Intelligent taxi scheduling method and system based on double attention mechanism

Abstract

The invention provides an intelligent taxi scheduling method and system based on a double attention mechanism, and the method comprises the steps: dividing an urban traffic network into hexagonal grid units through an intelligent agent, and combining the local state of each grid unit with the state of an adjacent grid through an adjacent matrix, dynamically generating traffic environment state vectors according to the traffic environment data; obtaining a local supply and demand state and a global supply and demand state based on the traffic environment state vector; the model is trained through the global supply and demand state, and an optimal scheduling strategy is obtained; an execution action is generated through the optimal scheduling strategy; and updating the strategy by executing the action. The method shows remarkable superiority in a dynamic complex traffic environment, and an efficient and feasible solution is provided for scheduling and repositioning optimization of an urban intelligent traffic system.

Description

Technical Field

[0001] The present invention relates to the technical field of traffic data processing, and in particular to an intelligent taxi dispatching method and system based on a dual attention mechanism. Background Art

[0002] With the acceleration of urbanization, the pressure on urban transportation systems is increasing, especially during peak travel periods, when the contradiction between supply and demand of taxis is particularly prominent. In recent years, with the development of big data and artificial intelligence technologies, intelligent dispatching systems have become an effective means to solve traffic dispatching problems. However, how to achieve reasonable dispatching and relocation of taxis in a dynamic and complex urban environment is still an urgent problem to be solved in urban traffic management.

[0003] At present, most taxi dispatch systems mainly rely on historical data or simple heuristic methods for supply and demand forecasting and dispatching decisions. These methods are usually based on static models and cannot adapt to the ever-changing demand patterns in cities. These methods lack sufficient flexibility when facing high-dimensional and dynamic environments, making it difficult to achieve real-time decision-making and multi-dimensional optimization. For example, traditional supply and demand forecasting methods are difficult to accurately predict future travel demand in a dynamically changing demand environment, resulting in a mismatch between taxi supply and demand. On the other hand, the idle rate of taxis is also difficult to reduce, and many taxis are idle at the edge of the city and cannot be relocated to areas with higher demand in time. Therefore, it is necessary to focus on intelligent taxi dispatching and optimization methods based on multi-agent reinforcement learning.

[0004] In order to solve the current problems such as difficulty in adapting to the dynamic and complex traffic environment, insufficient information fusion, imperfect optimization of the collaboration mechanism and poor adaptability, it is still necessary to develop more effective methods to improve the overall optimization of the urban taxi dispatch system. Summary of the invention

[0005] In view of the above situation, the main purpose of the present invention is to propose an intelligent taxi dispatching method and system based on a dual attention mechanism to solve the above technical problems.

[0006] The present invention proposes an intelligent taxi dispatching method based on a dual attention mechanism, the method comprising the following steps: Step 1: Use the intelligent agent to divide the urban traffic network into hexagonal grid units; Through the adjacency matrix, the traffic environment data is dynamically converted into a traffic environment state vector by combining the local state of each grid unit and the state of the adjacent grids; Step 2: Based on the traffic environment state vector, a dual attention mechanism is used to make predictions to obtain the local supply and demand state and the global supply and demand state; Step 3: Input the global supply and demand status into the decision layer for processing and combine it with the critic network to obtain the global value; According to the local supply and demand status, actions are obtained through the local attention mechanism; Based on the action, a global reward is generated through the environment simulator, the state and position of the taxi are updated in the environment based on the global reward, and feedback information from the environment is obtained; Based on the advantage function and combined with the feedback information of the environment, the intelligent agent is trained to obtain the trained intelligent agent, and the optimal scheduling strategy is obtained through the trained intelligent agent; Step 4: Input the optimal scheduling strategy into the execution layer to generate execution actions; By executing actions, the policy is updated.

[0007] The present invention also proposes an intelligent taxi dispatching system based on a dual attention mechanism, the system comprising: Preprocessing module for: Use intelligent agents to divide the urban transportation network into hexagonal grid cells; Through the adjacency matrix, the traffic environment data is dynamically converted into a traffic environment state vector by combining the local state of each grid unit and the state of the adjacent grids; Dual attention prediction module for: Based on the traffic environment state vector, the dual attention mechanism is used to predict the local supply and demand state and the global supply and demand state. Training modules for: The global supply and demand status is input into the decision layer for processing, and combined with the critic network to obtain the global value; According to the local supply and demand status, actions are obtained through the local attention mechanism; Based on the action, a global reward is generated through the environment simulator, the state and position of the taxi are updated in the environment based on the global reward, and feedback information from the environment is obtained; Based on the advantage function and combined with the feedback information of the environment, the intelligent agent is trained to obtain the trained intelligent agent, and the optimal scheduling strategy is obtained through the trained intelligent agent; Policy update module, used to: Input the optimal scheduling strategy into the execution layer to generate execution actions; By executing actions, the policy is updated.

