Dynamic task distribution and settlement method and system based on self-operation platform
By using a dynamic task distribution and settlement method on a self-operated platform, and leveraging long short-term memory networks and deep reinforcement learning models, the problem of low order allocation efficiency in multimodal transport systems has been solved, enabling unified management and efficient cost settlement for various logistics systems.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-12
- Publication Date
- 2026-04-03
AI Technical Summary
Existing technologies are insufficient to effectively connect multimodal transport systems and plan cross-transport process control, resulting in low efficiency in transport order allocation.
A dynamic task distribution and settlement method based on a self-operated platform is adopted. By collecting the transportation capacity and weather information of the logistics system through the self-operated platform, and using long short-term memory network and deep reinforcement learning model, orders are dynamically allocated and settlement is carried out to achieve unified management of multiple logistics systems.
It enables order monitoring and management across multiple logistics systems (sea, land, and air) and cross-system fee settlement, greatly improving the efficiency of customers' logistics resource matching.
Smart Images
Figure CN121787754A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of smart logistics technology, and more specifically, to a method and system for dynamic task distribution and settlement based on a self-operated platform. Background Technology
[0002] To build an ecosystem for multimodal transport integrating road, rail, and sea, implement a "single bill of lading" system for multimodal transport, construct a green and efficient transportation system, improve the national railway, highway, and waterway network, and vigorously promote the development of a "single bill of lading" system for multimodal transport. Unlike the traditional allocation of transport orders and control of transport processes for a single mode, a single transport order spanning multiple transport systems not only requires connecting the platforms of each transport system, but also requires planning the transport process control issues across transport systems. Summary of the Invention
[0003] This invention provides a method and system for dynamic task distribution and settlement based on a self-operated platform, solving the technical problems in related technologies.
[0004] This invention provides a dynamic task distribution and settlement method based on a self-operated platform, comprising the following steps:
[0005] Step 100: Collect partial or all of the capacity information of all logistics systems associated with the self-operated platform and partial or all of the weather information of the area where the logistics systems are located at each scheduling time; encode the information collected at each scheduling time to obtain an observation vector.
[0006] Step 200: Record the scheduling information generated at each scheduling moment, and encode the scheduling information recorded at each scheduling moment to obtain a scheduling vector;
[0007] Step 300: Define the current time as time v, and combine the observation vector and scheduling vector from time vn to time v into a comprehensive feature sequence. The i-th sequence unit of the comprehensive feature sequence includes the observation vector and scheduling vector at time z, where z = v-n + i-1.
[0008] Step 400: Input the comprehensive feature sequence into the dynamic task model, and the dynamic task model outputs scheduling information.
[0009] Step 500: Based on the scheduling information, the orders are distributed to the corresponding logistics systems, and then the logistics systems settle the logistics costs for the distributed orders.
[0010] Furthermore, if tn < 1, then the missing sequence units in the comprehensive feature sequence are filled in, and the filled sequence units are zero vectors or vectors with all components being 1.
[0011] Furthermore, the scheduling information includes information on the transportation vehicles in the logistics system to which all ongoing orders are assigned, and the logistics transportation allocation for all ongoing orders is completed through the scheduling information.
[0012] Furthermore, the dynamic task model includes:
[0013] Advanced manager, based on historical observation sequence. 0：t and scheduling sequence a 0：t-1 Estimate the posterior probability distribution b of the state of the environment at the current time t. t The formula for calculating (s) is as follows:
[0014]
[0015] b t (s) is the posterior probability distribution of the state at time t;
[0016] P(o t |s): is the observation model, given state s, the probability of observing observation Ot at the current time t;
[0017] P(s|s′,a t-1 This is a state transition model, in which scheduling information a is executed. t-1 Then, the probability of transitioning from state s′ to state s;
[0018] b t-1 (s′) is the posterior distribution of the state at the previous time t-1;
[0019] The calculation formula for the low-level controller is as follows:
[0020]
[0021] V(b t (s), a) represents the value function, a t This represents the scheduling information at time t.
