtain a fourth judgment result, and calculating the distance constraint reward value in the current state according to the fourth judgment result; judging whether the abnormal vector building element satisfies the secondary conflict constraint condition during the movement process to obtain a fifth judgment result, and calculating the secondary constraint reward value in the current state according to the fifth judgment result; summing up the boundary constraint reward value, the buffer constraint reward value, the distance constraint reward value and the secondary constraint reward value to obtain the action reward information of each group.

[0015] According to the method for handling the capping conflict of vector building elements based on deep reinforcement learning provided by the present invention, based on the action prediction information, controlling the abnormal vector building element to perform a position movement action until the optimal position, including: determining the target action of the abnormal vector building element according to the action prediction information; wherein, the target action is one of a variety of alternative movement schemes obtained by uniform division in advance; controlling the abnormal vector building element to move equidistantly until the optimal position according to the target action.

[0016] The method for handling the capping conflict of vector building elements based on deep reinforcement learning provided by the present invention determines and constructs a constrained Delaunary triangulation and a skeleton line segment among various vector building elements within a target spatial region, and generates a spatial conflict detection field model; connects the center points of each vector building element with the center points of other surrounding vector building elements to obtain a connecting line segment, and combines the spatial conflict detection field model to detect the capping conflict of the vector building elements, and determines the abnormal vector building elements with capping conflicts; determines the current state information of the abnormal vector building elements, and performs data grouping according to their area scale to set the constraint condition parameters and action reward information for each group, and combines the pre-constructed DQN model to obtain action prediction information; according to the action prediction information, controls the abnormal vector building elements to execute position movement actions until the optimal position. Since in the capping conflict handling link, the abnormal vector building elements are treated as agents, and through the DQN model, the abnormal vector building elements can efficiently solve the capping conflict through autonomous interaction with the environment, which not only improves the processing efficiency and accuracy of the capping conflict, but also makes the capping conflict handling process easier to implement and improves the feasibility of the capping conflict handling link. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0018] Figure 1 is a schematic flowchart of the method for handling the capping conflict of vector building elements based on deep reinforcement learning provided by the embodiments of the present invention; Figure 2 is a schematic structural diagram of the constrained Delaunary triangulation and the skeleton line segment in different situations; Figure 3 is a schematic diagram of the capping conflict demonstration, conflict detection representation, and conflict resolution; Figure 4 is a schematic diagram of the implementation principle of the entire capping conflict handling link; Figure 5 is a schematic diagram of the moving direction of the agent; Figure 6 is a schematic diagram of four buffer zones obtained according to four neighboring building elements around the agent; Figure 7 is a schematic diagram of the determination principle of the first moving buffer zone in the regular mode; Figure 8Schematic diagram of the principle for determining the second moving buffer in the non - regular mode; Figure 9 Schematic diagram of the identification status of the standard distance range; Figure 10 Schematic diagram of the principle for determining the maximum value of the distance before and after the movement of the agent; Figure 11 Schematic diagram of the sub - regions obtained by dividing a certain target space area; Figure 12 Schematic diagram of the detection result of the capping conflict of three sub - regions; Figure 13 Schematic diagram of the result of screening the detected capping conflicts in each sub - region and labeling them with numbers; Figure 14 Schematic diagram of the result of displaying the two agents with the largest area scale among the six groups in the entire study area; Figure 15 Schematic diagram of the result of DQN calculation for the two agents with the largest area scale among the six groups in the entire study area; Figure 16 Schematic diagram of the change situation of the reward function obtained by statistically analyzing the reward values obtained in the DQN model calculation process; Figure 17 Schematic diagram of the change situation of the loss function obtained by statistically analyzing the loss values obtained in the DQN model calculation process; Figure 18 Schematic diagram of the result of extracting the surface elements in the study area using the arcgis10.8 software platform and performing topological analysis on the existing spatial capping conflicts; Figure 19 It is Figure 18 Schematic diagram of the selected area in Figure 20 It is for Figure 18 Schematic diagram of the result of dealing with the spatial capping conflict for the selected part in Figure 21 Schematic diagram of the existence of various capping conflicts between individual buildings and the elimination of the corresponding capping conflicts. Detailed implementation manners

[0019] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below with reference to the accompanying drawings in the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present invention without making creative efforts shall fall within the protection scope of the present invention.

[0020] The following combines with Figures 1 to 21Describe the detailed solution of the vector building element capping conflict handling method based on deep reinforcement learning provided by the embodiments of the present invention.

[0021] Figure 1 It is a schematic flowchart of the vector building element capping conflict handling method based on deep reinforcement learning provided by the embodiments of the present invention.