[0008] Compared with the prior art, the present invention has the following beneficial effects: 1. This paper proposes a two-level dispatching architecture model framework. Through the collaborative work of the decision-making layer and the execution layer, it realizes the effective collaboration of multiple agents in a complex urban traffic environment. At the same time, it divides taxi allocation and vehicle relocation into two problems. Through the combination of global and local information, it optimizes demand allocation and empty vehicle movement respectively. This framework significantly improves the dispatching efficiency in urban traffic and reduces the empty vehicle rate.

[0009] 2. The present invention introduces a dual attention mechanism to extract global and local supply and demand information, thereby guiding the dispatching behavior of the intelligent agent. This mechanism can dynamically capture the changes in demand patterns in the city and adjust the dispatching strategy in real time, reducing the complexity of the system.

[0010] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description or learned through embodiments of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 This is a flow chart of the intelligent taxi dispatching method based on the dual attention mechanism proposed by the present invention; Figure 2 This is a model framework diagram of the intelligent taxi dispatching method based on the dual attention mechanism proposed in the present invention; Figure 3 This is the core algorithm diagram of the intelligent taxi dispatching method based on the dual attention mechanism proposed in the present invention; Figure 4 This is a comparison chart of order response rates of the intelligent taxi dispatching method based on the dual attention mechanism proposed in the present invention; Figure 5 This is a comparison chart of platform revenues of the smart taxi dispatching method based on the dual attention mechanism proposed in the present invention; Figure 6 A line chart comparing passenger response rates in the ablation experiment for verifying the effectiveness of the dual attention mechanism proposed in the present invention; Figure 7 A line chart comparing the total platform revenue of the ablation experiment proposed in the present invention to verify the effectiveness of the dual attention mechanism; Figure 8 This is a framework diagram of the intelligent taxi dispatching system based on the dual attention mechanism proposed in the present invention. DETAILED DESCRIPTION

[0012] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0013] These and other aspects of the embodiments of the present invention will be apparent with reference to the following description and accompanying drawings. In these descriptions and accompanying drawings, some specific implementations of the embodiments of the present invention are specifically disclosed to represent some ways of implementing the principles of the embodiments of the present invention, but it should be understood that the scope of the embodiments of the present invention is not limited thereto.

[0014] See also Figure 1 and Figure 2 The embodiment of the present invention proposes an intelligent taxi dispatching method based on a dual attention mechanism, and the method comprises the following steps: Step 1: Use the intelligent agent to divide the urban traffic network into hexagonal grid units; Through the adjacency matrix, the traffic environment data is dynamically converted into a traffic environment state vector by combining the local state of each grid unit and the state of the adjacent grids; In this step, the traffic environment data is dynamically converted into a traffic environment state vector, and the corresponding process relationship is: ; in, represents the total number of grids, represents the time step, represents the traffic environment state vector, Both represent the traffic status in the grid area.

[0015] Step 2: Based on the traffic environment state vector, a dual attention mechanism is used to make predictions to obtain the local supply and demand state and the global supply and demand state; See also Figure 3 In step 2, based on the traffic environment state vector, the dual attention mechanism is used to predict the local supply and demand state and the global supply and demand state. The specific steps are as follows: Through the traffic environment state vector, local observation is obtained, and the relationship between the corresponding process is: ; in, represents a local observation, represents the number of idle vehicles in the current grid cell, Indicates the number of currently uncompleted orders. represents the set of states of adjacent grid cells; Based on local observations, the global observation is obtained, and the relationship between the corresponding process is: ; in, represents the global observation, represents the total number of grid cells; Based on the local supply and demand feature vectors of the grid unit and the adjacent grid units, the attention weights of the adjacent grids are obtained through the local attention mechanism. The relationship between the corresponding process is: ; in, represents the trainable weight matrix in the local attention layer, and Represent the supply and demand feature vectors of the current grid cell and the adjacent grid cells, represents the attention weight of adjacent grids, represents the local weight factor; Based on the global supply and demand feature vectors of the grid unit and its adjacent grid units, the attention weight of the grid is obtained through the global attention mechanism. The relationship between the corresponding process is: ; in, represents the attention weight of the grid, represents the global weight factor, represents the trainable weight matrix in the global attention layer, and Both represent the supply and demand feature vectors of the grid cells; Finally, according to the attention weights of adjacent grids and the attention weight of the grid, the local supply and demand state and the global supply and demand state are generated.

