Logistics scheduling method and device based on deep learning and adaptive belief propagation

Abstract

The invention discloses a logistics scheduling method and device based on deep learning and adaptive belief propagation, and is used for solving the technical problem that the quality of a finally output logistics scheduling scheme is poor due to the fact that an existing logistics scheduling method is liable to fall into a local optimal solution. The method comprises the steps of obtaining a logistics scheduling instance to be solved and converting the logistics scheduling instance into a logistics scheduling factor graph; based on a self-adaptive message passing mechanism, a graph attention network is adopted to generate dynamic parameters according to the factor graph, and target messages between variable nodes and neighbor factor nodes are output; weighting and aggregating the target message to obtain a belief value of each variable node, wherein each variable node contains a plurality of logistics scheduling execution actions and each action corresponds to the belief value; and selecting an action corresponding to the minimum belief value of each variable node to form a target logistics scheduling scheme.

Description

Technical Field

[0001] This invention relates to the field of logistics and distribution technology, and in particular to a logistics scheduling method and apparatus based on deep learning and adaptive belief propagation. Background Technology

[0002] Urban delivery, as a core link at the end of the supply chain, directly connects warehouses and end customers, and is a key support for ensuring the efficient operation of e-commerce retail, fresh food cold chain, and on-demand delivery. With consumer demand upgrading towards "instant and personalized" delivery, the daily volume of delivery orders within cities is experiencing explosive growth. Relying solely on manual planning or simple scheduling tools is no longer sufficient to meet the demands of large-scale delivery. Against this backdrop, the Vehicle Routing Problem with Time Windows (VRPTW) has become a core optimization direction for urban delivery scheduling. Its core scenario involves vehicles departing from fixed warehouses to deliver goods to customers scattered throughout the city, while simultaneously considering multi-dimensional constraints and cost optimization. This directly determines delivery efficiency, operating costs, and customer satisfaction, and is a critical problem that urgently needs to be solved efficiently in the modern urban logistics system.

[0003] VRPTW presents significant complexity and constraint relationships in urban delivery scenarios: From a constraint perspective, vehicles must strictly adhere to the customer-specified time window to avoid early delivery leading to unclaimed customers or late arrivals triggering penalties; simultaneously, they must meet hard constraints such as vehicle load limits, driver daily working hour limits, urban road traffic rules, and single-customer work hours. From an optimization objective perspective, while satisfying all feasibility constraints, it is necessary to minimize the total driving distance, reduce late arrival penalties, or optimize the number of vehicles, forming a complex optimization model of "multi-constraints and multi-objectives"; moreover, in reality, order volumes often reach tens to hundreds of orders, coupled with real-time road conditions and dynamic demands, further increasing the difficulty of developing scheduling solutions.

[0004] Existing logistics scheduling methods mostly use fixed parameters or static programming logic, but they cannot dynamically adjust the vehicle driving rhythm to cope with road condition fluctuations and operation time deviations. Furthermore, due to the lack of sufficient balance of multiple constraints such as load, working hours, and time windows, this method is prone to getting trapped in local optima, resulting in poor quality of the final output logistics scheduling solution. Summary of the Invention

[0005] This invention provides a logistics scheduling method and apparatus based on deep learning and adaptive belief propagation, which solves the technical problem that existing logistics scheduling methods are prone to getting trapped in local optima, resulting in poor quality of the final output logistics scheduling scheme.

[0006] The first aspect of this invention provides a logistics scheduling method based on deep learning and adaptive belief propagation, comprising:

[0007] Obtain the logistics scheduling instance to be solved, and transform the logistics scheduling instance to be solved to generate a logistics scheduling factor graph;

[0008] Based on the adaptive message passing mechanism, a graph attention network is used to generate dynamic parameters according to the logistics scheduling factor graph, and output the target message between each variable node and each neighbor factor node in the logistics scheduling factor graph.

[0009] The target messages between each variable node and each neighbor factor node are weighted and aggregated to output the belief value corresponding to each variable node. The variable node includes multiple logistics scheduling execution actions, and each logistics scheduling execution action has a corresponding belief value.

[0010] In each of the belief values ​​associated with each logistics scheduling execution action corresponding to each variable node, the logistics scheduling execution action corresponding to the smallest belief value is selected as the target logistics scheduling execution action, and a target logistics scheduling scheme is formed based on multiple target logistics scheduling execution actions.

[0011] Optionally, the adaptive message passing mechanism, employing a graph attention network to dynamically generate parameters based on the logistics scheduling factor graph, outputs the target message between each variable node and its neighboring factor nodes in the logistics scheduling factor graph, including:

[0012] Initialize the message of each variable node in the logistics scheduling factor graph, and output the initial message between each variable node and each of its neighboring factor nodes;

[0013] The current iteration number is counted in real time, and the initial messages between each variable node and each neighbor factor node, as well as the logistics scheduling factor graph, are used as inputs to the graph attention network to generate edge-specific damping factors and neighbor-specific weights between each variable node and each neighbor factor node at each time step corresponding to the current iteration number.

[0014] The adaptive message passing mechanism is used to calculate the intermediate message between each variable node and each neighbor factor node based on the edge-specific damping factor and the neighbor-specific weight between each variable node and each neighbor factor node at each time step.

[0015] Determine whether the current iteration count has reached a preset threshold.

[0016] If so, the intermediate message between each variable node and each of the neighboring factor nodes is taken as the corresponding target message.

[0017] Optionally, it also includes:

[0018] If the current iteration count does not reach the preset threshold, then the intermediate message between each variable node and each of the neighboring factor nodes is used as the corresponding new initial message;

[0019] Jump to execute the real-time statistics of the current iteration number, and take the initial message between each variable node and each neighbor factor node, as well as the logistics scheduling factor graph, as the input of the graph attention network to generate the edge-specific damping factor and neighbor-specific weight between each variable node and each neighbor factor node at each time step corresponding to the current iteration number;

[0020] Until the current iteration count reaches the preset threshold, the intermediate messages between each variable node and each neighbor factor node determined when the current iteration count reaches the preset threshold are taken as the corresponding target messages.

[0021] Optionally, the logistics scheduling factor diagram is specifically as follows:

[0022] ;

[0023] in, A diagram of logistics scheduling factors; For a set of variable nodes; It is the set of neighbor factor nodes, i.e., the set of constraint function nodes; Let be the set of edges, representing the relationships between variables and factors.

[0024] Optionally, the adaptive message passing mechanism specifically includes:

[0025] ;

[0026] in, At time step t, the new message that variable node i sends to its neighbor factor node a, which is the intermediate message between variable node i and its neighbor factor node a; At the t-th time step, the edge-specific damping factor of the edge from variable node i to neighbor factor node a, that is, the edge-specific damping factor between variable node i and neighbor factor node a; At the (t-1)th time step, the message sent by variable node i to neighbor factor node a represents the initial message between variable node i and neighbor factor node a. In the set of neighbor factors of variable node i, except Other neighbor factor nodes; Let i be the neighboring factor node of the variable node i that is currently sending a message; Let i be the set of neighboring factors of variable node i; For the t-th time step, the neighbor-specific weights between variable node i and its neighbor factor node a; At time step t-1, the message sent by neighbor factor node b to variable node i represents the initial message between neighbor factor node b and variable node i.