[0022] As Figure 1 shown, the vector building element capping conflict handling method based on deep reinforcement learning provided by the embodiments of the present invention mainly includes the following steps: Step 101: Determine and construct the constrained Delaunary triangulation and skeleton line segments among each vector building element in the target space region, and generate a spatial conflict detection field model.

[0023] It can be understood that the construction of the spatial conflict detection field model provides an effective processing basis for the detection of spatial capping conflicts.

[0024] Step 102: Connect the center points of each vector building element with the center points of other surrounding vector building elements to obtain the connecting line segments, and combine with the spatial conflict detection field model to determine the abnormal vector building elements with capping conflicts.

[0025] In this embodiment, the connecting line segments can be geometrically compared with the triangles and skeleton line segments in the spatial conflict detection field model, so as to realize the detection of capping conflicts.

[0026] Step 103: Determine the current state information of the abnormal vector building elements, group the data according to the area scale of the abnormal vector building elements to set the constraint condition parameters and action reward information for each group, and combine with the pre-constructed DQN model to obtain the action prediction information.

[0027] In this embodiment, the data can be grouped according to the area scale of the abnormal vector building elements, and then the relevant constraint condition parameters and action reward information can be set.

[0028] Step 104: According to the action prediction information, control the abnormal vector building elements to execute the position movement action until the optimal position.

[0029] In one embodiment, determining the constrained Delaunary triangulation among each vector building element in the target space region specifically includes: First, perform node encryption processing on the edges of each vector building element in the target space region, and then construct a planar Delaunary triangulation among each vector building element to obtain the constrained Delaunary triangulation.

[0030] It is understandable that the basis of the spatial conflict detection field model is the constrained Delaunay triangulation in the field of computational geometry. The establishment of the constrained Delaunay triangulation includes two aspects. On the one hand, the circumcircle of each triangle does not contain any other points within the face. On the other hand, for the diagonal of the convex quadrilateral formed by every two adjacent triangles, after mutual exchange, the minimum angle of the six interior angles does not increase.

[0031] Because in practice, too few points will lead to the situation of long and narrow triangles in the constructed Delaunay triangulation. For example Figure 2 the situation shown in sub-diagram a in

[0032] To avoid the above defects, in this embodiment, it is necessary to encrypt the nodes on the edges of the vector building elements before constructing the constrained Delaunary triangulation. Specifically, reference can be made to Figure 2 sub-diagram b in Figure 2 This rule avoids the generation of long and narrow triangles in the generation of triangles. The constrained Delaunary triangulation generated after node encryption can be seen in sub-diagram c in

[0033] In one embodiment, determining the skeleton line segments between each vector building element within the target spatial region specifically includes: First, determine the number of adjacent triangles for each triangle in the constrained Delaunary triangulation. According to the number of adjacent triangles, all the triangles in the constrained Delaunary triangulation are divided into three types, namely, a triangle adjacent to one triangle, adjacent to two triangles, and adjacent to three triangles, etc. Different types of triangles construct different skeleton line segments.

[0034] Then, traverse all the triangles. According to the type corresponding to each triangle, determine the segmentation curve of the blank area between each vector building element, and use the segmentation curve as the skeleton line segment of the target space.

[0035] In one embodiment, determining the abnormal vector building elements with capping conflicts specifically includes: First, determine whether the connecting line segment intersects with the skeleton line segment in the spatial conflict detection field model, or whether it completely intersects with any triangle within the constrained Delaunary triangulation in the spatial conflict detection field model, to obtain the first judgment result.

[0036] Finally, based on the first judgment result, determine the abnormal vector building elements with capping conflicts.

[0037] In a specific implementation, based on the first judgment result, determining the abnormal vector building elements with capping conflicts specifically includes: If the first judgment result is not completely yes, it is determined that there is an overlapping conflict between the vector building elements, and the abnormal vector building elements with overlapping conflicts are obtained.

[0038] It can be understood that for the overlapping conflicts existing in space, for example Figure 3 in the situation shown in sub-diagram e, for the extraction of conflicting line segments, if the connecting line segment formed between the two center points does not intersect with the skeleton line segment or does not completely intersect with the triangle, it can be determined that there is an overlapping conflict between the abnormal vector building elements corresponding to the two center points. The detection process can refer to Figure 3 sub-diagram f in. The situation after resolving the conflict is as shown in Figure 3 sub-diagram g in.

[0039] In one embodiment, the DQN model specifically includes: An online network, which is used to predict the expected return values corresponding to each alternative action through the original fixed parameters, and output action prediction information according to the expected return values.

