Reinforcement learning driven distributed trajectory stream analysis elastic partitioning method

Abstract

The application discloses a distributed trajectory stream analysis elastic partitioning method based on reinforcement learning driving, which converts the elastic partitioning problem into a grid index parameter dynamic adjustment problem, and the partitioning of trajectory data is not dependent on static division rules, but realizes the elastic partitioning by dynamically adjusting the grid index parameters to adapt to the changes of the trajectory stream data in time and space.In the application, the agent continuously selects the optimal grid configuration according to the current state of the system through the reinforcement learning strategy, and balances between exploration and utilization to optimize load balancing and resource allocation.The application can dynamically cope with sudden traffic or regional density changes, effectively improve the performance and flexibility of distributed trajectory stream analysis, adaptively adjust the data partitioning according to real-time data distribution, optimize load balancing and resource allocation, solve the limitations of the traditional static partitioning scheme, and significantly improve the processing efficiency and system performance of the trajectory stream analysis.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of distributed computing, and particularly relates to a distributed trajectory stream analysis elastic partitioning method based on reinforcement learning driving. BACKGROUND

[0002] With the popularity of GPS (Global Positioning System) devices and smartphones, the trajectory data generated by mobile objects such as vehicles and pedestrians has shown explosive growth. Such data has become an important basic resource for urban computing and intelligent transportation systems, which can support real-time traffic monitoring, congestion prediction, and road network optimization and other application scenarios. In order to efficiently process massive trajectory stream data, existing research has developed a distributed trajectory analysis framework, as described in the literature [Alarabi L. Summit: A Scalable System for Massive Trajectory Data Management. Proceedings of the 26th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems. 2018: 612-613], which improves data processing capacity through multi-node parallel computing.

[0003] In distributed trajectory stream analysis, the data partitioning strategy plays a crucial role, and the partitioning of data directly affects the workload balance, resource utilization efficiency, and overall performance of the system. Traditional partitioning methods, as described in the literature [Shaikh S A, Kitagawa H, Matono A, et al. A Distributed and Scalable Framework for Low-Latency Continuous Trajectory Stream Processing. IEEE Access Journal, 2024], mostly rely on static partitioning, that is, trajectory data is divided according to certain fixed rules before processing begins. However, in actual scenarios, trajectory data exhibits dynamic spatiotemporal characteristics, and traffic flow rapidly fluctuates during peak hours such as morning and evening rush hours, resulting in the original static division being unable to adapt to real-time changes. This contradiction causes an imbalance in the allocation of computing resources, ultimately leading to a decrease in system throughput and an increase in response delay.

[0004] Therefore, designing an elastic partitioning method for distributed trajectory stream analysis has become an urgent need in academia and industry. First, the system needs to adapt to changes in trajectory distribution while maintaining spatial proximity. Many existing dynamic methods commonly used in distributed systems alleviate the load imbalance problem by dividing hot keys into sub-keys, but these methods ignore spatial relationships, which are crucial for trajectory analysis tasks such as similarity queries and clustering. Although some spatial-aware methods can maintain spatial proximity, they still rely on static partitioning frameworks and cannot dynamically adjust. Therefore, when a certain partition is overloaded or encounters sudden traffic, static schemes cannot effectively redistribute the load, leading to performance bottlenecks.

[0005] Secondly, the elastic partition method also needs to solve the problems of cold start and concept drift. The cold start problem refers to that the machine learning model cannot make accurate partition decisions at the beginning of training due to the lack of sufficient training data, resulting in unstable system performance. Concept drift refers to the change of data features over time, such as seasonal fluctuations of traffic flow or the impact of unexpected events. If the partition strategy is not adjusted in time, the system performance will decrease, and even the existing partition strategy will be invalid. The spatio-temporal coupling in trajectory stream makes drift detection and response more complex, and special mechanisms must be designed to ensure the long-term stability and efficiency of the partition strategy. SUMMARY

[0006] In view of the above, the present application provides a distributed trajectory stream analysis elastic partition method based on reinforcement learning driving, which can solve the problems of uneven load, excessive memory pressure and throughput decline of existing static partition methods in dynamic trajectory stream data analysis.