[0022] Furthermore, the posterior probability distribution of the state at time t is estimated by modeling using a first neural network. This first neural network includes a forget gate, an input gate, cell state parameters, and an output gate. The output f of the forget gate... t The calculation formula is:
[0023] f t =σ(W f ·[h t-1 x t ]+b f )
[0024] Among them, W f This is the weight matrix of the forget gate, b f It is the bias term of the forget gate, [ht-1 x t [] is the hidden state h from the previous time step. t-1 and the observation vector x at the current time step t The concatenation, where σ is the Sigmoid activation function, n≥t≥1;
[0025] x t It is equal to the first data vector at the p-th sampling time point, and the current time is defined as v;
[0026] p = V - n + t - 1
[0027] t represents the current time step of the Long Short-Term Memory network. The Long Short-Term Memory network takes an observation vector as input at each time step.
[0028] The output i of the input gate t With candidate cell state The calculation formula is:
[0029] i t =σ(W i ·[h t-1 x t ]+b i )
[0030]
[0031] Among them, W i It is the weight matrix of the input gate, b i W is the bias term of the input gate, tanh is the hyperbolic tangent activation function, and W is the bias term of the input gate. C It is the weight matrix for calculating the candidate cell states, b C It is a bias term for the candidate cell state;
[0032] Cell state C t It is from the previous cell state C t-1 The weighted synthesis based on candidate cell states is calculated using the following formula:
[0033]
[0034] Where * represents element-wise multiplication;
[0035] The activation value of the output gate o t The hidden state h at the current time step t The calculation formula is:
[0036] o t =σ(W o ·[h t-1 x t ]+b o )
[0037] h t =o t *tan h(C t )
[0038] Among them, W o It is the weight matrix of the output gate, b o It is the bias term of the output gate;
[0039] The hidden state h at the last time step n The process involves calculating the posterior probability distribution vector Y using the fully connected layer. The calculation formula is as follows:
[0040] Y = W U h n +b Y
[0041] Among them, W U It is the weight matrix of the fully connected layer, b U It is the bias vector.
[0042] Furthermore, the calculation formula for the second neural network is as follows:
[0043] H (0) =x (1)
[0044] u (L) =σ(W u X (L) +W u H (L-1) +b u )
[0045] r (L) =σ(W r X (L) +W r H (L-1) +b r )
[0046] C (L) =tanh(W c X (L) +W c r (L) ⊙H (L-1) +b c )
[0047] H (L) =(1-u (L) )⊙C (L) +u (L) ⊙H (L-1)
[0048] OUT two =σ(W two *H(L) +b two )
[0049] Where, x (1) and X (L) Let W represent the posterior probability distribution vectors of the 1st and Lth states of the input, respectively, where n ≥ L ≥ 1. u W r W c and W two Both represent trainable weight parameters, b u b r b c and b two Both represent trainable bias parameters, ⊙ represents the dot product, and u (L) Let r represent the Lth first intermediate feature. (L) Let C represent the Lth second intermediate feature. (L) H represents the Lth third intermediate feature. (L-1) and H (L) OUT represents the (L-1)th and Lth output features, respectively. two Let represent the second output vector, where the value of the i-th component of the second output vector represents the Q-value of the i-th action, tanh represents the tanh function, and σ represents the Sigmoid function.
[0050] Let the current time be v. The first state posterior probability distribution vector input to the second neural model is the state posterior probability distribution vector at time v-n+1.
[0051] Furthermore, the training steps for the first and second neural networks include:
[0052] Step 101: Model the Q-network, which is a combination of the first neural network and the second neural network;
[0053] Copy the Q network to obtain the targetQ network;
[0054] Randomly obtain one experience point from the experience pool;
[0055]
[0056] S represents t+1 The maximum Q-value output by the targetQ network;
[0057] according to and Calculate the loss and update the Q network;
[0058]
[0059] State St And the first action A t The output Q value of the lower Q network, State S t And the first action A t The output Q value of the target network;
[0060] γ is the discount factor, a value between 0 and 1, used to balance the weights of immediate and future rewards; the default value is 0.6.
[0061] Step 102: Every fixed number of network updates, update the targetQ network so that its parameters are the same as the parameters of the current Q network;
[0062] The default value for the fixed interval of network updates is 10.
[0063] Step 103 continues until the Q-network converges or the set number of iterations is reached, at which point the process terminates.