[0027] A second aspect of the present invention provides a logistics scheduling device based on deep learning and adaptive belief propagation, comprising:

[0028] The acquisition module is used to acquire the logistics scheduling instance to be solved, and to transform the logistics scheduling instance to be solved to generate a logistics scheduling factor graph.

[0029] The output module is used to generate dynamic parameters based on the logistics scheduling factor graph using an adaptive message passing mechanism and a graph attention network, and output the target message between each variable node and each neighbor factor node in the logistics scheduling factor graph.

[0030] The aggregation module is used to perform weighted aggregation of the target messages between each variable node and each neighbor factor node, and output the belief value corresponding to each variable node. The variable node includes multiple logistics scheduling execution actions, and each logistics scheduling execution action has a corresponding belief value.

[0031] The constitutive module is used to select the logistics scheduling execution action corresponding to the smallest belief value from the belief values ​​associated with each logistics scheduling execution action corresponding to each variable node as the target logistics scheduling execution action, and to constitute a target logistics scheduling scheme based on multiple target logistics scheduling execution actions.

[0032] Optionally, the adaptive message passing mechanism specifically includes:

[0033] ;

[0034] in, At time step t, the new message that variable node i sends to its neighbor factor node a, which is the intermediate message between variable node i and its neighbor factor node a; At the t-th time step, the edge-specific damping factor of the edge from variable node i to neighbor factor node a, that is, the edge-specific damping factor between variable node i and neighbor factor node a; At the (t-1)th time step, the message sent by variable node i to neighbor factor node a represents the initial message between variable node i and neighbor factor node a. In the set of neighbor factors of variable node i, except Other neighbor factor nodes; Let i be the neighboring factor node of the variable node i that is currently sending a message; Let i be the set of neighboring factors of variable node i; For the t-th time step, the neighbor-specific weights between variable node i and its neighbor factor node a; At time step t-1, the message sent by neighbor factor node b to variable node i represents the initial message between neighbor factor node b and variable node i.

[0035] A computer device provided in a third aspect of the present invention includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the logistics scheduling method based on deep learning and adaptive belief propagation as described in any of the preceding claims.

[0036] The fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed, implements the steps of the logistics scheduling method based on deep learning and adaptive belief propagation as described in any of the preceding claims.

[0037] The fifth aspect of the present invention provides a computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, wherein, when the program instructions are executed by a computer, the computer performs the steps of the logistics scheduling method based on deep learning and adaptive belief propagation as described in any of the preceding claims.

[0038] As can be seen from the above technical solutions, the present invention has the following advantages:

[0039] The above-mentioned technical solution of the present invention provides a logistics scheduling method based on deep learning and adaptive belief propagation. It obtains a logistics scheduling instance to be solved and transforms it to generate a logistics scheduling factor graph. Based on an adaptive message passing mechanism, a graph attention network is used to dynamically generate parameters according to the logistics scheduling factor graph, outputting the target message between each variable node and each neighboring factor node in the logistics scheduling factor graph. The target messages between each variable node and each neighboring factor node are weighted and aggregated to output the belief value corresponding to each variable node. Each variable node includes multiple logistics scheduling execution actions, and each logistics scheduling execution action has a corresponding belief value. From the belief values ​​associated with each logistics scheduling execution action corresponding to each variable node, the logistics scheduling execution action corresponding to the smallest belief value is selected as the target logistics scheduling execution action. Based on multiple target logistics scheduling execution actions, a target logistics scheduling scheme is constructed. Based on the above scheme, the present invention leverages the dynamic adaptability of the graph attention network to the constraints and dynamic changes of the scheduling scenario, and the automatic adjustment of dynamic parameters by the adaptive message passing mechanism, thereby freeing message updates from the limitations of traditional fixed parameters and reducing reliance on manual parameter tuning, thus outputting a high-quality logistics scheduling scheme. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 The flowchart illustrates the steps of a logistics scheduling method based on deep learning and adaptive belief propagation, as provided in Embodiment 1 of the present invention.

[0042] Figure 2 A schematic diagram of a factor graph provided in Embodiment 1 of the present invention;

[0043] Figure 3 This is a schematic diagram illustrating the use of an adaptive message passing mechanism to calculate a new message, as provided in Embodiment 1 of the present invention.

[0044] Figure 4 This is a general framework diagram of the centralized, non-autoregressive parameter joint inference mechanism provided in Embodiment 1 of the present invention;

[0045] Figure 5 This is a structural block diagram of a logistics scheduling device based on deep learning and adaptive belief propagation, provided in Embodiment 2 of the present invention. Detailed Implementation

[0046] This invention provides a logistics scheduling method and apparatus based on deep learning and adaptive belief propagation, which solves the technical problem that existing logistics scheduling methods are prone to getting trapped in local optima, resulting in poor quality of the final output logistics scheduling scheme.

[0047] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0048] Terminology Explanation:

[0049] Constraint Optimization Problem (COP): Under the premise of satisfying several equality or inequality constraints, the problem seeks to assign values ​​to variables so that the objective function reaches its minimum (or maximum) value. It is widely used in operations research, chip design, scheduling planning and other fields.

[0050] Factor graph: A bipartite graph structure where circular nodes represent variables and square nodes represent constraint functions. Edges connecting nodes indicate that the variable appears within the scope of the constraint. Factor graphs can intuitively represent the multiplicative solutions of complex functions and are the basic framework for message passing algorithms.

[0051] Belief Propagation (BP), also known as the sum-product algorithm, is an iterative message-passing algorithm for probabilistic reasoning and constraint optimization on a factor graph. Variable nodes and function nodes update their local beliefs by sending messages to each other, ultimately calculating the optimal value for each variable.

[0052] Damping Factor: A coefficient introduced in the traditional backpropagation (BP) algorithm to mix the messages from the previous round with the current calculation message, smoothing the iteration process and mitigating oscillations. The damping factor is usually a manually set global constant, making it difficult to automatically adapt to different problems.

[0053] Attention mechanism: A mechanism in deep learning that assigns different weights based on the importance of the input content. In graph neural networks, it can be used to learn the different influences of neighboring nodes to achieve dynamic aggregation.

[0054] Graph Attention Network (GAT): A neural network for processing graph data. It uses an attention mechanism to assign attention weights to each node's neighbors, enabling flexible feature aggregation. This invention employs a multi-layer GAT to encode the factor graph.

[0055] Figure 1 A flowchart illustrating the steps of a logistics scheduling method based on deep learning and adaptive belief propagation provided in Embodiment 1 of the present invention.

[0056] This invention provides a logistics scheduling method based on deep learning and adaptive belief propagation, comprising:

[0057] Step 101: Obtain the logistics scheduling instance to be solved, and transform the logistics scheduling instance to be solved to generate a logistics scheduling factor graph.

[0058] The logistics scheduling instance to be solved is a specific logistics delivery task scenario, which includes all the basic information required to complete the task (such as warehouse, customer, vehicle, constraints, etc.).