[0007] A distributed trajectory stream analysis elastic partition method based on reinforcement learning driving, comprising the following steps:

[0008] (1) In the offline stage, the agent collects experience and updates the action selection strategy through historical trajectory data;

[0009] (2) In the online stage, the agent periodically selects actions in the hierarchical action space based on the trajectory distribution state in the current time period and according to the learned action selection strategy, and finally determines the partition action of the current period, while receiving the reward signal returned by the trajectory stream analysis system;

[0010] (3) The agent extracts experience from the buffer according to the state of the current period, and updates the action selection strategy in real time according to the experience;

[0011] (4) The trajectory stream analysis system receives the partition action selected by the agent, constructs a dynamic grid index based on the given lower and upper coordinates of the spatial domain, and divides the entire space into multiple grid units with the action selected by the agent as the grid division parameter;

[0012] (5) The trajectory stream analysis system dynamically divides the real-time trajectory stream according to the fixed time window size, generates a number of sub-trajectories, and calculates the corresponding grid index key for each sub-trajectory;

[0013] (6) The trajectory stream analysis system distributes the sub-trajectories to the corresponding computing nodes in parallel according to the grid index key, to realize grid-based load balancing and multiple trajectory analysis task processing.

[0014] Further, the specific implementation of step (1) is as follows:

[0015] S11: The agent collects the distribution of historical trajectory data using the space-time context collector, and constructs a state matrix accordingly;

[0016] S12: The agent randomly selects an action in the action space, obtains the load balancing degree and memory usage irrelevant to the specific analysis task fed back by the trajectory stream analysis system, and calculates the low-level reward obtained by weighting the two;

[0017] S13: The agent combines the current state, the selected action, and the corresponding reward into experience;

[0018] S14: The experience is dimensionally reduced, and then the reduced experience is clustered;

[0019] S15: The experience optimized by step S14 is stored in an offline buffer, and the offline buffer uses a size-aware pruning strategy to preferentially remove redundant experience from over-representative clusters;

[0020] S16: The agent iteratively updates the action selection strategy using the experience in the offline buffer as a sample.

[0021] Further, the specific implementation of step (2) is as follows:

[0022] S21: The agent uses the space-time context collector to count the trajectory distribution in the current time period, and constructs a state matrix accordingly;

[0023] S22: The agent performs action selection in the hierarchical action space based on the state matrix and the action selection strategy, and determines the partition action of the current period;

[0024] S23: The trajectory stream analysis system executes the action selected in the previous step and obtains the feedback reward signal, and the agent combines the state matrix, the selected action, and the corresponding reward into experience and stores it in the online buffer. The online buffer uses cosine similarity for state comparison, and uses the online DP-Medoids (Dynamic Programming Center Point) algorithm for clustering. In each cluster, the representative quality of the experience is maintained through a cluster balance eviction strategy.

[0025] Further, in step S22, the agent first selects an action in the coarse-grained action space, which corresponds to a large range of grid configurations. The coarse-grained action selection is based on the reward value weighted by the low-level reward and the high-level reward. The high-level reward is the throughput of the specific task. Then, for each coarse-grained selected action, the agent further selects in the fine-grained action space to refine the grid configuration. The fine-grained action selection is based on the high-level reward.

[0026] Further, in step (3), the agent selects a set of experiences with the maximum cosine similarity to the current state from the online buffer and the offline buffer to update the action selection policy of the agent, and two independent mean square error loss minimizations are used for the high-level reward and the low-level reward predicted by the agent respectively; meanwhile, the agent handles the cold start problem and the concept drift problem through the experience replay mechanism to ensure the continuous optimization of the action selection policy.

[0027] Further, in step (4), the trajectory stream analysis system first calculates the grid size according to the maximum spatial span and the grid division parameter (i.e. the number of grids) selected by the agent based on the given lower and upper coordinates of the spatial domain, and then divides the spatial region into multiple grid units according to the grid size to construct a dynamic grid index, which will be used for subsequent trajectory stream allocation and parallel analysis task processing. The grid division parameter determines the granularity of the grid, and the dynamic grid index is adjusted in real time according to the action to achieve elastic partitioning, so as to flexibly cope with different loads and traffic changes.