[0016] Step 3: Input the global supply and demand status into the decision layer for processing and combine it with the critic network to obtain the global value; According to the local supply and demand status, actions are obtained through the local attention mechanism; Based on the action, a global reward is generated through the environment simulator, the state and position of the taxi are updated in the environment based on the global reward, and feedback information from the environment is obtained; Based on the advantage function and combined with the feedback information of the environment, the intelligent agent is trained to obtain the trained intelligent agent, and the optimal scheduling strategy is obtained through the trained intelligent agent; In step 3, the global supply and demand state is input into the decision layer for processing, and the global value is obtained by combining with the critic network. In the process of obtaining the global value, the mean square error is used as the loss function, and the expression of the mean square error is: ; in, represents the parameters of the critic network, represents the discount factor, Represents the reward information after the agent executes, Indicates the period value, Indicates that given at time step The value estimated by the critic network in the state Indicates that given at time step The value estimated by the critic network in the state represents the total length of the time step, represents the mean square error function; In the process of obtaining actions based on the local supply and demand status through the local attention mechanism, the following sub-steps are also included: Based on the parameters of the policy network and the local observations to generate action distribution, the corresponding process has the following relationship: ; in, represents the probability of the actor network choosing an action under local observation, Indicates the actions that the taxi can choose; Through the taxi's strategy function, we can get the taxi's action, and the relationship corresponding to this process is: ; in, Indicates that the taxi is at time step Select the action, represents the taxi’s policy function, Indicates that the taxi is at time step Status at the time; Using the actions and status of the taxi, we can get the value function optimization strategy estimated by the critic network. The corresponding relationship is: ; in, represents the taxi’s policy parameters, Indicates that it has been processed by the value calculation function. Represents the weight of the function used by each taxi when making decisions; Based on the action, the global reward is generated through the environment simulator. The specific steps are as follows: Define the balance indicator, the definition is: ; in, Represents the balance indicator, Indicates that at time step The number of passenger demands in the area below, Indicates that at time step The number of idle vehicles in the area below; The local reward is defined by the balance indicator, which is defined as: ; in, Indicates partial reward; The global reward is obtained based on the local reward, and the corresponding process has the following relationship: ; in, represents the global reward; Based on the advantage function and combined with the feedback information of the environment, the agent is trained to obtain the trained agent, and the optimal scheduling strategy is obtained through the trained agent. The process also includes the following sub-steps: The function of optimizing strategy parameters is constructed through the advantage function, and the relationship between the corresponding process is: ; in, represents the advantage function, represents the function used to optimize the strategy parameters, represents the parameters of the actor network, It means to solve the expected joint distribution of state and action under the current strategy. Indicates the status under the current strategy. Represents the action under the current strategy; In the process of advantage function calculation, it is enhanced by generalized advantage estimation, and the relationship between the corresponding process is: ; in, represents the advantage error, Indicates that at time step Instant rewards when represents the advantage error at the next time step, represents the index of the time step; During the training process, the loss function of the actor network is constructed based on the advantage function, and the relationship between the corresponding process is: ; in, represents the loss function of the actor network, Represents the operation of taking the expected value of all samples at multiple time steps, represents the pruning hyperparameter, Indicates that a shear operation has been performed; represents the probability of taking an action under the same local observation based on the old strategy; The parameters of the actor network are updated through the loss function of the actor network, and the relationship between the corresponding process is: ; in, represents the learning rate, represents the gradient; The total loss function expression is: ; in, represents the loss function of the critic network, represents the entropy term, represents the weight hyperparameter of the loss function of the critic network, The weight hyperparameter representing the entropy term Step 4: Input the optimal scheduling strategy into the execution layer to generate execution actions; Update the strategy by executing actions; In step 4, the optimal scheduling strategy is input into the execution layer to generate execution actions. The corresponding process relationship is: ; in, Represents the specific actions sampled from the optimal scheduling strategy; By executing actions, the strategy is updated, and the corresponding relationship in the process is: ; in, Represents the updated policy parameters.