[0059] The logistics scheduling factor graph is a graph structure used to model logistics scheduling problems. It consists of variable nodes, factor nodes (neighboring factor nodes), and edges, and is used to intuitively express the relationship between scheduling decisions and constraints. Among them, variable nodes are nodes in the factor graph that represent scheduling decision variables that need to be optimized, corresponding to specific scheduling action parameters (such as departure time, travel time, etc.); factor nodes are nodes in the factor graph that represent constraints that need to be satisfied, corresponding to rules or restrictions in the scheduling process (such as time windows, load, etc.); edges are line segments connecting variable nodes and factor nodes, used to indicate that a certain decision variable is directly affected by a certain constraint.

[0060] It should be noted that obtaining the COP instance to be solved (i.e., the logistics scheduling instance to be solved) includes core elements such as warehouse location, customer location and demand (e.g., cargo weight, service time window), available vehicle information (e.g., load limit, driver working hours), and road traffic rules. Based on these elements, the transformation is completed by analyzing the decision variables and constraints: the scheduling decisions to be optimized (e.g., vehicle departure time, travel time between customers, service start time for each customer, etc.) are mapped to variable nodes in the logistics scheduling factor graph, and the constraints to be satisfied (e.g., customer time window limits, vehicle load constraints, driver working hour limits, path continuity requirements, etc.) are mapped to factor nodes. Then, edges are used to connect related variable nodes and factor nodes (e.g., connecting the "customer A service start time" variable node to the "customer A time window" factor node), ultimately generating the logistics scheduling factor graph. For example, Figure 2As shown, the factor graph accurately depicts the interactions between variables and constraints in the problem, providing a computational graph basis for subsequent message passing. It can be represented as:

[0061] ;

[0062] in, A diagram of logistics scheduling factors; For a set of variable nodes; It is the set of neighbor factor nodes, i.e., the set of constraint function nodes; Let be the set of edges, representing the relationships between variables and factors.

[0063] Figure 2 middle Different factor nodes represent constraints or functional relationships, used to characterize the association rules between variables (such as constraint equations, probability distributions, etc.). Different variable nodes represent decision variables that need to be optimized or inferred (such as departure time and travel time in logistics scheduling, or state variables and parameters in other fields), and are the core objects of information transmission and optimization.

[0064] Step 102: Based on the adaptive message passing mechanism, a graph attention network is used to generate dynamic parameters according to the logistics scheduling factor graph, and the target message between each variable node and each neighbor factor node in the logistics scheduling factor graph is output.

[0065] It should be noted that, based on the adaptive message passing mechanism, a graph attention network is used to dynamically generate parameters according to the logistics scheduling factor graph. Compared with the limitations of traditional scheduling methods that rely on fixed parameters, the graph attention network can adaptively capture the constraint relationships (such as the linkage between customer time windows and vehicle travel time) and dynamic changes (such as real-time road condition fluctuations and temporary order additions) in the logistics scheduling scenario. It automatically generates more adaptable edge-specific parameters (i.e., edge-specific damping factors and neighbor-specific weights between variable nodes and neighbor factor nodes), eliminating the need for repeated manual parameter testing. This significantly reduces reliance on manual intervention while providing accurate parameter support for message passing. In turn, it outputs the target message between each variable node and each neighbor factor node in the logistics scheduling factor graph. Due to the dynamic adaptation of parameters, this target message can accurately reflect the degree to which the variable node satisfies the constraints of the corresponding neighbor factor node (such as "vehicle departure time" satisfying "driver working hours constraint" and "morning peak traffic constraint"), avoiding the constraint satisfaction evaluation bias caused by traditional fixed parameters.

[0066] Specifically, step 102 may include the following sub-steps:

[0067] S21. Initialize the message of each variable node in the logistics scheduling factor graph and output the initial message between each variable node and each neighboring factor node.

[0068] S22. Real-time statistics of the current iteration number, and use the initial messages between each variable node and each neighbor factor node, as well as the logistics scheduling factor graph, as input to the graph attention network to generate the edge-specific damping factor and neighbor-specific weight between each variable node and each neighbor factor node at each time step corresponding to the current iteration number.

[0069] S23. An adaptive message passing mechanism is adopted to calculate the intermediate message between each variable node and each neighbor factor node based on the edge-specific damping factor and the neighbor-specific weight between each variable node and each neighbor factor node at each time step.

[0070] S24. Determine whether the current iteration count has reached the preset threshold.

[0071] S25. If so, then the intermediate message between each variable node and each neighboring factor node shall be taken as the corresponding target message.

[0072] Optionally, it also includes:

[0073] If the current iteration count has not reached the preset threshold, the intermediate message between each variable node and each neighboring factor node will be used as the corresponding new initial message.

[0074] Jump to execute real-time statistics of the current iteration number, and take the initial messages between each variable node and each neighbor factor node, as well as the logistics scheduling factor graph, as input to the graph attention network to generate the steps of edge-specific damping factor and neighbor-specific weight between each variable node and each neighbor factor node at each time step corresponding to the current iteration number;

[0075] Until the current iteration count reaches a preset threshold, the intermediate messages between each variable node and its neighboring factor nodes determined when the current iteration count reaches the preset threshold are taken as the corresponding target messages.

[0076] The preset threshold number of times can be set as needed, and this invention does not limit it.

[0077] It should be noted that, as Figure 3 As shown, to address the limitation of traditional DBP (Dynamic Bayesian Network) using a globally fixed damping factor for all messages, this invention proposes a fine-grained adaptive message update rule. In each BP iteration, when the variable node... To a certain neighbor factor node When sending a message, an adaptive message passing mechanism is used to calculate the new message between the variable node and its neighboring factor nodes:

[0078] ;

[0079] in, At time step t, the new message that variable node i sends to its neighbor factor node a, which is the intermediate message between variable node i and its neighbor factor node a; Let represent the edge-specific damping factor of the edge from variable node i to neighbor factor node a at time step t, which determines the retention of historical information in the current message. The ratio is used to finely control the smoothness of message updates and suppress oscillations; At the (t-1)th time step, the message sent by variable node i to neighbor factor node a represents the initial message between variable node i and neighbor factor node a. In the set of neighbor factors of variable node i, except Other neighbor factor nodes; Let i be the neighboring factor node of the variable node i that is currently sending a message; Let i be the set of neighboring factors of variable node i; For time step t, the neighbor-specific weights between variable node i and its neighbor factor node a. Used to characterize the construction of the giver factor The new message came from another neighbor factor. The importance weights of messages and the weights of each neighbor satisfy normalization. ; The message sent by neighbor factor node b to variable node i in the previous step represents the message sent by neighbor factor node b to variable node i at time step t-1, i.e., the initial message between neighbor factor node b and variable node i.

[0080] Figure 3 In At time step t, from the factor node , (Source factor node) passed to variable node And for the target factor node Neighbor-specific weights.

[0081] Through the above mechanism, this invention introduces two types of dynamically adjustable core parameters:

[0082] Variable - Side-specific damping factor : Used to control the mixing ratio of new and old messages when variable node i sends a message to neighbor factor node a. A larger value indicates that more historical information is retained, thus smoothing out drastic changes; Smaller values ​​emphasize the current calculated value, promoting exploration. Different variable-factor pairs and different iteration steps can each have their own... To adapt to local dynamics.