[0064] Furthermore, a logistics system refers to logistics organizations that do not belong to the same system, including rail logistics systems, freight rail logistics systems, river logistics systems, ocean logistics systems, and air logistics systems.
[0065] A dynamic task distribution and settlement system based on a self-operated platform, used to execute the steps in the aforementioned dynamic task distribution and settlement method based on a self-operated platform, includes:
[0066] The information acquisition module collects capacity information and weather information for the region where the logistics system is located from all logistics systems associated with the self-operated platform at each scheduling time; and encodes the information collected at each scheduling time to obtain an observation vector.
[0067] The scheduling information recording module records the scheduling information generated at each scheduling moment, and encodes the scheduling information recorded at each scheduling moment to obtain a scheduling vector;
[0068] The sequence generation module defines the current time as time t, and combines the observation vector and scheduling vector from time tn to time t into a comprehensive feature sequence. The i-th sequence unit of the comprehensive feature sequence includes the observation vector and scheduling vector at time z.
[0069] Model memory, which is used to store dynamic task models;
[0070] The executor inputs the comprehensive feature sequence into the dynamic task model, and the dynamic task model outputs scheduling information.
[0071] The order settlement module distributes orders to the corresponding logistics system based on the scheduling information, and then the logistics system settles the logistics costs for the distributed orders.
[0072] The cost settlement module is used to receive logistics costs for orders from the logistics system, and then summarize the logistics costs for the same order to calculate the total cost of the order.
[0073] A storage medium storing non-transitory computer-readable instructions, which, when executed by a computer, are capable of performing the steps in the aforementioned dynamic task distribution and settlement method based on a self-operated platform.
[0074] The beneficial effects of this invention are as follows:
[0075] This invention requires only one system to monitor and manage orders across multiple logistics systems, including land, sea, and air, and to settle order fees across systems, greatly improving the efficiency of matching logistics resources for customers. Attached Figure Description
[0076] Figure 1 This is a flowchart of a dynamic task distribution and settlement method based on a self-operated platform according to the present invention. Detailed Implementation
[0077] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.
[0078] At least one embodiment of the present invention discloses a dynamic task distribution and settlement method based on a self-operated platform, such as... Figure 1 As shown, it includes the following steps:
[0079] Step 100: Collect partial or all of the capacity information of all logistics systems associated with the self-operated platform and partial or all of the weather information of the area where the logistics systems are located at each scheduling time; encode the information collected at each scheduling time to obtain an observation vector.
[0080] The time interval between adjacent scheduling times is fixed, with a default time interval of 4 hours.
[0081] This invention considers situations where it is impossible to fully observe all information about the environment, and only partial observations can be obtained. The actual condition of goods, weather changes, traffic congestion, and other factors are difficult to know completely. Therefore, the information collected is generally local information, but the possibility of collecting global information cannot be ruled out.
[0082] Even when obtaining partial information, it is still necessary to interpolate default values to fill in the missing information and ensure the consistency of the dimension of the encoded observation vector.
[0083] Step 200: Record the scheduling information generated at each scheduling moment, and encode the scheduling information recorded at each scheduling moment to obtain a scheduling vector;
[0084] It should be noted that the first scheduling moment is the startup moment. At this time, no scheduling information records are generated. Instead, a default scheduling vector is generated directly. The default scheduling vector can be a zero vector or a vector with all components being 1.
[0085] The method of information encoding is feature engineering, which uses the corresponding feature engineering method to encode and generate features for information of various modalities.
[0086] Step 300: Define the current time as time v, and combine the observation vector and scheduling vector from time vn to time v into a comprehensive feature sequence. The i-th sequence unit of the comprehensive feature sequence includes the observation vector and scheduling vector at time z, where z = v-n + i-1.
[0087] If tn < 1, then the missing sequence units in the comprehensive feature sequence are filled in. The filled sequence units are zero vectors or vectors with all components being 1.
[0088] Step 400: Input the comprehensive feature sequence into the dynamic task model, and the dynamic task model outputs scheduling information.