[0040] A target network, which is used to generate a target return value through the current fixed parameters to reduce the correlation between the predicted value and the target value during training. The parameters of the target network are copied from the online network according to a set period (such as several rounds of iteration) to maintain the relative stability of the target value and achieve the purpose of a stable training process.

[0041] In this embodiment, the DQN model can calculate the corresponding reward value for the actions performed by the intelligent agent with the help of a double neural network structure, store the experience obtained in each step of the action, and finally update and fit the parameters of the double network through experience replay and the calculation of the loss function, so as to make the best judgment and guidance on the actions of the intelligent agent, thereby effectively solving the spatial overlapping conflict.

[0042] The implementation principle of the entire overlapping conflict handling link can be referred to Figure 4 , such as Figure 4 shown. In the early stage, a field model is constructed by using a constrained Delaunary triangulation network and skeleton line segments to obtain abnormal vector building elements. Then, by determining various constraint conditions such as map boundary constraints, moving buffer constraints, moving distance constraints, and secondary conflict constraints, as well as the corresponding reward function, it is used to guide the movement and decision-making of the intelligent agent.

[0043] Based on the fact that the intelligent agent is in the initial state S t at time t, the action a t is executed. After the execution, it is necessary to judge whether the above-mentioned various constraint conditions are satisfied. If the above-mentioned various constraint conditions are satisfied, the next step action is stopped; otherwise, the action is continued to be selected. After that, the total reward value r t and the state information S t+1 at the next moment are calculated, and the experience segment (St , a t , r t , S t+1 are stored in the experience pool for subsequent experience replay. During experience replay, experience segments can be randomly sampled from the experience pool and provided to the current network and the target network for update and fitting.

[0044] Subsequently, through the set loss function the difference between the expected return value and the target return value is measured, and the parameters of the online network are updated using the gradient descent method to reduce the loss and continuously optimize the prediction ability of the current network.

[0045] In one embodiment, according to the action prediction information, the abnormal vector building element is controlled to perform a position movement action until the optimal position, specifically including: First, according to the action prediction information, the target action of the abnormal vector building element is determined; among them, the target action is one of the multiple alternative movement schemes uniformly divided in advance.

[0046] Then, according to the target action, the abnormal vector building element is controlled to move equidistantly until the optimal position.

[0047] In this embodiment, to ensure that no deformation occurs during the movement, referring to Figure 5 , in this embodiment, the entire building is used as the agent S, and its action is a deterministic action. In this embodiment, the action of the agent S is defined as an equidistant action in eight uniform directions, specifically as shown by the arrows in Figure 5 , and the distance size is determined according to the cartographic specification.

[0048] At the same time, with the help of the greedy strategy in the DQN model to balance the exploration and exploitation ratio in the action, when the agent takes an action each time, with the probability of it conducts random exploration, and with the probability of it makes maximum use of the result predicted by the neural network, that is: (1) where a t represents the action of the agent in the current state, represents randomly selecting an action from the action space A, represents the probability, is a constant between 0 and 1, represents selecting the action that maximizes the Q (i.e., return) value in the current state.

[0049] In this embodiment, the DQN model introduces a dual-network structure of an online network and a target network to predict the actions of the agent and continuously fit the actual action situation, so as to guide each action of the agent. The online network is a neural network specifically used to generate action selection probabilities during the training process of the agent, providing a numerical basis for the selection of the agent's actions.

[0050] Through the experience replay mechanism, the online network continuously extracts a certain batch of experience data in each episode to participate in the parameter update work of gradient descent, and controls the size of the update step through the learning rate The control updates the size of the step.

[0051] As a copy introduced in the DQN model, the target network is mainly used to calculate the target Q value. The specific calculation formula is: (2) In the formula, is the target Q value, is the maximum Q value predicted by the target network for the next state S t+1 , r t is the total reward value obtained for the current action, is the discount factor, used to control the weight of future rewards, represents the action at the next moment.

[0052] By means of the different Q values calculated by the dual-network, the loss function is calculated between the online network and the target network, and the minimum error is continuously achieved to gradually fit the real target Q value. And the parameters of the target network are periodically synchronized and updated from the online network to avoid the training oscillation problem caused by frequent update of the target value.

[0053] The constraint condition is a limiting rule used to determine the optimal position of the agent in the process of deep reinforcement learning. By setting various constraint conditions for the DQN model, it is judged whether the movement result of the agent reaches the optimal position. The constraint condition parameters of this embodiment specifically include: the map boundary constraint condition, the movement buffer constraint condition, the movement distance constraint condition, and the secondary conflict constraint condition.