[0083] Neighbor-specific weights This is used to measure the importance of messages provided by different neighbor factor nodes b when variable node i updates the message sent to neighbor factor node a. Traditional backpropagation (BP) adds all neighbor messages equally, while this invention assigns a weight coefficient to each neighbor factor. This allows the algorithm to identify and highlight important constraint information. For example, it can assign higher weights to neighboring factors that are "more constrained" or have already stabilized, and reduce the weights of unstable or less influential factors, thereby intelligently integrating multi-source information and guiding the search towards a better direction.

[0084] It's worth mentioning that manually setting such a massive number of parameters for each variable, each edge, and each time step would be extremely difficult. , A task that is virtually impossible to accomplish. With a... Taking a factor graph with one variable and a maximum degree d running for T time steps as an example, theoretically, it needs to determine... The continuous selection of hyperparameter values—something traditional hyperparameter tuning methods cannot handle. Therefore, this invention introduces a deep neural network model. As an intelligent controller, namely a graph attention network, it is used to automatically infer the aforementioned adaptive parameters.

[0085] Among them, the deep neural network controller: model Taking the factor graph structure and backpropagation (BP) messages (i.e., the initial messages between each variable node and its neighboring factor nodes) as input, the network directly outputs all damping factors and neighbor weights required for each iteration, specifically the edge-specific damping factors and neighbor-specific weights between each variable node and its neighboring factor nodes at each time step corresponding to the current iteration number. In particular, the network input integrates information from both the global problem structure and the real-time message state, including:

[0086] (i) Structural information of the factor graph G (represented in a format acceptable to graph neural networks, such as an adjacency list or matrix).

[0087] (ii) The complete set of historical messages up to the current iteration, including both variable-to-factor message sequences and factor-to-variable message sequences.

[0088] By jointly encoding the factor graph structure and past message dynamics, the model is able to capture the pattern of BP message evolution over iteration.

[0089] network At the beginning of each iteration, the damping factors of all edges that need to be updated in that step are automatically inferred based on the above input. and all corresponding neighbor weights Subsequently, the BP algorithm uses these adaptive parameters to combine and transmit new messages, replacing the original manually set rules, thus completing this round of message passing. The model's inference and BP iterations alternate: with each BP iteration, the model updates its output parameters for the next step based on the new message state, and so on. Leveraging the powerful representation and generalization capabilities of deep learning, the BP message update process can achieve online adaptive adjustment based on problem characteristics and real-time status, realizing a paradigm shift from "manually designed heuristics" to "data-driven intelligent decision-making."

[0090] A potential problem with the aforementioned Deep Attentive Backpropagation (BP) framework is that it requires calling the neural network for inference at every time step. For inference processes with a large number of iterations, this stepwise autoregressive inference can lead to high time overhead and difficulty in gradient propagation (long sequences are prone to gradient vanishing / exploding). To address this, this invention introduces a series of improvement strategies in its implementation, enabling the algorithm to maintain its adaptive characteristics while significantly improving its operational efficiency and scalability.

[0091] Based on the above, this invention further proposes a centralized, non-autoregressive joint parameter inference mechanism, which reduces the call frequency of the deep learning model from once per step to once per round (several steps), significantly reducing computational and memory overhead. The overall framework of this mechanism is as follows: Figure 4 As shown, r represents the number of iterations.

[0092] Specifically, the algorithm employs a multi-independent iteration (restart) architecture, executing a total of R iterations, each of which can be considered a complete restart of the algorithm. By repeatedly searching from different random initial states, multiple iterations significantly reduce the risk of the algorithm getting trapped in local optima, improving the global quality and stability of the solution.

[0093] In the i-th iteration (i=1,2,…,R), the algorithm performs at most T steps of BP message passing and updating. Each step t (t=1,2,…,T) updates two types of messages: variable-to-factor messages and factor-to-variable messages, denoted by [symbols notation missing]. and The calculation of these messages depends on the message state of the previous step t−1 and the parameters deduced by this invention. , This invention employs centralized batch parameter inference: for the current i-th iteration, before starting the first step of message update, a deep neural network infers all the damping factors and neighbor weight sets required for the entire iteration in one go. , Then, each of the T iterations is executed in the order of the pre-inferred parameters.

[0094] The parameter inference is implemented using a Graph Attention Network (GAT). The input to the GAT model consists of two parts: (i) factor graph structure information, describing variable nodes, factor nodes, and their connections; and (ii) the complete message sequence of the previous iteration, i.e., the message update trajectory of all steps in the (i-1)th iteration (i.e., the initial messages between each variable node and its neighboring factor nodes). Here, it is assumed that the previous iteration has been completed and the corresponding final state message has been obtained (if i=1 indicates no previous iteration, a random initialization message or a zero message can be used as input). Through its powerful graph structure awareness and attention mechanism, GAT integrates the factor graph topology and past message dynamics to output a complete set of parameters guiding the message update at each step in the current iteration i. , .

[0095] This design is based on a crucial assumption: the message state after the final convergence of the previous iteration has sufficiently encoded the historical information of all previous steps and the current solution space of the problem, enough for the neural network to predict the optimization direction required for the next iteration. In other words, this invention believes that during multiple independent restarts of backpropagation (BP), it is not necessary to call the network at every iteration; calls are only needed between iterations. The neural network uses the message evolution trajectory of the previous complete iteration to infer the parameters for the next iteration, thereby "decoupling" the step-by-step inference process that was originally synchronized with BP and completing it in a batch manner. This improvement significantly reduces the calling frequency of deep models, greatly alleviating the time and memory burden of long sequence inference.

[0096] Therefore, this invention refers to the improved deep learning adaptive backpropagation algorithm as Fast Deep Belief Propagation (FDBP). FDBP has the following significant advantages over the original stepwise inference DABP algorithm:

[0097] Extremely high computational efficiency: By merging multiple model inferences in each BP iteration into a single iteration, the number of calls to the deep neural network is reduced from T to 1. This significant reduction in computational overhead means the algorithm can fully utilize the parallel capabilities of modern GPUs, achieving a remarkable speedup. Actual tests show that, with the same number of restarts, FDBP's average solution speed is approximately 2.87 times faster than the original algorithm.

[0098] Excellent memory efficiency: Centralized inference avoids storing all intermediate hidden states (e.g., the stepwise states of an RNN) within the backpropagation iteration, and GPU memory usage no longer increases linearly with the number of iterations. Furthermore, optimizations such as mixed precision further reduce GPU memory usage. Therefore, FDBP can handle larger factor graph instances under the same hardware conditions, solving the scalability problem of previous methods due to GPU memory exhaustion.

[0099] Guaranteed solution quality: Despite the change in parameter inference method, FDBP can still effectively learn the structural features and optimization trends of the problem by utilizing the complete information from the previous iteration. Experimental results show that FDBP achieves comparable, and in some cases even better, quality in the final solution compared to the original stepwise inference algorithm, and converges faster. In other words, it achieves higher efficiency while maintaining solution quality.