[0028] Further, in step (5), the trajectory stream analysis system first divides the real-time trajectory stream into multiple sub-trajectories by giving a time window size, each sub-trajectory being a continuous sequence of spatio-temporal points with a length not exceeding the time window size, and each spatio-temporal point in the sequence containing the ID of the moving object, the latitude and longitude coordinates, and the timestamp; then, the grid index key of the sub-trajectory is calculated according to the spatial position information of the sub-trajectory and the dynamic grid index, so as to map the sub-trajectory to the corresponding grid unit, which will be used in the subsequent trajectory stream analysis process to distribute the sub-trajectory to the appropriate computing node for parallel processing.

[0029] Further, in step (6), each grid unit is assigned to a computing node, which receives and processes the sub-trajectory related to the grid unit according to the predefined partitioning strategy. Different computing nodes process the sub-trajectories assigned to them in parallel, and are scheduled according to the grid granularity, spatial distribution, computing node processing capacity, and real-time load to achieve load balancing. Meanwhile, the trajectory stream analysis system supports the execution of multiple trajectory analysis tasks, including trajectory range query, trajectory similarity query, and density-based trajectory clustering task. Each computing node executes the corresponding analysis operation according to the assigned sub-trajectory and the task type when executing these tasks. The trajectory range query task analyzes the behavior of objects within a specific region by filtering trajectories within the region, the trajectory similarity query task is used to compare the similarity of different trajectories to help identify similar behavior patterns, and the density-based trajectory clustering task discovers traffic flow dense areas by analyzing the spatial density of trajectories.

[0030] A computer device comprises a memory and a processor, the memory has a computer program stored therein, and the processor is configured to execute the computer program to implement the above-mentioned distributed trajectory stream analysis elastic partitioning method based on reinforcement learning driving.

[0031] A computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the above-mentioned distributed trajectory stream analysis elastic partitioning method based on reinforcement learning driving.

[0032] In order to solve the problem of partitioning elasticity adjustment, the present application converts the elastic partitioning problem into a grid index parameter dynamic adjustment problem. Specifically, in the present application, the partitioning of trajectory data does not depend on static division rules, but realizes elastic partitioning by dynamically adjusting the grid index parameters to adapt to the changes of trajectory stream data in time and space. At the same time, the present application models the dynamic adjustment problem of grid index parameters as CBwK (Contextual Bandit Problem with Resource Constraints), in which the agent continuously selects the optimal grid configuration according to the current state of the system through the reinforcement learning strategy, and balances between exploration and utilization to optimize load balancing and resource allocation; in this way, the system can dynamically cope with sudden traffic or regional density changes, effectively improving the performance and flexibility of distributed trajectory stream analysis.

[0033] Therefore, the method of the present application can adaptively adjust data partitioning according to real-time data distribution through the reinforcement learning model, optimize load balancing and resource allocation, and solve the limitations of traditional static partitioning schemes, thereby significantly improving the processing efficiency and system performance of trajectory stream analysis. In addition, the method of the present application can also effectively support various trajectory analysis tasks such as trajectory range query, trajectory similarity query and density-based trajectory clustering task, further improving the flexibility and practicality of the system. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 FIG. 1 is a flowchart of the method of the present application. DETAILED DESCRIPTION

[0035] In order to more specifically describe the present application, the technical solutions of the present application will be described in detail below in combination with the drawings and specific embodiments.

[0036] The embodiment provides a distributed trajectory stream analysis elastic partitioning method based on reinforcement learning driving, and the specific implementation steps are as shown in Figure 1

[0037] Step S11: In the offline stage, the agent collects experience through historical trajectory data and updates the action selection strategy.

[0038] ​Specifically, the agent collects the distribution of historical trajectory data through the spatio-temporal context collector and constructs a state matrix S. Then, the agent randomly selects an action A in the action space, which corresponds to a specific configuration parameter (i.e., grid partition parameter) of the trajectory stream partition index, thereby affecting the way of computing resource mapping (such as allocating grid cells to different computing nodes); by executing the action A, the agent actually selects and tries between different partition schemes, thereby affecting the way of trajectory stream partitioning and the load distribution result; the agent obtains the load balancing degree R b and memory usage R i reward, and calculates the weighted sum of the two to obtain the low-level reward . Then, the agent combines the state matrix, the selected action, and the corresponding reward into experience, and performs principal component analysis dimensionality reduction on the experience to effectively capture the core mode and reduce the dimension of the experience. Further, the Gaussian mixture model is used to cluster the reduced experience, and the clustering result determines the organization method of the experience. In order to maintain the balance of data distribution, the buffer adopts a size-aware pruning strategy, which preferentially removes redundant experience from over-representative clusters. Finally, these optimized experiences are stored in the offline buffer, ensuring the retention of representative historical experiences and effectively alleviating the problem of catastrophic forgetting. Finally, the agent updates its action selection policy θ according to the experience data in the offline buffer. This step simulates different trajectory stream distributions to train the reinforcement learning model, enabling it to identify potential load balancing configurations.