[0054] In one embodiment, taking any group as an example, the constraint condition parameters of each group are set, specifically including: On the first hand, according to the map boundary corresponding to the target space area, the map boundary constraint condition is determined.

[0055] As the carrier of map data, when the agent moves, it is necessary to ensure that the agent is within the map range, and it is necessary to ensure that the agent does not exceed the map boundary during the movement of the agent.

[0056] In a second aspect, according to the distances between the abnormal vector building elements and the surrounding building clusters, as well as the distribution pattern of the surrounding building clusters, the mobile buffer constraint conditions are determined.

[0057] Maintaining the distribution pattern among the vector building elements is an important constraint in resolving capping conflicts. In this embodiment, a constraint is mainly imposed on the movement range of the vector building elements within the target spatial region to ensure that the distribution of the vector building elements after movement and change is within a reasonable range, and to prevent the agent from over-exploring the space.

[0058] The specific approach is to detect the distances from each neighboring building element in the surrounding building clusters around the agent to the agent, and select one of the minimum distances as the basis to create a buffer zone as the movement range of the agent, and this range can at least accommodate twice the size of the agent itself. For example, Figure 6 As shown, four buffer zones can be designed based on the four neighboring building elements around the agent, namely the four areas numbered 201, 202, 203, and 204. However, only the buffer zone numbered 201 cannot accommodate twice the size of the agent in all four directions. Therefore, the construction of the buffer zone numbered 201 is not established, and the construction of the other buffer zones is established. In this embodiment, different calculation methods are adopted for regular and irregular building cluster distribution patterns.

[0059] In a specific implementation, according to the distances between the abnormal vector building elements and the surrounding building clusters, as well as the distribution pattern of the surrounding building clusters, the mobile buffer constraint conditions are determined, specifically including: If the distribution pattern of the surrounding building clusters is a regular pattern, the constraints on the movement distance and range of the vector building elements are relatively strong. When establishing the buffer zone, the principle of minimizing the range is followed. Then, the minimum distance between the abnormal vector building element and the neighboring building elements in the surrounding building clusters is determined, and a first mobile buffer zone with a side length three times this minimum distance and capable of accommodating twice the size of the abnormal vector building element itself is constructed. Moreover, the buffer zone range in this embodiment is determined by the agent with the largest single area that has a spatial capping conflict within the target spatial region, so as to ensure that the buffer zone established by this agent can be applicable to other elements and improve the operation efficiency of the algorithm.

[0060] For example, Figure 7 As shown, the size of the buffer zone is determined by the minimum distance d between the agent and the neighboring building elements in the surrounding building clusters. A rectangular area with a side length of 2d is constructed based on this minimum distance, and on this basis, it is expanded outward by 1.5 times to construct a rectangular area with a side length of 3d and capable of accommodating twice the size of the abnormal vector building element itself as the first mobile buffer zone.

[0061] If the distribution pattern of the surrounding building complexes is an irregular pattern and the distribution position after the agent moves does not need to be overly precise, then determine the maximum distance between the abnormal vector building element and the adjacent building elements in the surrounding building complexes, and construct a second movement buffer with a side length four times the maximum distance and capable of accommodating twice the size of the abnormal vector building element itself.

[0062] As Figure 8 shown, for Figure 8 d1, d2, d3, and d4 in

[0063] take the maximum distance d1 as the basis for determining the buffer, construct a rectangular buffer with a side length twice that of d1, and at the same time expand it outward by a factor of 2 on this basis to construct a rectangular buffer with a side length four times that of d1 and capable of accommodating twice the size of the abnormal vector building element itself, that is, the second movement buffer, so as to ensure that this range is applicable to most building complexes.

[0064] In this embodiment, according to the cartographic specification, determine the standard distance range between the abnormal vector building element and other buildings, and determine the movement distance constraint conditions. Figure 9 As

[0065] shown. In this embodiment, during the movement process, further constrain the movement distance of the agent to ensure that the distance between the vector building elements after movement does not exceed the limit of the map cartography distance. At the same time, considering the timeliness and complexity of the algorithm operation, only calculate the movement distance constraint for the agent with the largest area that has an overlapping conflict, so that it can meet the movement requirements of most agents.

[0066] In this embodiment, the rule for establishing the movement distance constraint is as follows: First, calculate the maximum distance before and after the movement. For the maximum distance D, it is composed of the distance d5 from the geometric center of the agent to its farthest boundary and a certain value d6 in the standard distance range. Specifically, as Figure 10 shown, the maximum distance D before and after the agent moves can be specifically expressed as follows: (3) Fourthly, according to the state of the overlapping conflict between the abnormal vector building element and other vector building elements during the movement process, determine the secondary conflict constraint conditions.