[0089] Dynamic task models include:
[0090] Advanced Manager: The role of the advanced manager is to estimate the abstract state or sub-goals of the environment;
[0091] Low-level controller: The low-level controller selects specific actions based on the state estimated by the high-level manager;
[0092] Advanced manager, based on historical observation sequence. 0：t and scheduling sequence a 0：t-1 Estimate the posterior probability distribution b of the state of the environment at the current time t. t The formula for calculating (s) is as follows:
[0093]
[0094] b t (s): is the posterior probability distribution of the state at time t;
[0095] P(o t |s): This is the observation model, given state s, the observation o at the current time t.t The probability of;
[0096] P(s|s′,a t-1 This is a state transition model, in which scheduling information a is executed. t-1 Then, the probability of transitioning from state s′ to state s;
[0097] b t-1 (s′) is the posterior distribution of the state at the previous time t-1;
[0098] The calculation formula for the low-level controller is as follows:
[0099]
[0100] V(b t (s), a) represents the value function, a t This represents the scheduling information at time t.
[0101] In one embodiment of the present invention, the posterior probability distribution of the state at time t is estimated by modeling a first neural network. The first neural network includes a forget gate, an input gate, a cell state, and an output gate. The output f of the forget gate is... t The calculation formula is:
[0102] f t =σ(W f ·[h t-1 x t ]+b f )
[0103] Among them, W f This is the weight matrix of the forget gate, b f It is the bias term of the forget gate, [h t-1 x t [] is the hidden state h from the previous time step. t-1 and the observation vector x at the current time step t The concatenation, where σ is the Sigmoid activation function, n≥t≥1;
[0104] x t It is equal to the first data vector at the p-th sampling time point, and the current time is defined as v;
[0105] p = V - n + t - 1
[0106] t represents the current time step of the Long Short-Term Memory network. The Long Short-Term Memory network takes an observation vector as input at each time step.
[0107] The output i of the input gate t With candidate cell state The calculation formula is:
[0108] it =σ(W i ·[h t-1 x t ]+b i )
[0109]
[0110] Among them, W i It is the weight matrix of the input gate, b i W is the bias term of the input gate, tanh is the hyperbolic tangent activation function, and W is the bias term of the input gate. C It is the weight matrix for calculating the candidate cell states, b C It is a bias term for the candidate cell state;
[0111] Cell state C t It is from the previous cell state C t-1 The weighted synthesis based on candidate cell states is calculated using the following formula:
[0112]
[0113] Where * represents element-wise multiplication;
[0114] The activation value of the output gate o t The hidden state h at the current time step t The calculation formula is:
[0115] o t =σ(W o ·[h t-1 x t ]+b o )
[0116] h t =o t *tan h(C t )
[0117] Among them, W o It is the weight matrix of the output gate, b o It is the bias term of the output gate;
[0118] The hidden state h at the last time step n The process involves calculating the posterior probability distribution vector Y using the fully connected layer. The calculation formula is as follows:
[0119] Y = W U h n +b U
[0120] Among them, W U It is the weight matrix of the fully connected layer, b Y It is the bias vector.
[0121] In one embodiment of the present invention, the calculation formula for the second neural network is as follows:
[0122] H (0) =x (1)
[0123] u (L) =σ(W u X (L) +W u H (L-1) +b u )
[0124] r (L) =σ(W r X (L) +W r H (L-1) +b r )
[0125] C (L) =tanh(W c X (L) +W c r (L) ⊙H (L-1) +b c )
[0126] H (L) =(1-u (L) )⊙C (L) +u (L) ⊙H (L-1)
[0127] OUT two =σ(W two *H (L) +b two )
[0128] Where, x (1) and X (L) Let W represent the posterior probability distribution vectors of the 1st and Lth states of the input, respectively, where n ≥ L ≥ 1. u W r W c and W two Both represent trainable weight parameters, b u b r b c and b two Both represent trainable bias parameters, ⊙ represents the dot product, and u (L) Let r represent the Lth first intermediate feature. (L) Let C represent the Lth second intermediate feature. (L) H represents the Lth third intermediate feature. (L-1) and H (L)OUT represents the (L-1)th and Lth output features, respectively. two Let represent the second output vector, where the value of the i-th component of the second output vector represents the Q-value of the i-th action, tanh represents the tanh function, and σ represents the Sigmoid function.