[0039] Step S12: In the online phase, the agent periodically determines the trajectory distribution state in the current time period and sequentially performs action selection in the hierarchical action space according to the learned policy, finally determines the partition action in this period, and receives the reward signal returned by the trajectory stream analysis system. Specifically:

[0040] S21: The spatio-temporal context collector statistics the trajectory distribution in the current time period and constructs a state matrix S based on the statistics ;

[0041] S22: The agent selects actions in the hierarchical action space based on the learned policy θ according to the state matrix S. Specifically, the agent first selects a coarse-grained action in the coarse-grained action space, which corresponds to a large range of grid configurations. The coarse-grained action selection is based on the predicted reward , the high-level reward R h is the throughput of the specific task; the coarse-grained action space is defined as:

[0042]

[0043] wherein: is the interval between actions, and The maximum and minimum values of the predefined action.

[0044] Then, for each coarse-grained selection action , the agent makes further selection in the fine-grained action space to refine the grid configuration; the fine-grained action selection is based on the reward R h predicted by the agent; the fine-grained action space is defined as:

[0045]

[0046] wherein: is an interval around the coarse-grained action , defined as:

[0047]

[0048] This hierarchical action selection method can effectively balance the size of the search space and the accuracy of policy adjustment, ensuring that the agent can achieve the best configuration when optimizing load balancing and resource allocation; during this process, the agent uses an exploration-exploitation strategy to balance the exploration of unknown configurations and the utilization of known efficient configurations. In the exploration phase, the agent will select configurations with the highest uncertainty and information content to collect more knowledge about system performance to improve the policy more effectively; in the exploitation phase, the agent uses the learned knowledge to select configurations that are expected to bring high returns to optimize the performance of the current system. Through this balance between exploration and exploitation, the agent can effectively improve the efficiency of grid configuration selection and thus improve the overall performance of the trajectory stream analysis system.

[0049] S23: Perform the action selected in the previous step on the trajectory stream analysis system, and obtain the low-level reward and high-level reward R h from the system feedback, combine them with the state matrix and the selected action into experiences. Subsequently, these experiences are stored in the online buffer. The online buffer uses cosine similarity for state comparison and uses the online DP-Medoids algorithm for clustering; within each cluster, the representative quality of the experiences is maintained through a cluster balance eviction strategy.

[0050] Step S13: The agent extracts experiences from the buffer according to the current period's state, and updates the action selection strategy in real time according to the experiences.

[0051] Considering the current state matrix S, a set of experiences with the highest cosine similarity to the current state is selected from the online and offline buffers to update the agent's policy, and this update minimizes the two independent mean square error losses:

[0052]

[0053]

[0054] wherein: is the number of experiences for each draw, is the low-level reward predicted by the agent, is the high-level reward predicted by the agent.

[0055] Step S14: The trajectory stream analysis system receives the partitioned action selected by the agent, and based on the given lower and upper coordinate limits of the spatial domain, constructs a dynamic grid index with the action selected by the agent as the grid partition parameter, and equally divides the entire space into multiple grid cells. The grid index will be adjusted in real time according to the action to achieve elastic partitioning.

[0056] Specifically, the system first calculates the grid size g based on the given lower and upper coordinate limits of the spatial domain , the maximum spatial span (i.e., the maximum difference in longitude or latitude), and the number of grids A selected by the agent:

[0057]

[0058] Then, according to the calculated grid size g, the spatial region is divided into multiple grid cells, each with a size of These grid cells will be used for subsequent trajectory stream allocation and parallel analysis task processing, and the grid index will be adjusted in real time according to the action to achieve elastic partitioning.

[0059] Step S15: The trajectory stream analysis system dynamically divides the real-time trajectory stream according to a predetermined fixed time window, generates a number of sub-trajectories, and calculates the corresponding grid index key for each sub-trajectory.