[0067] During the process of moving the agents with spatial capping conflicts, it is inevitable to generate capping with other surrounding agents, resulting in new spatial conflicts. Therefore, in this embodiment, secondary conflict constraint conditions are constructed for the agents, and the secondary conflict is regarded as the strictest constraint.

[0068] Finally, the map boundary constraint conditions, movement buffer constraint conditions, movement distance constraint conditions, and secondary conflict constraint conditions are used as the constraint condition parameters for each group.

[0069] In one embodiment, taking any group as an example, the action reward information for each group is set, specifically including: On the first hand, it is judged whether the abnormal vector building element satisfies the map boundary constraint conditions during the movement process, and a second judgment result is obtained. According to the second judgment result, the boundary constraint reward value in the current state is calculated.

[0070] In this embodiment, if the agent exceeds the map boundary, the reward value is set to -3; if not, the reward value is set to 0. That is to say, the boundary constraint reward value can be expressed as follows: (4) On the second hand, it is judged whether the abnormal vector building element satisfies the movement buffer constraint conditions during the movement process, and a third judgment result is obtained. According to the third judgment result, the buffer constraint reward value in the current state is calculated.

[0071] Since the distribution pattern ensures the correctness and authenticity of the map data, once the movement result of the agent exceeds the range of the movement buffer, the reward value is set to -7 to guide it to continue moving to avoid exceeding the movement buffer. If it does not exceed the movement buffer, the reward value can be set to 0. Therefore, the buffer constraint reward value can be expressed as follows: (5) On the third hand, it is judged whether the abnormal vector building element satisfies the movement distance constraint conditions during the movement process, and a fourth judgment result is obtained. According to the fourth judgment result, the distance constraint reward value in the current state is calculated.

[0072] After the agent takes an action and moves, if it exceeds the specified distance, that is, it does not satisfy the movement distance constraint conditions, the reward value is set to -3. On the contrary, the reward value is set to 0. Therefore, the distance constraint reward value can be expressed as: (6) On the fourth hand, it is judged whether the abnormal vector building element satisfies the secondary conflict constraint conditions during the movement process, and a fifth judgment result is obtained. According to the fifth judgment result, the secondary constraint reward value in the current state is calculated.

[0073] In this embodiment, the absolute value of the reward value generated by the secondary production should be greater than the reward values of other constraint items, so as to guide the agent to quickly learn to avoid the occurrence of secondary conflicts. When a capping conflict occurs, the reward value is set to -10, and on the contrary, the reward value is set to 0. Therefore, the secondary constraint reward value can be specifically expressed as: (7) Finally, sum up the boundary constraint reward value, buffer constraint reward value, distance constraint reward value and secondary constraint reward value to obtain the action reward information for each group.

[0074] In this embodiment, all the reward values are superimposed. When all the constraints are met, that is, when the constraint reward value obtained by the action is 0, the correct result can be output. Specifically, the action reward information (i.e., the total reward value) can be expressed as follows: (8) Among them, represents the boundary constraint reward value, represents the buffer constraint reward value, represents the distance constraint reward value, represents the secondary constraint reward value.

[0075] Next, taking a certain target space area as an example, the implementation process of the above scheme will be described in detail. Regarding the spatial characteristics of the distribution of vector building elements in the target space area, such as Figure 11 shown in the figure. Among them, the building layouts in Area A and Area B are relatively regular and are defined as relatively regular areas; while the building distribution in Area C shows a certain degree of irregularity and is classified as a relatively irregular area. For these two types of areas with different characteristics, DQN deep reinforcement learning calculations are carried out respectively, aiming to comprehensively verify the generality and effectiveness of the DQN model in solving problems under different spatial forms.

[0076] For Area A and Area B with relatively regular building distributions and Area C with irregular distributions, in this embodiment, constraint Delaunary triangulations and skeleton line segments are respectively constructed for each sub-area to detect the spatial capping conflicts existing between the vector building elements in the three sub-areas. The capping conflict detection results are as Figure 12 shown in the figure.

[0077] In the entire study area, due to the unique building patterns of individual vector building elements, they have been in a state of mutual capping during the process of map zooming. In this embodiment, this situation is regarded as a reasonable existence and is not calculated as a capping conflict. It needs to be manually deleted. The statistical results of the number of abnormal vector building elements with specific capping conflicts are shown in Table 1.

[0078] Table 1 Capping conflict statistical results

[0079] Screen the detected capping conflicts in each sub-region and highlight them with numerical labels. The results can be seen in Figure 13 . It is not difficult to find that the detected conflict situations in the regular Areas A and B can be seen in Table 1. For the irregular Area C, after excluding reasonable capping conflicts, the number of elements with capping conflicts is 8. The total number of detected capping conflicts in the three sub-regions A, B, and C is 23.

