Construction method of double-path collaborative decision network for multi-agent collaborative path optimization

Abstract

The invention provides a construction method of a double-path collaborative decision network for multi-agent collaborative path optimization, and the method comprises the steps: obtaining multi-source heterogeneous real-time business data and historical business data, and constructing a standardized data set; converting the standardized data set into a dynamic environment input sequence; at least two domain agents perform full-process business logical reasoning based on the dynamic environment input sequence to generate a decision chain structure covering the full life cycle of the business; implementing full-flow path planning simulation based on the decision chain structure, and generating an initial path scheme including multi-resource collaborative scheduling; the environment perception intelligent agent puts the initial distribution path scheme into an actual logistics distribution scene for execution, and loads real-time environment change information to determine a logistics decision chain weight; and the logistics strategy agent optimizes the initial distribution path scheme based on the logistics decision chain weight to obtain a logistics business knowledge graph so as to construct a logistics double-path collaborative decision network.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a method for constructing a dual-path collaborative decision-making network for multi-agent collaborative path optimization. Background Art

[0002] With the rapid development and intelligent transformation of the logistics industry, multi-agent collaboration is becoming increasingly widespread in logistics and distribution scenarios. Through the coordinated operation of multiple agents, multi-agent systems can efficiently schedule logistics resources and optimize delivery routes to meet the growing demand for logistics services. However, current multi-agent logistics systems still face numerous technical challenges.

[0003] At the data level, logistics operations involve heterogeneous data from multiple sources, including real-time traffic conditions, order information, and vehicle status. Traditional data collection and processing methods struggle to effectively integrate this data, resulting in poor data consistency and low information utilization. Furthermore, due to the diverse and dynamic nature of data sources, noise and missing values ​​in the data can severely impact the accuracy of subsequent decision-making. For example, in complex urban traffic environments, inaccurate real-time traffic data can lead to inappropriate delivery route planning, increasing transportation costs and delaying delivery times.

[0004] In terms of decision-making mechanisms, existing multi-agent decision-making methods typically employ a single decision-making model, such as rule-based or learning-based decision-making. Rule-based methods struggle to adapt to the dynamic logistics environment and lack adaptability. Learning-based methods, such as deep reinforcement learning, while capable of achieving autonomous decision-making to a certain extent, often lack interpretability and struggle to incorporate logistics expertise and experience, resulting in decision-making outcomes that may be out of sync with actual business needs. Furthermore, the collaborative decision-making mechanisms among multi-agents are still imperfect, and agents lack effective information exchange and collaboration during the decision-making process. This prevents the full utilization of the advantages of multi-agent systems and makes it difficult to achieve globally optimal decision-making results.

[0005] When it comes to route optimization, existing route planning algorithms mostly assume a static environment, making them incapable of handling real-time changes in road conditions and dynamic adjustments to orders during the logistics and delivery process. When faced with unexpected traffic incidents or unexpected new orders, route planning solutions often cannot be adjusted in a timely manner, resulting in reduced delivery efficiency. Furthermore, existing route optimization methods lack systematicity and holistic approach when considering the coordinated scheduling of multiple resources (such as vehicles, personnel, and storage resources), making it difficult to achieve optimal resource allocation. Summary of the Invention

[0006] The purpose of this application is to propose a method for constructing a dual-path collaborative decision-making network for multi-agent collaborative path optimization, which is used to solve or alleviate the technical problems existing in the existing technology.

[0007] A method for constructing a dual-path collaborative decision-making network for multi-agent collaborative path optimization, comprising:

[0008] Acquire multi-source heterogeneous real-time business data and historical business data to build standardized data sets;

[0009] converting the standardized data set into a dynamic environment input sequence;

[0010] At least two domain agents perform full-process business logic reasoning based on the dynamic environment input sequence to generate a decision chain structure covering the entire life cycle of the business;

[0011] Implementing full-process path planning simulation based on the decision chain structure to generate an initial path plan including multi-resource collaborative scheduling;

[0012] The environment-aware agent executes the initial delivery path plan in the actual logistics delivery scenario, loading real-time environmental change information to determine the weight of the logistics decision chain;

[0013] The logistics strategy agent optimizes the initial distribution path plan based on the logistics decision chain weight to obtain the logistics business knowledge graph to build a logistics two-way collaborative decision network.

[0014] The specific technical benefits of this application's technical solution are as follows:

[0015] 1. This solution constructs a standardized data set by acquiring real-time and historical business data from multiple sources and converting it into a dynamic environment input sequence. This process enables unified processing of complex data, eliminates data inconsistencies, and improves data quality and information utilization. This standardized data set provides a reliable foundation for subsequent intelligent decision-making, avoids errors caused by data issues, and significantly improves the accuracy and stability of data-driven decision-making.

[0016] 2. This solution utilizes at least two domain agents to perform full-process business logic reasoning based on dynamic environment input sequences, generating a decision chain structure that covers the entire business lifecycle. This decision chain not only enables effective collaboration among multiple agents but also incorporates logistics expertise and experience through logical reasoning, making the decision-making process explainable. Compared to traditional single-agent decision models, this solution's decision chain structure enables a global perspective, enabling multi-agent collaborative optimization, improving the scientific and rationality of decision-making, and better meeting the actual needs of logistics operations.