[0017] See also Figure 8 The embodiment of the present invention also provides an intelligent taxi dispatching system based on a dual attention mechanism, the system comprising: Preprocessing module for: Use intelligent agents to divide the urban transportation network into hexagonal grid cells; Through the adjacency matrix, the traffic environment data is dynamically converted into a traffic environment state vector by combining the local state of each grid unit and the state of the adjacent grids; Dual attention prediction module for: Based on the traffic environment state vector, the dual attention mechanism is used to predict the local supply and demand state and the global supply and demand state. Training modules for: The global supply and demand status is input into the decision layer for processing, and combined with the critic network to obtain the global value; According to the local supply and demand status, actions are obtained through the local attention mechanism; Based on the action, a global reward is generated through the environment simulator, the state and position of the taxi are updated in the environment based on the global reward, and feedback information from the environment is obtained; Based on the advantage function and combined with the feedback information of the environment, the intelligent agent is trained to obtain the trained intelligent agent, and the optimal scheduling strategy is obtained through the trained intelligent agent; Policy update module, used to: Input the optimal scheduling strategy into the execution layer to generate execution actions; By executing actions, the policy is updated.

[0018] Furthermore, in order to verify the effectiveness of the model proposed in this invention and effectively simulate the supply and demand dynamics in urban traffic, an environmental simulator is designed to train the reinforcement learning algorithm for traffic relocation. The steps are as follows: Step 1: At each time step, the simulator extracts order information for the corresponding time period from the historical data set and distributes these orders to the hexagonal grid cells divided by the city. The order generation process is based on the historical demand distribution of the region and combined with the traffic conditions at the time to ensure that the generated orders are consistent with the supply and demand fluctuations in reality; Step 2: In the current time step, the simulator performs order-vehicle matching operations in each grid area based on order requirements and the status of available vehicles. The matching process uses an optimized matching algorithm that comprehensively considers factors such as order priority, vehicle current location, and estimated arrival time, aiming to maximize matching efficiency and passenger satisfaction; Step 3: When the supply and demand relationship in a certain area is unbalanced, the simulator relocates idle vehicles to areas with higher demand based on the learned scheduling strategy. The relocation strategy not only depends on the demand situation in the current area, but also uses global state information to predict demand changes in future time steps to ensure the forward-looking and global optimality of scheduling; Step 4: At the end of each time step, the simulator updates the distribution of vehicles and passenger demand in each area based on the results of order matching and vehicle repositioning. This update provides new input for the next time step, ensuring that the simulator can dynamically reflect the spatiotemporal changes in traffic supply and demand; By looping steps one, two, three, and four, the simulator can effectively simulate the supply and demand changes and vehicle scheduling in the real traffic system, providing a reliable environment for the training of reinforcement learning algorithms.

[0019] For further information, see Figure 4 and Figure 5In order to verify the effectiveness of the invention and its role in improving the order response rate and platform revenue, the order response rate and platform revenue of the invention are compared with those of 7 other algorithms under different taxi fleet sizes ranging from 1,000 to 10,000. The algorithms used to compare the invention include Random algorithm, cDQN algorithm, cDQN algorithm, DDPG algorithm, MADDPG algorithm, cDQN algorithm, MASAC algorithm and MAPPO algorithm. The results show that MADARL has shown excellent performance in most experimental settings, especially in large-scale taxi dispatching tasks, outperforming other baseline algorithms.

[0020] Further, please refer to Table 1. In order to verify the independent contribution of the local attention mechanism and the global attention mechanism in the dual attention mechanism algorithm proposed in the present invention in the MADARL algorithm, the purpose of this ablation experiment is to verify the independent contribution of the local attention mechanism and the global attention mechanism in the MADARL algorithm; Table 1 Ablation experiment results

[0021] The experimental results show that as the system scales up, the order response rate (ORR) and platform cumulative revenue (PCR) increase. These results show that the dual attention mechanism can effectively combine local and global information to improve the response speed and resource allocation efficiency of the scheduling system. Especially in large-scale scheduling tasks, the dual mechanism significantly improves the overall performance. For further information, see Figure 6 as well as Figure 7 ,The experimental results show that the hybrid dual attention mechanism (SA-GC) with local attention mechanism and global attention mechanism performs well in all indicators, especially in large-scale systems, significantly improving order response rate, platform revenue and comprehensive index. Compared with other experimental groups, the dual attention mechanism can better balance the local and global information flow and improve the collaborative decision-making ability of intelligent agents. Through specific numerical improvements, the effectiveness of the MADARL algorithm in processing complex multi-agent tasks is further verified.