[0080] Sort the areas of the selected abnormal vector building elements with capping conflicts among them to construct relevant parameters. For the grouping of regular and irregular regions, see Table 2 specifically.

[0081] Table 2 Area Ranking and Grouping of Capping Conflict Elements in the Study Area

[0082] Combined with the statistics of the areas of the capping conflict elements in the three sub-regions in Table 2, the relevant parameters within each group can be seen in Table 3.

[0083] Table 3 Buffer Constraints and Moving Distance Constraints for Regular and Irregular Regions

[0084] Specifically display the two agents with the largest area scales among the six groups in the entire study area to provide a basis for judging the subsequent experimental results. For details, see Figure 14 .

[0085] For the regular regions of Areas A and B and the irregular region of Area C, combined with the moving distance constraint and buffer range constraint parameters in Table 3, perform DQN deep reinforcement learning calculations. The specific parameters are shown in Table 4.

[0086] Table 4 DQN Model Parameter Settings for Regular and Irregular Regions

[0087] For the agents in the study area, in each round, the maximum number of moving steps for each agent is set to 15 times. Move each agent with a capping conflict in a fixed order, and store the reward experience obtained in each step in the experience pool until all agents reach the moving step limit or meet the constraint conditions as the action end point. Then reset to the initial layout and perform the calculation for the next round to perform the calculation and processing of spatial capping conflicts. The DQN calculation results of some agents in the study area are as Figure 15 shown. The selected agents have the same arrangement and distribution as the agents in Figure 14 . Conduct an effect comparison based on this.

[0088] By collecting the operation result data of the DQN model, the number of feasible solutions output by the model and the operation time loss are shown in Table 5.

[0089] Table 5 DQN calculation results for regular and irregular regions

[0090] Statistical analysis is respectively carried out on the reward values and loss values obtained in the calculation process of the DQN model, and the changes of its reward function and loss function are as Figure 16 and Figure 17 shown.

[0091] Through Figure 15 analysis, it can be seen that for the regular region, the previous capping conflicts have been successfully eliminated. As can be seen from Table 5, by grouping the data to maximize the adaptation of the constraint conditions, the effective solutions output by the model corresponding to all groups of data are maintained above 1900 times, and the running time of each single round always remains at a low level; for the irregular region, from Figure 15 the calculation results, it can still be seen that reinforcement learning is still feasible in dealing with the spatial capping conflicts in the irregular region. As can be seen from the analysis of Table 5, the success rates of the two groups of data in the irregular region are maintained at a high level and have high calculation efficiency.

[0092] Combined with Figure 16 and Figure 17 shown, from the changes of the reward function and the loss function, for the regular region, the reward functions of the four groups of data all achieve gradual convergence before 500 rounds, and the reward values fluctuate gently afterwards. Combining with Figure 17 the changes of the loss function in it, it can be seen that the changes of the loss functions of the four groups of data are closely related to the changes of the reward function. The loss values of the loss functions all gradually decrease and show large fluctuations in the early action process, but then the loss values of the four groups of data start to decrease and gradually tend to be stable. Although the loss values of Group 1 and Group 3 in the regular region, and Group 1 in the irregular region do not finally stabilize at a low level, according to the number of effective solutions output in Table 5 and the peak values of the convergence of the reward function, it can be seen that the model maintains a high level of effective solution output.

[0093] From Figure 17It can be seen that in the processing of these three groups of data, many actions are performed on them. Although the loss difference finally remains at a relatively high level, in combination with the reasonable output scheme quantity in Table 5, it does not affect the output of the actual reasonable scheme and has little impact on the overall calculation efficiency and effect. For the irregular area, through the analysis of the reward function image and the loss function image of the two groups of data, it can be found that the function change situations of the two groups of data are similar to those in the regular area. Similarly, before the 500th round of the experiment, the reward values of the two gradually start to converge, and then the reward values start to converge and tend to be flat. The difference is that the reward function of the first group of data in the irregular area fluctuates more frequently, and its loss function converges more slowly, which corresponds to the amount of data to be processed itself. As can be seen from Table 5, the amount of data processed in this group is the largest in the entire study area. Because the larger the amount of data, the more difficult it is for the agent to explore the correct position in the environment, the longer the time required, and the more frequent the reward value fluctuations. However, overall, the time consumption caused by this scale of data to the model does not have a substantial impact.