[0060] Specifically, given a time window size , the real-time trajectory stream is divided into multiple sub-trajectories , each of which is a continuous trajectory segment that satisfies the time condition , i.e., the time span of each sub-trajectory does not exceed the specified window size .

[0061] Here, is a spatiotemporal point, represented as , where is the unique identifier of the moving object, and are the longitude and latitude coordinates of the point, is the timestamp of the point.

[0062] Then, for each sub-trajectory , its grid index key is calculated according to its spatial position (i.e. longitude and latitude) and the constructed grid index. The calculation of the grid index key depends on the selected grid partitioning strategy, and maps each sub-trajectory to the corresponding grid cell according to its spatial position. This grid index key will be used in the subsequent trajectory stream analysis process to distribute the sub-trajectories to appropriate computing nodes for parallel processing.

[0063] Step S16: The trajectory stream analysis system distributes the sub-trajectories to the corresponding computing nodes in parallel according to their grid index keys, to achieve grid-based load balancing and processing of multiple trajectory analysis tasks.

[0064] Specifically, each grid cell is assigned to a computing node, which receives and processes the sub-trajectories related to it according to the predefined partitioning strategy. This way, different computing nodes can process the assigned sub-trajectories in parallel, improving processing efficiency. At the same time, the system supports the execution of multiple trajectory analysis tasks, including trajectory range queries, trajectory similarity queries, and density-based trajectory clustering tasks. When executing these tasks, each computing node performs the corresponding analysis operation according to the assigned sub-trajectories and task type. For example, the trajectory range query task analyzes the behavior of objects within a specific area by filtering trajectories within the area, the trajectory similarity query compares the similarity of different trajectories to help identify similar behavior patterns, and the density-based trajectory clustering task discovers areas of high traffic density by analyzing the spatial density of trajectories. Through this grid-based distributed data distribution mechanism, the system can effectively distribute computing load, improve processing speed, and ensure efficient execution of different analysis tasks in a distributed environment. This mechanism not only speeds up the processing of trajectory streams, but also ensures the rational use of resources and efficient scheduling of tasks.

[0065] The above description of the embodiments is to facilitate the understanding and application of the present application by those skilled in the art. Those skilled in the art can easily make various modifications to the above embodiments and apply the general principles described herein to other embodiments without inventive labor. Therefore, the present application is not limited to the above embodiments, and any improvements and modifications made to the present application by those skilled in the art should be within the scope of protection of the present application.

Claims

1. A distributed trajectory stream analysis elastic partitioning method based on reinforcement learning driving, characterized in that, Comprise the following steps: (1) Offline stage, the agent collects experience through historical trajectory data and updates the action selection strategy; (2) Online stage, the agent periodically selects actions in the hierarchical action space based on the trajectory distribution state in the current time period and the learned action selection strategy, and finally determines the partition action of this period, while receiving the reward signal returned by the trajectory stream analysis system; (3) The agent extracts experience from the buffer according to the state of the current period, and updates the action selection strategy in real time according to the experience; (4) The trajectory stream analysis system receives the partition action selected by the agent, and based on the given lower and upper coordinates of the spatial domain, constructs a dynamic grid index with the action selected by the agent as the grid division parameter, and divides the entire space into multiple grid units; (5) The trajectory stream analysis system dynamically divides the real-time trajectory stream according to the fixed time window size to generate several sub-trajectories, and calculates the corresponding grid index key for each sub-trajectory; (6) The trajectory stream analysis system distributes the sub-trajectories in parallel to the corresponding computing nodes according to the grid index key, so as to realize grid-based load balancing distribution and multiple trajectory analysis task processing.

2. The method of claim 1, wherein, The specific implementation of step (1) is as follows: S11: The agent uses the space-time context collector to count the distribution of historical trajectory data and constructs a state matrix accordingly; S12: The agent randomly selects an action in the action space, obtains the load balancing degree and memory usage unrelated to specific analysis tasks from the trajectory stream analysis system, and calculates the low-level reward obtained by weighting the two; S13: The agent combines the current state, the selected action and the corresponding reward into experience; S14: Dimensionality reduction is performed on the experience, and then the dimensionality reduced experience is clustered; S15: Store the experience optimized by step S14 in the offline buffer, and the offline buffer uses a size-aware pruning strategy to preferentially remove redundant experience from over-representative clusters; S16: The agent iteratively updates the action selection strategy using the experience in the offline buffer as a sample.