[0100] In summary, this invention proposes a highly efficient and scalable deep learning-enabled backpropagation (BP) algorithm architecture. Its core idea is to replace the inefficient stepwise autoregressive inference in the original algorithm with a centralized parameter inferr based on graph attention networks, achieving "batch processing" of parameter prediction. This architectural innovation fundamentally solves the computational and memory bottlenecks of the original deep BP algorithm, achieving several times the efficiency improvement while maintaining or even enhancing solution performance. This invention lays the foundation for the application of large-scale co-operation in real-world scenarios.

[0101] In this embodiment, if the current iteration count has not reached the preset threshold, the intermediate message between each variable node and each neighboring factor node is used as the corresponding new initial message, and the process jumps to step S22 until the current iteration count reaches the preset threshold. Then, the intermediate message between each variable node and each neighboring factor node determined when the current iteration count reaches the preset threshold is used as the corresponding target message.

[0102] Step 103: Perform weighted aggregation on the target messages between each variable node and each neighboring factor node, and output the belief value corresponding to each variable node. The variable node includes multiple logistics scheduling execution actions, and each logistics scheduling execution action has a corresponding belief value.

[0103] The target message is the final stable message between the variable node and its neighboring factor nodes after iterative updates via adaptive message passing, reflecting the degree to which the variable satisfies the corresponding constraints.

[0104] Weighted aggregation is an operation that sums or merges the target messages between variable nodes and their neighboring factor nodes according to the importance weight of each constraint, and is used to integrate information from multiple constraints.

[0105] The belief value is the weighted aggregation of the target messages of all neighboring factor nodes of the variable node, and is used to quantitatively evaluate the quality of the variable's actions (the smaller the value, the better the action).

[0106] The logistics scheduling execution action is the specific scheduling decision option corresponding to the variable node, such as selecting 6:00 or 7:00 for vehicle departure time, and selecting 0.5 hours or 1 hour for travel time between customers.

[0107] It should be noted that the target messages between each variable node and each neighboring factor node are weighted and summed according to the importance weight of each constraint (i.e., a weighted aggregation process), and the corresponding belief value for each variable node is output. This belief value comprehensively reflects the degree to which the variable node satisfies multi-dimensional constraints (such as time window, load, and working hours). Among them, the variable node corresponds to multiple logistics scheduling execution actions (for example, the vehicle departure time can be selected as 6:00 or 7:00, the travel time from customer A to B can be selected as 0.8 hours or 1.0 hours, and the unloading operation time can be selected as 0.2 hours or 0.3 hours, etc.). Since each execution action has different performance in satisfying constraints (for example, a departure at 6:00 is more in line with the traffic rules before the morning rush hour, but the driver's working hours are different), the corresponding belief value is generated through the weighted aggregation of target messages to quantify its advantages and disadvantages.

[0108] Step 104: Select the logistics scheduling action corresponding to the smallest belief value from the belief values ​​associated with each logistics scheduling execution action corresponding to each variable node as the target logistics scheduling execution action, and construct the target logistics scheduling scheme based on multiple target logistics scheduling execution actions.

[0109] The target logistics scheduling scheme is a complete scheduling plan composed of the target execution actions corresponding to the minimum belief value of each variable node, which satisfies all constraints and has the optimal overall cost.

[0110] It should be noted that among the belief values ​​associated with multiple logistics scheduling execution actions corresponding to each variable node, the logistics scheduling execution action corresponding to the smallest belief value is selected as the target execution action. Then, based on the target execution actions of all variable nodes, a target logistics scheduling scheme that satisfies multiple constraints and is cost-optimal can be constructed.

[0111] For comparison of technical effectiveness, existing technologies can be used as a reference. In urban delivery scenarios, vehicles need to depart from warehouses and complete services within the time windows specified by each customer, while being constrained by vehicle load, driver working hours, road conditions, and operation time. This task can be abstracted as a vehicle routing problem with time windows (VRPTW), aiming to optimize the overall cost, including total travel distance, lateness penalties, or number of vehicles, while satisfying feasibility constraints such as capacity and time windows. Due to the large order volume and the dynamic nature of road conditions and demand, common difficulties in reality include some customers arriving late or early, high total mileage, and insufficient adaptation to new orders and road events.

[0112] Furthermore, Constraint Optimization Problems (COPs) are a common type of complex problem, widely applied in fields such as artificial intelligence, operations research, chip design, and logistics scheduling. The goal of a COP is to find the solution that minimizes the objective function value from all possible combinations of variable assignments, while satisfying a set of constraints. Formally, a COP instance typically consists of triples... Definition: Where Represents a set of variables. This represents the set of possible values ​​for each variable. For a set of constraint (or cost) functions; each constraint function f∈ It defines the cost of certain variable assignment combinations within its scope. The ultimate goal of COP is to find a variable assignment scheme with the minimum global total cost.

[0113] A factor graph is a bipartite graph model used to describe Coding Operations (COPs): the graph contains variable nodes and factor nodes (representing constraint functions), and edges represent the interactions between variables and factors. Through factor graphs, complex global cost functions can be decomposed into the product or sum of multiple factor functions, thus intuitively showing the dependencies between variables and constraints, and providing a structural foundation for message passing algorithms.

[0114] Min-sum BP algorithm: The min-sum BP algorithm is one of the core message-passing algorithms for solving COP. This algorithm approximates the optimal solution by iteratively propagating messages on the factor graph. The specific process is as follows:

[0115] 1) A variable node sends a message to each of its neighboring factor nodes, and this message is equal to the sum of the messages received from all neighboring factors except the receiving factor;

[0116] 2) The factor node sends a message to each of its neighboring variable nodes. The message is equal to the minimum cost of the factor under all variable assignment combinations except the target variable (i.e., the local cost of the factor plus the message previously sent by the corresponding variable). Then, the cost of the assignment cases corresponding to these minimum costs is taken as the message.

[0117] After several iterations, each variable node accumulates the "belief" from all relevant factor nodes, representing the total cost after integrating various messages. Finally, each variable selects the assignment that minimizes its total belief as the output solution. This process can be viewed as finding the minimum value of a distribution on a graph structure. However, the traditional Min-sum BP algorithm faces significant challenges on factor graphs with loops: because messages propagate cyclically within loops and continuously accumulate errors, the algorithm often struggles to converge, or converges to a suboptimal local solution. This severely limits its effectiveness in solving complex real-world problems.

[0118] To alleviate the above problems, researchers have proposed improved algorithms such as Damped Belief Propagation (DBP). DBP introduces a constant damping factor when updating messages at variable nodes. ∈0,1, proportionally combine the old messages from the previous round with the newly calculated messages from this round. The message is then mixed and the result is used as the message to be sent in the current round. This damped update strategy smooths out message changes, suppresses oscillations to some extent, increases the convergence probability of BP on the cyclic factor graph, and improves the quality of the solution.