[0094] Overall, the DQN model shows a high success rate in solving the spatial overlay conflict. The rapid convergence of its reward function proves that the application of the DQN model in this embodiment enables the agent to quickly select the target position through interaction with the environment in a single batch of small-scale iterations, demonstrating efficient processing capabilities, reflecting the high applicability of the DQN model in solving such problems, and reflecting a relatively high level of technical application and practical value.

[0095] For the traditional processing scheme of spatial overlay conflict, in this embodiment, the arcgis10.8 software platform is used to extract the planar features in the research area and conduct topological analysis on the spatial overlay conflicts existing in the research area. The platform detects a total of 1372 errors in the research area, and the results are as Figure 18 shown.

[0096] Through Figure 19 the selected part of the irregular area of the research area shown, that is Figure 18 in, the spatial overlay conflict is processed. With the help of the built-in algorithms such as shifting and deleting in the platform, the processing results of the irregular area are as Figure 20 shown. For the overlay conflicts between multiple vector building features such as No. 3 and No. 5 in Figure 19 and the overlay conflicts between single buildings such as No. 1, No. 2, and No. 4 in Figure 19 , the processing of the overlay conflicts, as shown in sub-graphs i, j, and k in Figure 21 , also uses the shifting and deleting algorithms to manually solve the conflicts. Specifically, it can be referred to Figure 21There are three sub - figures l, m, and n. However, the use of the deletion algorithm cannot guarantee the accuracy of the original map information. Generally speaking, due to the inaccurate extraction of planar buildings by the platform, a large amount of redundant planar data is generated, as shown in Table 6.

[0097] Table 6 Comparison of Detection and Processing Quantity Results

[0098] Through experimental comparison and data analysis, it is shown that in this embodiment, reinforcement learning is used to automatically and separately process each abnormal vector building element involved in conflicts, improving the ability to adapt to complex scenarios. At the same time, the occurrence of secondary conflicts is avoided, greatly retaining the authenticity of the map while increasing the success rate of conflict resolution. In addition, in terms of optimizing the timeliness of the algorithm, in this embodiment, the DQN model is used to uniformly perform grouped conflict processing on all conflict elements within the target spatial region, avoiding the huge time consumption caused by manually solving each conflict one by one in the traditional method, and greatly improving the timeliness of conflict processing.

[0099] In summary, the method for handling vector building element capping conflicts based on deep reinforcement learning provided by the embodiment of the present invention has at least the following advantages: First, it improves the accuracy of detecting spatial capping conflicts in a large range. By constructing a constrained Delaunary triangulation and skeleton line segments, the relationship strength between adjacent planar elements is strengthened, and compared with traditional methods, the accuracy of detecting capping conflicts is improved.

[0100] Second, it improves the timeliness of resolving spatial capping conflicts. With the mode of the DQN model's autonomous interaction with the environment and the setting of constraint conditions, abnormal vector building elements involved in capping conflicts are prompted to autonomously find the optimal position to resolve conflict problems, greatly improving the working efficiency of the capping conflict handling link.

[0101] Third, the DQN model has wide applicability and diverse output results. According to the distribution characteristics of planar buildings in different research areas, differential constraint rules are formulated to scientifically determine the reasonable layout of conflict elements, enabling the DQN model to fully adapt to the characteristics of each region. Also, the model can output multiple reasonable layout schemes within the constraint framework, providing sample data for reference in the final result.

[0102] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features. These modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for handling capping conflicts of vector building elements based on deep reinforcement learning, characterized in that, Including: Determine and construct the constrained Delaunary triangulation and skeleton line segments among various vector building elements within the target spatial region to generate a spatial conflict detection field model; Connect the center points of each vector building element with the center points of other surrounding vector building elements to obtain connection line segments, and in combination with the spatial conflict detection field model, determine the abnormal vector building elements with capping conflicts; Determine the current state information of the abnormal vector building elements, group the data according to the area scale of the abnormal vector building elements to set the constraint condition parameters and action reward information for each group, and in combination with the pre-constructed DQN model, obtain action prediction information; According to the action prediction information, control the abnormal vector building elements to execute position movement actions until the optimal position.

2. The method for handling capping conflicts of vector building elements based on deep reinforcement learning according to claim 1, wherein Determine the constrained Delaunary triangulation among various vector building elements within the target spatial region, including: Perform node encryption processing on all existing edges within the target spatial region, and construct a planar Delaunary triangulation among various vector building elements to obtain the constrained Delaunary triangulation.