[0022] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0023] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0024] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. An intelligent taxi dispatching method based on dual attention mechanism, characterized in that: The method comprises the following steps: Step 1: Use the intelligent agent to divide the urban traffic network into hexagonal grid units; Through the adjacency matrix, the traffic environment data is dynamically converted into a traffic environment state vector by combining the local state of each grid unit and the state of the adjacent grids; Step 2: Based on the traffic environment state vector, a dual attention mechanism is used to make predictions to obtain the local supply and demand state and the global supply and demand state; Step 3: Input the global supply and demand status into the decision layer for processing and combine it with the critic network to obtain the global value; According to the local supply and demand status, actions are obtained through the local attention mechanism; Based on the action, a global reward is generated through the environment simulator, the state and position of the taxi are updated in the environment based on the global reward, and feedback information from the environment is obtained; Based on the advantage function and combined with the feedback information of the environment, the intelligent agent is trained to obtain the trained intelligent agent, and the optimal scheduling strategy is obtained through the trained intelligent agent; Step 4: Input the optimal scheduling strategy into the execution layer to generate execution actions; By executing actions, the policy is updated.

2. The intelligent taxi dispatching method based on the dual attention mechanism according to claim 1 is characterized in that: In step 1, the traffic environment data is dynamically converted into a traffic environment state vector, and the relationship between the corresponding process is: ； in, represents the total number of grids, represents the time step, represents the traffic environment state vector, Both represent the traffic status in the grid area.

3. The intelligent taxi dispatching method based on dual attention mechanism according to claim 2 is characterized in that: In step 2, based on the traffic environment state vector, a dual attention mechanism is used to make predictions to obtain the local supply and demand state and the global supply and demand state. The specific steps are as follows: Through the traffic environment state vector, local observation is obtained, and the relationship between the corresponding process is: ； in, represents a local observation, represents the number of idle vehicles in the current grid cell, Indicates the number of currently uncompleted orders. represents the set of states of adjacent grid cells; Based on local observations, the global observation is obtained, and the relationship between the corresponding process is: ； in, represents the global observation, represents the total number of grid cells; Based on the local supply and demand feature vectors of the grid unit and the adjacent grid units, the attention weights of the adjacent grids are obtained through the local attention mechanism. The relationship between the corresponding process is: ； in, represents the trainable weight matrix in the local attention layer, and Represent the supply and demand feature vectors of the current grid cell and the adjacent grid cells, represents the attention weight of adjacent grids, represents the local weight factor; Based on the global supply and demand feature vectors of the grid unit and its adjacent grid units, the attention weight of the grid is obtained through the global attention mechanism. The relationship between the corresponding process is: ； in, represents the attention weight of the grid, represents the global weight factor, represents the trainable weight matrix in the global attention layer, and Both represent the supply and demand feature vectors of the grid cells; According to the attention weights of adjacent grids and the attention weight of the grid, the local supply and demand state and the global supply and demand state are generated.

4. The intelligent taxi dispatching method based on dual attention mechanism according to claim 3 is characterized in that: In step 3, the global supply and demand state is input into the decision layer for processing, and the global value is obtained by combining with the critic network. In the process of obtaining the global value, the mean square error is used as the loss function, and the expression of the mean square error is: ； in, represents the parameters of the critic network, represents the discount factor, Represents the reward information after the agent executes. Indicates the period value, Indicates that given at time step The value estimated by the critic network in the state Indicates that given at time step The value estimated by the critic network in the state represents the total length of the time step, represents the mean square error function.