[0129] Let the current time be v, and the first state posterior probability distribution vector input to the second neural model is the state posterior probability distribution vector at time v-n+1.
[0130] The action space contains all possible actions that can be performed, and an action represents the means of transport assigned to each order.
[0131] The order identifier is an ID, and the action sequence contains all the transportation vehicles corresponding to the IDs. If the number of orders is less than the total number of IDs, it will not affect the generation of the action.
[0132] In one embodiment of the present invention, a first neural network and a second neural network are trained using the DQN method, and the reward function for Q-learning is as follows:
[0133] R(s t a t ) = w1·transportation efficiency(a t -w2·Transportation Costs(s) t a t -w3·Delay Penalty (a t )
[0134] R(s t a t ) indicates that in state s t Perform action a t The resulting rewards;
[0135] Transportation efficiency (a) t This can be defined as the amount of transportation completed per unit time under this scheduling action, used to maximize overall throughput.
[0136] Transportation costs (s) t a t ) is in the current state s t Next, execute action a t The resulting transportation costs include route distance, energy consumption, etc.
[0137] Delay penalty (a t () is a penalty for delayed order delivery, used to ensure service quality;
[0138] w1, w2, and w3 are the corresponding weighting coefficients used to balance the importance of each objective.
[0139] By maximizing this reward function, the intelligent scheduler can simultaneously pursue higher transportation efficiency (increased throughput), lower transportation costs (energy savings), and smaller delay penalties (ensuring service quality), seeking the optimal balance among the three to achieve efficient multimodal transport scheduling.
[0140] In one embodiment of the present invention, the training step includes:
[0141] Step 101: Model the Q-network, which is a combination of the first neural network and the second neural network;
[0142] Copy the Q network to obtain the targetQ network;
[0143] Randomly obtain one experience point from the experience pool;
[0144]
[0145] S represents t+1 The maximum Q-value output by the targetQ network;
[0146] according to and Calculate the loss and update the Q network;
[0147]
[0148] State S t And the first action A t The output Q value of the lower Q network, State S t And the first action A t The output Q value of the target network;
[0149] γ is the discount factor, a value between 0 and 1, used to balance the weights of immediate and future rewards; the default value is 0.6.
[0150] Step 102: Every fixed number of network updates, update the targetQ network so that its parameters are the same as the parameters of the current Q network;
[0151] The default value for the fixed interval of network updates is 10.
[0152] Step 103 continues until the Q-network converges or the set number of iterations is reached, at which point the process terminates.
[0153] The aforementioned "done" refers to the completion of all transportation tasks, and the default value for the number of iterations is 1000.
[0154] In one embodiment of the present invention, the scheduling information includes information on an order being assigned to a transportation vehicle in a logistics system, and the logistics transportation allocation of the order can be completed through the scheduling information.
[0155] In one embodiment of the present invention, the scheduling information includes information on the transportation vehicles of the logistics system to which all ongoing orders are assigned, and the logistics transportation allocation of all ongoing orders can be completed through the scheduling information.
[0156] In one embodiment of the present invention, a logistics system refers to a logistics organization that does not belong to the same system, such as a train logistics system, a freight truck logistics system, a river logistics system, a marine logistics system, and an air logistics system.
[0157] Step 500: Based on the scheduling information, the orders are distributed to the corresponding logistics systems, and then the logistics systems settle the logistics costs for the distributed orders.
[0158] In at least one embodiment of the present invention, a dynamic task distribution and settlement system based on a self-operated platform is provided, comprising:
[0159] The information acquisition module collects capacity information and weather information for the region where the logistics system is located from all logistics systems associated with the self-operated platform at each scheduling time; and encodes the information collected at each scheduling time to obtain an observation vector.
[0160] The scheduling information recording module records the scheduling information generated at each scheduling moment, and encodes the scheduling information recorded at each scheduling moment to obtain a scheduling vector;
[0161] The sequence generation module defines the current time as time t, and combines the observation vector and scheduling vector from time tn to time t into a comprehensive feature sequence. The i-th sequence unit of the comprehensive feature sequence includes the observation vector and scheduling vector at time z.
[0162] Model memory, which is used to store dynamic task models;
[0163] The executor inputs the comprehensive feature sequence into the dynamic task model, and the dynamic task model outputs scheduling information.