[0119] Although damping techniques have alleviated the oscillations of the backpropagation (BP) and improved convergence to some extent, the DBP still has two inherent drawbacks:

[0120] 1) The damping factor is static and singular, making it difficult to accommodate different situations: The optimal damping coefficient is highly dependent on the graph structure and constraint characteristics of the specific problem instance, usually requiring adjustment on a case-by-case basis using methods such as grid search or evolutionary algorithms, a cumbersome process. Fixed global... It cannot adapt to the dynamically changing context during message passing. For example, smaller damping may be needed in the early stages of an iteration to encourage exploration, while larger damping may be needed in the later stages to stabilize messages and promote convergence. However, DBP uses a globally uniform... The inability to adjust in real time according to different variables or iteration stages limits the algorithm's generalization ability and overall performance.

[0121] 2) Equal weighting of message combinations, lacking differentiation: In message computation at variable nodes, DBP, like traditional BP, directly sums all messages from neighboring factors without distinguishing their importance. In reality, different constraints may have varying degrees of influence on variables; some "tight constraints" or key factors have a greater guiding role in the solution, while other neighbor information may contain more noise. If an average aggregation strategy is still used, it will limit the algorithm's ability to accurately model complex association patterns, thus affecting search efficiency and solution quality.

[0122] Besides DBP, recent studies have attempted to integrate deep learning into the BP algorithm to improve performance. For example, neural networks are used to predict BP messages or variable update rules. However, most of these methods rely on supervised learning, requiring a large number of pre-computed optimal solutions as training labels. Obtaining the exact optimal solution for a large number of COP instances is inherently difficult (NP-hard problems typically require exponential time to solve), making these supervised learning methods less feasible in practical applications. Furthermore, offline-trained models are prone to performance drops due to distribution differences when facing new problems, exhibiting out-of-distribution generalization. Therefore, there is an urgent need for a new improved BP algorithm that can dynamically adjust message passing parameters (such as damping factors and neighbor weights), achieve efficient learning without expensive labeled data, and support GPU parallel acceleration, thereby overcoming the performance bottleneck of BP in solving COP problems.

[0123] Based on the above, existing technologies (mainly referring to traditional Min-sum BP and its damping variant DBP) have the following main shortcomings:

[0124] 1) Convergence problem: On factor graphs containing loops, the BP algorithm is prone to oscillation or non-convergence. Messages continue to propagate in the closed loop and amplify errors, making it difficult for the algorithm to stabilize to a fixed point.

[0125] 2) The quality of the solution is unstable: Even if the backpropagation eventually converges, it often gets trapped in local minima due to multiple propagations in the loop, resulting in poor quality of the solution and a significant gap from the global optimum.

[0126] 3) Reliance on empirical parameter tuning: Although DBP alleviates the convergence problem by introducing a damping factor, the optimal... The choice of damping strength is highly dependent on expert experience and extensive trial and error. Different problems, and even the same problem at different iteration stages, may require different damping strengths, while a fixed damping strength... The inability to dynamically adjust limits the algorithm's performance. Furthermore, the strategy of applying an indiscriminate average weighting to all neighbor messages of a variable ignores the differences in the importance of different constraints, further limiting the BP's exploration capabilities.

[0127] To address the aforementioned issues of poor convergence, suboptimal solution quality, and strong dependence on parameter tuning, this invention aims to introduce deep learning technology to design an adaptive message passing mechanism for dynamically optimizing key parameters in the backpropagation (BP) process. This method can automatically adjust the damping factor and neighbor weights without requiring manually labeled data, significantly enhancing the algorithm's convergence stability, improving solution quality, and reducing reliance on manual parameter tuning. Ultimately, it achieves efficient, robust, and adaptive solutions to constrained optimization problems.

[0128] This invention employs Belief Propagation (BP) as its core solution algorithm. By aggregating information in parallel and locally on a graph structure through message passing, it decomposes the global objective and constraints into several "local factors acting on subsets of variables," thus efficiently approximating the global optimal solution for large-scale problems. BP has both a sum-product form for finding marginal values ​​and a min-sum form more suitable for optimization, the latter directly handling "cost minimization" objectives. Furthermore, BP naturally possesses incremental update capabilities for new observations or structural changes, adapting to the engineering requirements of rolling optimization and rapid recalculation. In addition, this algorithm dynamically optimizes the BP message passing process using a data-driven approach, thereby overcoming the bottlenecks of traditional methods in terms of convergence and solution quality. The overall architecture of ABP comprises three core modules: factor graph modeling, an adaptive message passing mechanism, and a deep neural network controller.

[0129] In its implementation, the BP modeling steps involve factorizing VRPTW and constructing a factor graph: variable nodes include "vehicle departure time, customer service start time, arc selection between vehicle and customer, station service duration, and remaining vehicle capacity"; factor nodes include "time window factor, capacity factor, flow conservation and unique service factor, path feasibility factor, driving cost factor, and lateness penalty factor." Then, min-sum BP message passing is executed: variable nodes summarize messages from neighboring factors to form a local cost assessment of their own values; factor nodes perform "minimum aggregation" of their local costs with messages from other neighboring variables before returning them to the variable nodes. After iteration until convergence or the stopping criterion is met, the "value corresponding to the minimum marginal cost" is selected for each variable node, resulting in a directly implementable delivery decision, including vehicle departure time, vehicle-customer access order and arc selection, and service duration at each station; if necessary, a consistency check and lightweight patch are performed at the end to obtain a globally optimal solution that satisfies constraints.

[0130] Despite the advantages mentioned above, standard BP is prone to oscillations, unstable convergence, or convergence to poor-quality solutions on graphs containing cyclic factors. Traditional damped belief propagation (DBP) smooths the iteration by mixing previous messages into message updates, which can improve convergence to some extent. However, it has two shortcomings: first, the globally fixed damping coefficient is difficult to adapt to the dynamic changes of different instances and different iteration stages; second, it treats all neighbor messages equally, failing to reflect the differences in the degree of influence of different neighbors on variables, leading to insufficient information utilization.

[0131] To address the aforementioned bottlenecks, this invention proposes an adaptive backpropagation (BP) approach, introducing a data-driven parameter adjustment mechanism without altering the basic BP framework. The core of this mechanism includes:

[0132] 1) Adaptive message damping: GAT dynamically provides a damping coefficient for each message edge. ∈[0,1]. When updating the variable node, use... The "previous round of old messages" and the "current round of newly calculated messages" are linearly weighted to obtain the smoothed current message; The focus is on stability during periods of high activity. The small-scale approach focuses on exploration, thereby adaptively suppressing oscillations and improving convergence speed during the iteration process.

[0133] 2) Neighbor-weighted adaptive: The attention mechanism assigns differentiated weights to different neighbors, highlighting message sources that are more sensitive to the target or strong constraints, and avoiding uniform weighting that dilutes key information.

[0134] 3) Centralized batch parameter inference: Reduces the call to the deep model from every step update to once per round (several steps), and centrally predicts the damping and weights required for an entire round between iterations, significantly reducing computation and memory overhead.

[0135] 4) Multiple independent restarts and self-supervised training: Multiple random restarts reduce the probability of getting stuck in local minima; self-supervised or unsupervised objectives are used for online learning to get rid of the dependence on the optimal solution label, while improving the adaptability to distribution changes.