5. The intelligent taxi dispatching method based on dual attention mechanism according to claim 4 is characterized in that: In the process of obtaining actions based on the local supply and demand status through the local attention mechanism, the following sub-steps are also included: Based on the parameters of the policy network and the local observations to generate action distribution, the corresponding process has the following relationship: ； in, represents the probability of the actor network choosing an action under local observation, Indicates the actions that the taxi can choose; Through the taxi's strategy function, we can get the taxi's action, and the relationship corresponding to this process is: ； in, Indicates that the taxi is at time step Select the action, represents the taxi’s policy function, Indicates that the taxi is at time step Status at the time; Using the actions and status of the taxi, we can get the value function optimization strategy estimated by the critic network. The corresponding relationship is: ； in, represents the taxi’s policy parameters, Indicates that it has been processed by the value calculation function. Represents the weight of the function used by each taxi when making decisions.

6. The intelligent taxi dispatching method based on dual attention mechanism according to claim 5 is characterized in that: In step 3, a global reward is generated through an environment simulator according to the action, and the specific steps are as follows: Define the balance indicator, the definition is: ； in, Represents the balance indicator, Indicates that at time step The number of passenger demands in the area below, Indicates that at time step The number of idle vehicles in the area below; The local reward is defined by the balance indicator, which is defined as: ； in, Indicates partial reward; The global reward is obtained based on the local reward, and the corresponding process has the following relationship: ； in, Represents the global reward.

7. The intelligent taxi dispatching method based on dual attention mechanism according to claim 6 is characterized in that: In step 3, based on the advantage function and in combination with the feedback information of the environment, the agent is trained to obtain a trained agent, and the optimal scheduling strategy is obtained through the trained agent. The process also includes the following sub-steps: The function of optimizing strategy parameters is constructed through the advantage function, and the relationship between the corresponding process is: ； in, represents the advantage function, represents the function used to optimize the strategy parameters, represents the parameters of the actor network, It means to solve the expected joint distribution of state and action under the current strategy. Indicates the status under the current strategy. Represents the action under the current strategy; In the process of advantage function calculation, it is enhanced by generalized advantage estimation, and the relationship between the corresponding process is: ； in, represents the advantage error, Indicates that at time step Instant rewards when represents the advantage error at the next time step, represents the index of the time step; During the training process, the loss function of the actor network is constructed based on the advantage function, and the relationship between the corresponding process is: ； in, represents the loss function of the actor network, Represents the operation of taking the expected value of all samples at multiple time steps, represents the pruning hyperparameter, Indicates that a shear operation has been performed; represents the probability of taking an action under the same local observation based on the old strategy; The parameters of the actor network are updated through the loss function of the actor network, and the relationship between the corresponding process is: ； in, represents the learning rate, represents the gradient; The total loss function expression is: ； in, represents the loss function of the critic network, represents the entropy term, represents the weight hyperparameter of the loss function of the critic network, Represents the weight hyperparameter of the entropy term.

8. The intelligent taxi dispatching method based on dual attention mechanism according to claim 7 is characterized in that: In step 4, the optimal scheduling strategy is input into the execution layer to generate execution actions. The corresponding process has the following relationship: ； in, Represents the specific actions sampled from the optimal scheduling strategy.

9. The intelligent taxi dispatching method based on dual attention mechanism according to claim 8 is characterized in that: In step 4, the strategy is updated by executing actions, and the relationship between the corresponding process is: ； in, Represents the updated policy parameters.

10. An intelligent taxi dispatching system based on a dual attention mechanism, characterized in that: The system applies the intelligent taxi dispatching method based on the dual attention mechanism as described in any one of claims 1 to 9 above, and the system comprises: Preprocessing module for: Use intelligent agents to divide the urban transportation network into hexagonal grid cells; Through the adjacency matrix, the traffic environment data is dynamically converted into a traffic environment state vector by combining the local state of each grid unit and the state of the adjacent grids; Dual attention prediction module for: Based on the traffic environment state vector, the dual attention mechanism is used to predict the local supply and demand state and the global supply and demand state. Training modules for: The global supply and demand status is input into the decision layer for processing, and combined with the critic network to obtain the global value; According to the local supply and demand status, actions are obtained through the local attention mechanism; Based on the action, a global reward is generated through the environment simulator, the state and position of the taxi are updated in the environment based on the global reward, and feedback information from the environment is obtained; Based on the advantage function and combined with the feedback information of the environment, the intelligent agent is trained to obtain the trained intelligent agent, and the optimal scheduling strategy is obtained through the trained intelligent agent; Policy update module, used to: Input the optimal scheduling strategy into the execution layer to generate execution actions; By executing actions, the policy is updated.

Images