[0164] The order settlement module distributes orders to the corresponding logistics system based on the scheduling information, and then the logistics system settles the logistics costs for the distributed orders.
[0165] The cost settlement module is used to receive logistics costs for orders from the logistics system, and then summarize the logistics costs for the same order to calculate the total cost of the order.
[0166] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.
Claims
1. A dynamic task distribution and settlement method based on a self-operated platform, characterized in that, Includes the following steps: Step 100: Collect partial or all of the capacity information of all logistics systems associated with the self-operated platform and partial or all of the weather information of the area where the logistics systems are located at each scheduling time; encode the information collected at each scheduling time to obtain an observation vector. Step 200: Record the scheduling information generated at each scheduling moment, and encode the scheduling information recorded at each scheduling moment to obtain a scheduling vector; Step 300: Define the current time as time v, and combine the observation vector and scheduling vector from time vn to time v into a comprehensive feature sequence. The i-th sequence unit of the comprehensive feature sequence includes the observation vector and scheduling vector at time z, where z = v-n + i-1. Step 400: Input the comprehensive feature sequence into the dynamic task model, and the dynamic task model outputs scheduling information. Step 500: Based on the scheduling information, the orders are distributed to the corresponding logistics systems, and then the logistics systems settle the logistics costs for the distributed orders.
2. The dynamic task distribution and settlement method based on a self-operated platform according to claim 1, characterized in that, If tn < 1, then the missing sequence units in the comprehensive feature sequence are filled in. The filled sequence units are zero vectors or vectors with all components being 1.
3. The dynamic task distribution and settlement method based on a self-operated platform according to claim 1, characterized in that, The scheduling information includes information on the transportation vehicles in the logistics system to which all orders in progress are assigned. The logistics transportation allocation for all orders in progress is completed through the scheduling information.
4. The dynamic task distribution and settlement method based on a self-operated platform according to claim 1, characterized in that, Dynamic task models include: Advanced manager, based on historical observation sequence. 0：t and scheduling sequence a 0：t-1 Estimate the posterior probability distribution b of the state of the environment at the current time t. t The formula for calculating (s) is as follows: b t (s) is the posterior probability distribution of the state at time t; P(O t |s) is the observation model, given state s, the observation o at current time t. t The probability of; P(s|s′,a t-1 This is a state transition model, in which scheduling information a is executed. t-1 Then, the probability of transitioning from state s′ to state s; b t-1 (s′) is the posterior distribution of the state at the previous time t-1; The calculation formula for the low-level controller is as follows: V(b t (s), a) represents the value function, a t This represents the scheduling information at time t.
5. The dynamic task distribution and settlement method based on a self-operated platform according to claim 1, characterized in that, The posterior probability distribution of the state at time t is estimated by modeling using a first neural network. This first neural network includes a forget gate, an input gate, cell states, and an output gate. The output f of the forget gate... t The calculation formula is: f t =σ(W f ·[h t-1 ,x t ]+b f ) Among them, W f This is the weight matrix of the forget gate, b f It is the bias term of the forget gate, [h t-1 X t [] is the hidden state h from the previous time step. t-1 and the observation vector x at the current time step t The concatenation, where σ is the Sigmoid activation function, n≥t≥1; x t It is equal to the first data vector at the p-th sampling time point, and the current time is defined as v; p = V - n + t - 1 t represents the current time step of the Long Short-Term Memory network. The Long Short-Term Memory network takes an observation vector as input at each time step. The output i of the input gate t With candidate cell state The calculation formula is: i t =σ(W i ·[h t-1 ,x t ]+b i ) Among them, W j It is the weight matrix of the input gate, b i W is the bias term of the input gate, tanh is the hyperbolic tangent activation function, and W is the bias term of the input gate. C It is the weight matrix for calculating the candidate cell states, b C It is a bias term for the candidate cell state; Cell state C t It is from the previous cell state C t-1 The weighted synthesis based on candidate cell states is calculated using the following formula: Where * represents element-wise multiplication; The activation value of the output gate o t The hidden state h at the current time step t The calculation formula is: the t =σ(W o ·[h t-1 ,x t ]+b o ) h t = no t *fish(C) t ) Among them, W o It is the weight matrix of the output gate, b o It is the bias term of the output gate; The hidden state h at the last time step n The process involves calculating the posterior probability distribution vector Y using the fully connected layer. The calculation formula is as follows: Y=W U h n +b U Among them, W U It is the weight matrix of the fully connected layer, b U It is the bias vector.