3. The method of claim 1, wherein, The specific implementation of step (2) is as follows: S21: The agent uses the space-time context collector to count the trajectory distribution in the current time period and constructs a state matrix accordingly; S22: The agent selects actions in the hierarchical action space based on the state matrix and the action selection strategy, and determines the partition action of this period; S23: The trajectory stream analysis system executes the action selected in the previous step and obtains the feedback reward signal, and the agent combines the state matrix, the selected action and the corresponding reward into experience and stores it in the online buffer. The online buffer uses cosine similarity for state comparison and adopts online DP-Medoids algorithm for clustering. In each cluster, the representative quality of experience is maintained through cluster balance eviction strategy.

4. The method of claim 3, wherein, In the step S22, the agent first selects an action in the coarse-grained action space, which corresponds to a large range of grid configurations. The coarse-grained action selection is based on the reward value weighted by the low-level reward and the high-level reward, and the high-level reward is the throughput of the specific task. Then, for each coarse-grained selected action, the agent further selects in the fine-grained action space to refine the grid configuration. The fine-grained action selection is based on the high-level reward.

5. The method of claim 1, wherein, In the step (3), the agent selects a set of experiences with the maximum cosine similarity to the current state from the online buffer and the offline buffer to update the action selection policy of the agent according to the current state matrix. In the updating process, two independent mean square error loss minimizations are adopted for the predicted high-level reward and low-level reward of the agent, respectively. Meanwhile, the agent processes the cold start problem and the concept drift problem through the experience replay mechanism to ensure the continuous optimization of the action selection policy.

6. The method of claim 1, wherein, In the step (4), the trajectory stream analysis system first calculates the grid size according to the maximum spatial span and the grid partition parameter selected by the agent based on the given lower and upper limit coordinates of the spatial domain, and then divides the spatial region into multiple grid cells to construct a dynamic grid index according to the grid size. These grid cells will be used for subsequent trajectory stream allocation and parallel analysis task processing. The grid partition parameter determines the granularity of the grid, and the dynamic grid index will be adjusted in real time according to the action to realize elastic partitioning, so as to flexibly cope with different loads and traffic changes.

7. The method of claim 1, wherein, In the step (5), the trajectory stream analysis system first divides the real-time trajectory stream into multiple sub-trajectories by giving a time window size. Each sub-trajectory is a continuous sequence of spatiotemporal points, and its length does not exceed the time window size. Each spatiotemporal point in the sequence contains the ID of the moving object, the latitude and longitude coordinates, and the timestamp. Then, the trajectory stream analysis system calculates the grid index key of the sub-trajectory according to the spatial position information of the sub-trajectory and the dynamic grid index, and maps the sub-trajectory to the corresponding grid cell. The grid index key will be used in the subsequent trajectory stream analysis process to distribute the sub-trajectory to the appropriate computing node for parallel processing.

8. The method of claim 1, wherein, In the step (6), each grid cell is assigned to a computing node, which receives and processes the sub-trajectory related to the grid cell according to the predefined partitioning strategy. Different computing nodes process the sub-trajectories assigned to them in parallel, and are scheduled according to the grid granularity, spatial distribution, computing node processing capacity, and real-time load to achieve load balancing. Meanwhile, the trajectory stream analysis system supports the execution of multiple trajectory analysis tasks, including trajectory range query, trajectory similarity query, and density-based trajectory clustering task. Each computing node executes the corresponding analysis operation according to the assigned sub-trajectory and the task type when executing these tasks. The trajectory range query task analyzes the behavior of objects within a specific region by filtering trajectories within the region, the trajectory similarity query task compares the similarity of different trajectories to help identify similar behavior patterns, and the density-based trajectory clustering task discovers traffic dense areas by analyzing the spatial density of trajectories.

9. A computer device comprising a memory and a processor, said memory having stored therein a computer program, characterized in that: The processor is configured to execute the computer program to implement the method for distributed trajectory stream analysis elastic partitioning driven by reinforcement learning according to any one of claims 1-8.

10. A computer readable storage medium storing a computer program, characterized in that: The computer program is configured to be executed by the processor to implement the method for distributed trajectory stream analysis elastic partitioning driven by reinforcement learning according to any one of claims 1-8.

Images