[0136] Through the above design, the algorithm can output a scheduling scheme with higher time window satisfaction rate, lower total mileage, and more robustness to dynamic changes in VRPTW scenarios, while maintaining a rapid response capability to new orders and traffic disturbances.

[0137] In summary, this invention employs a linear embedding layer and a multi-head graph attention network to encode the factor graph and messages, inferring all damping factors and neighbor weights required for the entire BP iteration in one go. This avoids the stepwise predictions of the autoregressive model, significantly reducing memory usage and improving inference efficiency. This innovative centralized parameter inference architecture drastically reduces the number of deep model calls, resolving the performance bottleneck of the original DNN-BP method in long sequence inference.

[0138] Secondly, this invention designs unsupervised loss functions based on belief entropy minimization, cost smoothing regularization, and dynamic constraints, and employs online training to enable the model to autonomously learn the optimal damping factor and weights under unlabeled data conditions. By smoothing the cost function of the BP solution and combining it with feedback on constraint satisfaction, this invention achieves an end-to-end self-supervised learning framework that ensures effective model convergence without requiring expensive optimal solution labels.

[0139] Meanwhile, this invention can automatically adjust the weights and damping factors of each neighbor message based on the current BP message state, balancing global convergence stability with an emphasis on local importance constraints. Compared to a fixed strategy, the dynamically adjustable parameters allow the algorithm to achieve a balance between oscillation suppression and search exploration, significantly improving convergence speed and solution accuracy.

[0140] Furthermore, this invention utilizes mixed-precision training and computation, reducing floating-point arithmetic precision in exchange for smaller GPU memory usage. Simultaneously, it appropriately introduces graph partitioning techniques to divide the massive factor graph into several subgraphs for separate solving, or loads them into memory for processing in stages. These methods effectively reduce resource consumption when solving large-scale COP problems, improving the algorithm's scalability and practical value.

[0141] Compared to existing technologies, this invention significantly improves overall speed by jointly inferring damping factors and neighbor weights, reducing the number of deep network calls, and highly parallelizing the BP inference process, achieving an average speedup of over 2× (compared to existing deep BP algorithms). This increased speed enables the algorithm to explore a larger solution space and handle larger-scale COP instances within a limited time. Secondly, this invention employs centralized inference and mixed precision optimizations, decoupling memory usage from the number of BP iterations, thus preventing linear growth. Under the same hardware conditions, the algorithm can solve larger-scale problems that were previously inaccessible due to memory limitations, greatly improving the scalability of the solution. This invention uses a self-supervised training strategy, updating model parameters online by optimizing unsupervised objectives (such as smoothing solution costs). It eliminates the need for expensive optimal solution labels, lowering the model training threshold, and avoids performance degradation caused by mismatches between training distribution and actual problems through online learning, resulting in a wider applicability and stronger generalization. Furthermore, by dynamically adjusting neighbor weights and damping factors, this invention allows the algorithm to more accurately aggregate key constraint information, reducing message oscillations and improving solution accuracy. Practical tests show that the solution cost obtained by the method of this invention is often superior to existing best algorithms, approaching or even reaching the known optimal cost. Furthermore, the multiple random restart strategy further increases the probability of finding the global optimum, ensuring the quality of the solution. In addition, the framework provided by this invention is applicable to various types of COP scenarios, including stochastic constraint satisfaction problems, graph coloring problems, and scheduling planning. It performs excellently in benchmark tests of different structures and scales. Moreover, the framework has good extensibility, allowing for customization of the loss function and model structure according to specific problem characteristics (e.g., designing specialized regularization terms for domain-specific constraints), thereby further improving the solution performance. This means that the algorithm of this invention can serve as a general-purpose COP solution engine with broad adaptability in practical applications.

[0142] In this embodiment of the invention, a logistics scheduling method based on deep learning and adaptive belief propagation is provided. The method obtains a logistics scheduling instance to be solved and transforms it to generate a logistics scheduling factor graph. Based on an adaptive message passing mechanism, a graph attention network is used to dynamically generate parameters according to the logistics scheduling factor graph, outputting the target message between each variable node and its neighboring factor nodes. The target messages between each variable node and its neighboring factor nodes are weighted and aggregated to output the belief value corresponding to each variable node. Each variable node includes multiple logistics scheduling actions, and each logistics scheduling action has a corresponding belief value. From the belief values ​​associated with each logistics scheduling action corresponding to each variable node, the logistics scheduling action corresponding to the smallest belief value is selected as the target logistics scheduling action. Based on multiple target logistics scheduling actions, a target logistics scheduling scheme is constructed. Based on the above scheme, this invention leverages the dynamic adaptability of the graph attention network to the constraints and dynamic changes of the scheduling scenario, and the automatic adjustment of dynamic parameters by the adaptive message passing mechanism. This allows message updates to break free from the limitations of traditional fixed parameters, reducing reliance on manual parameter tuning, and thus outputting a high-quality logistics scheduling scheme.

[0143] Please see Figure 5 , Figure 5 This is a structural block diagram of a logistics scheduling device based on deep learning and adaptive belief propagation, provided in Embodiment 2 of the present invention.

[0144] This invention provides a logistics scheduling device based on deep learning and adaptive belief propagation, comprising:

[0145] The acquisition module 501 is used to acquire the logistics scheduling instance to be solved and transform the logistics scheduling instance to be solved to generate a logistics scheduling factor graph.

[0146] Output module 502 is used to generate dynamic parameters based on the logistics scheduling factor graph using a graph attention network based on an adaptive message passing mechanism, and output the target message between each variable node and each neighbor factor node in the logistics scheduling factor graph.

[0147] The aggregation module 503 is used to perform weighted aggregation of the target messages between each variable node and each neighbor factor node, and output the belief value corresponding to each variable node. The variable node includes multiple logistics scheduling execution actions, and each logistics scheduling execution action has a corresponding belief value.

[0148] The constitutive module 504 is used to select the logistics scheduling execution action corresponding to the smallest belief value from the belief values ​​associated with each logistics scheduling execution action corresponding to each variable node as the target logistics scheduling execution action, and to form a target logistics scheduling scheme based on multiple target logistics scheduling execution actions.

[0149] Furthermore, the adaptive message passing mechanism is as follows:

[0150] ;

[0151] in, At time step t, the new message that variable node i sends to its neighbor factor node a, which is the intermediate message between variable node i and its neighbor factor node a; At the t-th time step, the edge-specific damping factor of the edge from variable node i to neighbor factor node a, that is, the edge-specific damping factor between variable node i and neighbor factor node a; At the (t-1)th time step, the message sent by variable node i to neighbor factor node a represents the initial message between variable node i and neighbor factor node a. In the set of neighbor factors of variable node i, except Other neighbor factor nodes; Let i be the neighboring factor node of the variable node i that is currently sending a message; Let i be the set of neighboring factors of variable node i; For the t-th time step, the neighbor-specific weights between variable node i and its neighbor factor node a; At time step t-1, the message sent by neighbor factor node b to variable node i represents the initial message between neighbor factor node b and variable node i.

[0152] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described device and module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0153] This invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program; when the computer program is executed by the processor, the processor performs the steps of the logistics scheduling method based on deep learning and adaptive belief propagation as described in any of the above embodiments.