6. The dynamic task distribution and settlement method based on a self-operated platform according to claim 5, characterized in that, The calculation formula for the second neural network is as follows: H (0) =x (1 ) u (L) =σ(W u X (L) +W u H (L-1) +b u ) r (L) =σ(W r X (L) +W r H (L-1) +b r ) C (L) =tanh(W c X (L) +W c r (L) ⊙H (L-1) +b c ) H (L) =(1-u (L) )⊙C (L) +u (L) ⊙H (L-1) OUT two =σ(W two *H 9L) +b two ) Where, x (1) and X (L) Let W represent the posterior probability distribution vectors of the 1st and Lth states of the input, respectively, where n ≥ L ≥ 1. u W r W c and W two Both represent trainable weight parameters, b u b r b c and b two Both represent trainable bias parameters, ⊙ represents the dot product, and u (L) Let r represent the Lth first intermediate feature. (L) Let C represent the Lth second intermediate feature. (L) H represents the Lth third intermediate feature. (L -1) and H (L) OUT represents the (L-1)th and Lth output features, respectively. two Let represent the second output vector, where the value of the i-th component of the second output vector represents the Q-value of the i-th action, tanh represents the tanh function, and σ represents the Sigmoid function; Let the current time be v. The first state posterior probability distribution vector input to the second neural model is the state posterior probability distribution vector at time v-n+1.
7. The dynamic task distribution and settlement method based on a self-operated platform according to claim 6, characterized in that, The training steps for the first and second neural networks include: Step 101: Model the Q-network, which is a combination of the first neural network and the second neural network; Copy the Q network to obtain the targetQ network; Randomly obtain one experience point from the experience pool; S represents t+1 The maximum Q-value output by the targetQ network; according to and Calculate the loss and update the Q network; State S t And the first action A t The output Q value of the lower Q network, State S t And the first action A t The output Q value of the target network; γ is the discount factor, a value between 0 and 1, used to balance the weights of immediate and future rewards; the default value is 0.
6. Step 102: Every fixed number of network updates, update the targetQ network so that its parameters are the same as the parameters of the current Q network; Step 103 continues until the Q-network converges or the set number of iterations is reached, at which point the process terminates.
8. The dynamic task distribution and settlement method based on a self-operated platform according to claim 1, characterized in that, A logistics system refers to a logistics organization that does not belong to the same system, including rail logistics systems, freight rail logistics systems, river logistics systems, ocean logistics systems, and air logistics systems.
9. A dynamic task distribution and settlement system based on a self-operated platform, characterized in that, It is used to perform the steps in the dynamic task distribution and settlement method based on a self-operated platform as described in any one of claims 1-8, including: The information acquisition module collects capacity information and weather information for the region where the logistics system is located from all logistics systems associated with the self-operated platform at each scheduling time; and encodes the information collected at each scheduling time to obtain an observation vector. The scheduling information recording module records the scheduling information generated at each scheduling moment, and encodes the scheduling information recorded at each scheduling moment to obtain a scheduling vector; The sequence generation module defines the current time as time t, and combines the observation vector and scheduling vector from time tn to time t into a comprehensive feature sequence. The i-th sequence unit of the comprehensive feature sequence includes the observation vector and scheduling vector at time z. Model memory, which is used to store dynamic task models; The executor inputs the comprehensive feature sequence into the dynamic task model, and the dynamic task model outputs scheduling information. The order settlement module distributes orders to the corresponding logistics system based on the scheduling information, and then the logistics system settles the logistics costs for the distributed orders. The cost settlement module is used to receive logistics costs for orders from the logistics system, and then summarize the logistics costs for the same order to calculate the total cost of the order.
10. A storage medium storing non-transitory computer-readable instructions, characterized in that, When non-transitory computer-readable instructions are executed by a computer, the steps in a dynamic task distribution and settlement method based on a self-operated platform as described in any one of claims 1-8 can be performed.