3. The method for handling the capping conflict of vector building elements based on deep reinforcement learning according to claim 1, characterized in that, Determine the skeleton line segments among various vector building elements within the target spatial region, including: Determine the number of adjacent triangles of each triangle in the constrained Delaunary triangulation, and divide all the triangles in the constrained Delaunary triangulation into three types according to the number of adjacent triangles; Traverse all the triangles, and according to the type corresponding to each triangle, determine the segmentation curves of the blank areas among various vector building elements, and use the segmentation curves as the skeleton line segments within the target spatial region.

4. The method for handling the capping conflict of vector building elements based on deep reinforcement learning according to claim 1, characterized in that Determine the abnormal vector building elements with capping conflicts, including: Judge whether the connection line segments intersect with the skeleton line segments in the spatial conflict detection field model, or whether they completely intersect with any triangle within the constrained Delaunary triangulation in the spatial conflict detection field model, to obtain a first judgment result; According to the first judgment result, determine the abnormal vector building elements with capping conflicts.

5. The method for handling the capping conflict of vector building elements based on deep reinforcement learning according to claim 4, wherein According to the first judgment result, determine the abnormal vector building elements with capping conflicts, including: If the first judgment result is not completely yes, it is determined that there is a capping conflict among the vector building elements, and the abnormal vector building elements with capping conflicts are obtained.

6. The method for handling the capping conflict of vector building elements based on deep reinforcement learning according to claim 1, wherein The DQN model includes: An online network, which is used to predict the expected return values corresponding to each alternative action through the original fixed parameters, and output action prediction information according to the expected return values; A target network, which is used to generate a target return value through the current fixed parameters, and the parameters of the target network are copied from the online network according to a set period.

7. The method for handling capping conflicts of vector building elements based on deep reinforcement learning according to claim 1, wherein Set the constraint condition parameters for each group, including: Determine the map boundary constraint conditions according to the map boundary corresponding to the target spatial region; Determine the movement buffer constraint conditions according to the distance between the abnormal vector building elements and the surrounding building groups and the distribution pattern of the surrounding building groups; According to the regulations of the mapping specification, determine the standard distance range between the abnormal vector building elements and other buildings, and determine the moving distance constraint conditions; According to the state of the capping conflict generated between the abnormal vector building elements and other vector building elements during the movement process, determine the secondary conflict constraint conditions; Take the map boundary constraint conditions, the moving buffer constraint conditions, the moving distance constraint conditions, and the secondary conflict constraint conditions as the constraint condition parameters for each group.

8. The method for handling vector building element capping conflicts based on deep reinforcement learning according to claim 7, wherein According to the distance between the abnormal vector building elements and the surrounding building groups and the distribution pattern of the surrounding building groups, determine the moving buffer constraint conditions, including: If the distribution pattern of the surrounding building groups is a regular pattern, determine the minimum distance between the abnormal vector building elements and the adjacent building elements in the surrounding building groups, and construct a first moving buffer with a side length three times the minimum distance and accommodating twice the size of the abnormal vector building elements themselves, to obtain the moving buffer constraint conditions under regular distribution; If the distribution pattern of the surrounding building groups is an irregular pattern, determine the maximum distance between the abnormal vector building elements and the adjacent building elements in the surrounding building groups, and construct a second moving buffer with a side length four times the maximum distance and accommodating twice the size of the abnormal vector building elements themselves, to obtain the moving buffer constraint conditions under irregular distribution.

9. The method for handling vector building element capping conflicts based on deep reinforcement learning according to claim 7, wherein, Set the action reward information for each group, including: Judge whether the abnormal vector building elements satisfy the map boundary constraint conditions during the movement process, obtain a second judgment result, and calculate the boundary constraint reward value in the current state according to the second judgment result; Judge whether the abnormal vector building elements satisfy the moving buffer constraint conditions during the movement process, obtain a third judgment result, and calculate the buffer constraint reward value in the current state according to the third judgment result; Judge whether the abnormal vector building elements satisfy the moving distance constraint conditions during the movement process, obtain a fourth judgment result, and calculate the distance constraint reward value in the current state according to the fourth judgment result; Judge whether the abnormal vector building elements satisfy the secondary conflict constraint conditions during the movement process, obtain a fifth judgment result, and calculate the secondary constraint reward value in the current state according to the fifth judgment result; Sum up the boundary constraint reward value, the buffer constraint reward value, the distance constraint reward value, and the secondary constraint reward value to obtain the action reward information for each group.

10. The method for handling vector building element capping conflicts based on deep reinforcement learning according to claim 1, wherein, According to the action prediction information, control the abnormal vector building elements to execute the position movement action until the optimal position, including: According to the action prediction information, determine the target action of the abnormal vector building elements; among them, the target action is one of the multiple alternative movement plans evenly divided in advance; According to the target action, control the abnormal vector building elements to move equidistantly until the optimal position.

Images