[0154] This invention also provides a computer-readable storage medium storing a computer program / instructions thereon, which, when executed by a processor, implements the steps of the logistics scheduling method based on deep learning and adaptive belief propagation as described in any of the above embodiments.

[0155] This invention also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the logistics scheduling method based on deep learning and adaptive belief propagation as described in any of the above embodiments.

[0156] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0157] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0158] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A logistics scheduling method based on deep learning and adaptive belief propagation, characterized in that, The method comprises the following steps: acquiring a logistics scheduling instance to be solved, and converting the logistics scheduling instance to be solved to generate a logistics scheduling factor graph; based on an adaptive message passing mechanism, using a graph attention network to generate dynamic parameters according to the logistics scheduling factor graph, and outputting target messages between each variable node and each neighbor factor node in the logistics scheduling factor graph; respectively weighting and aggregating the target messages between each variable node and each neighbor factor node to output a belief value corresponding to each variable node, wherein the variable node includes a plurality of logistics scheduling execution actions, and each logistics scheduling execution action has a corresponding belief value; selecting a logistics scheduling execution action corresponding to the smallest belief value from the belief values associated with each logistics scheduling execution action of each variable node as a target logistics scheduling execution action, and constructing a target logistics scheduling scheme according to a plurality of target logistics scheduling execution actions. 2.The logistics scheduling method based on deep learning and adaptive belief propagation according to claim 1, wherein, The adaptive message passing mechanism uses a graph attention network to generate dynamic parameters according to the logistics scheduling factor graph, and outputs target messages between each variable node and each neighbor factor node in the logistics scheduling factor graph, which comprises: initializing the messages of each variable node in the logistics scheduling factor graph, and outputting initial messages between each variable node and each neighbor factor node; real-time statistics of the current iteration number, and the initial messages between each variable node and each neighbor factor node, and the logistics scheduling factor graph as the input of the graph attention network, to generate edge-specific damping factors and neighbor-specific weights between each variable node and each neighbor factor node at each time step corresponding to the current iteration number; using the adaptive message passing mechanism to calculate intermediate messages between each variable node and each neighbor factor node according to the edge-specific damping factors and neighbor-specific weights between each variable node and each neighbor factor node at each time step; determining whether the current iteration number reaches a preset number threshold; if yes, the intermediate messages between each variable node and each neighbor factor node are taken as the corresponding target messages. 3.The logistics scheduling method based on deep learning and adaptive belief propagation according to claim 2, characterized in that, Further comprising: if the current iteration number does not reach the preset number threshold, the intermediate messages between each variable node and each neighbor factor node are taken as the corresponding new initial messages; jumping to the step of real-time statistics of the current iteration number, and the initial messages between each variable node and each neighbor factor node, and the logistics scheduling factor graph as the input of the graph attention network, to generate edge-specific damping factors and neighbor-specific weights between each variable node and each neighbor factor node at each time step corresponding to the current iteration number; until the current iteration number reaches the preset number threshold, the intermediate messages between each variable node and each neighbor factor node determined when the current iteration number reaches the preset number threshold are taken as the corresponding target messages. 4.The logistics scheduling method based on deep learning and adaptive belief propagation according to claim 1, wherein, The logistics scheduling factor graph is specifically: ； wherein, is a factor graph for logistics scheduling; is a set of variable nodes; is a set of neighbor factor nodes, i.e. a set of constraint function nodes; is a set of edges, representing the association between variables and factors. 5.The logistics scheduling method based on deep learning and adaptive belief propagation according to claim 1, wherein, The adaptive message passing mechanism, in particular: ； wherein, is the new message sent by the variable node i to the neighboring factor node a at the t-th time step, i.e., the intermediate message between the variable node i and the neighboring factor node a; is the edge-specific damping factor for the edge from the variable node i to the neighboring factor node a at the t-th time step, i.e., the edge-specific damping factor between the variable node i and the neighboring factor node a; is the message sent by the variable node i to the neighboring factor node a at the t-1-th time step, representing the initial message between the variable node i and the neighboring factor node a; is the neighboring factor node in the neighbor factor set of the variable node i, other than ; is the neighboring factor node to which the variable node i currently sends a message; is the neighbor factor set of the variable node i; is the neighbor-specific weight between the variable node i and the neighboring factor node a at the t-th time step; is the message sent by the neighboring factor node b to the variable node i at the t-1-th time step, representing the initial message between the neighboring factor node b and the variable node i. 6.A logistics scheduling device based on deep learning and adaptive belief propagation, characterized by, Comprising: An acquisition module, configured to acquire a logistics scheduling instance to be solved, and convert the logistics scheduling instance to be solved to generate a logistics scheduling factor graph; An output module, configured to generate dynamic parameters according to the logistics scheduling factor graph based on an adaptive message passing mechanism and a graph attention network, and output target messages between each variable node and each neighbor factor node in the logistics scheduling factor graph; An aggregation module, configured to respectively aggregate the target messages between each variable node and each neighbor factor node to output a belief value corresponding to each variable node, wherein the variable node includes a plurality of logistics scheduling execution actions, and each logistics scheduling execution action has a corresponding belief value; A configuration module, configured to select a logistics scheduling execution action corresponding to a minimum belief value as a target logistics scheduling execution action in each logistics scheduling execution action associated with a belief value corresponding to each variable node, and configure a target logistics scheduling scheme according to a plurality of target logistics scheduling execution actions.

7. The deep learning and belief propagation based logistics scheduling apparatus according to claim 6, wherein, The adaptive message passing mechanism, in particular: ； wherein, is the new message sent by the variable node i to the neighboring factor node a at the t-th time step, i.e., the intermediate message between the variable node i and the neighboring factor node a; is the edge-specific damping factor for the edge from the variable node i to the neighboring factor node a at the t-th time step, i.e., the edge-specific damping factor between the variable node i and the neighboring factor node a; is the message sent by the variable node i to the neighboring factor node a at the t-1-th time step, representing the initial message between the variable node i and the neighboring factor node a; is the neighboring factor node of the variable node i, other than ; is the neighboring factor node of the variable node i, other than ; is the neighboring factor node of the variable node i, other than ; is the neighboring factor node of the variable node i, other than is the neighbor-specific weight between the variable node i and the neighboring factor node a at the t-th time step; is the message sent by the neighboring factor node b to the variable node i at the t-1-th time step, representing the initial message between the neighboring factor node b and the variable node i.

8. A computer device, comprising: Comprising a memory and a processor, the memory storing a computer program, and the computer program being executed by the processor to make the processor execute the steps of the logistics scheduling method based on deep learning and adaptive belief propagation according to any one of claims 1-5.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed to implement the logistics scheduling method based on deep learning and adaptive belief propagation according to any one of claims 1-5.

10. A computer program product, characterised in that, The computer program product comprises a computer program stored on a non-transitory computer readable storage medium, and the computer program comprises program instructions, wherein when the program instructions are executed by a computer, the computer executes the logistics scheduling method based on deep learning and adaptive belief propagation according to any one of claims 1-5.

Images