[0017] 3. This solution implements a full-process route planning simulation based on a decision chain structure, generating an initial route plan that incorporates the coordinated scheduling of multiple resources. Using an environmentally aware agent and real-time environmental information, it determines the weights of the logistics decision chain and optimizes the route plan. This dynamic adjustment mechanism enables route planning to respond in real time to unexpected situations such as changing road conditions and order adjustments, ensuring optimal delivery routes. Furthermore, the design of multi-resource coordinated scheduling achieves a globally optimized allocation of resources, including vehicles, personnel, and storage, improving resource utilization efficiency and reducing logistics costs, effectively addressing the limitations of traditional route optimization methods.

[0018] 4. The resulting dual-path collaborative decision-making network for logistics, constructed through this solution, integrates the advantages of the aforementioned technical links. Through the optimization of routing plans by a logistics strategy agent and the construction of a knowledge graph, the decision-making network is endowed with self-learning and adaptive capabilities. Compared to existing logistics systems, this dual-path collaborative decision-making network can continuously optimize decision-making strategies in complex and changing logistics environments, improving the overall performance and service quality of the logistics system, and providing a more efficient and reliable technical solution for the intelligent development of the logistics industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 A flowchart of a method for constructing a dual-path collaborative decision-making network for multi-agent collaborative path optimization according to an embodiment of the present application.

[0020] Figure 2 This is a structural diagram of a device for constructing a dual-path collaborative decision-making network for multi-agent collaborative path optimization according to an embodiment of the present application.

[0021] Figure 3 Schematic diagram of the structure of the electronic device of this embodiment.

[0022] Figure 4 This is the hardware structure of the electronic device in this embodiment. DETAILED DESCRIPTION

[0023] In order to enable those skilled in the art to better understand the technical solutions in the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the embodiments of the present application, all other embodiments obtained by ordinary technicians in this field should fall within the scope of protection of the embodiments of the present application.

[0024] like Figure 1 As shown, the embodiment of the present application provides a method for constructing a dual-path collaborative decision-making network for multi-agent collaborative path optimization, which includes:

[0025] Acquire multi-source heterogeneous real-time business data and historical business data to build standardized data sets;

[0026] converting the standardized data set into a dynamic environment input sequence;

[0027] At least two domain agents perform full-process business logic reasoning based on the dynamic environment input sequence to generate a decision chain structure covering the entire life cycle of the business;

[0028] Implementing full-process path planning simulation based on the decision chain structure to generate an initial path plan including multi-resource collaborative scheduling;

[0029] The environment-aware agent executes the initial delivery path plan in the actual logistics delivery scenario, loading real-time environmental change information to determine the weight of the logistics decision chain;

[0030] The logistics strategy agent optimizes the initial distribution path plan based on the logistics decision chain weight to obtain the logistics business knowledge graph to build a logistics two-way collaborative decision network.

[0031] In summary, this application decomposes logistics problems into computable decision-making sub-chains through multi-agent systems, generates initial solutions through virtual simulation, and optimizes decision weights through real-time feedback. Finally, with the help of knowledge graphs, dual-path collaboration of data and knowledge is achieved, enabling the system to continuously evolve in complex logistics scenarios and achieve dynamic optimal solutions for multi-resource collaborative paths. Exemplary technical implementations are as follows:

[0032] Through standardized integration of multi-source data, a unified foundation for cross-domain decision-making is established. For example, this involves acquiring heterogeneous data from multiple sources, such as real-time traffic conditions and order information, and then constructing a standardized data set through cleaning and alignment. The essence of this technology is to eliminate differences in data formats and spatiotemporal references. For example, geographic coordinates from different sources are unified into a GPS coordinate system, and data such as order timeliness and vehicle status are synchronized by timestamp to form a structured foundational data layer. This step provides unambiguous input for subsequent intelligent agent reasoning, resolving the decision-making errors caused by the "disorganized formats and spatiotemporal misalignment" of traditional logistics data.

[0033] By generating dynamic environment sequences, data is given spatiotemporal dynamic features. For example, the standardized data is converted into a dynamic environment input sequence, which is essentially a time-series and scenario-based modeling of the data. For example, real-time road condition data is sliced ​​into traffic flow sequences by time, and order data is constructed into time window sequences according to delivery time requirements. This processing gives static data the characteristics of "spatiotemporal evolution", making it easier for intelligent agents to capture dynamic associations in logistics scenarios (such as the impact of morning rush hour traffic conditions on delivery time), and provide dynamic environmental representation for full-process decision-making.

[0034] Through multi-agent collaborative reasoning, a domain knowledge-driven decision chain is constructed. For example, at least two domain agents (such as warehouse distribution, scheduling, and route planning agents) are used to perform logical reasoning on dynamic sequences. Essentially, the entire logistics process is broken down into independently solvable sub-problems, which are then integrated into a global solution through a decision chain. For example, the intelligent warehouse distribution agent generates the optimal warehouse distribution strategy based on order and warehouse data to solve the problem of "where to ship goods from the warehouse"; the scheduling agent generates a scheduling plan based on vehicle status and order timeliness to solve the problem of "who will perform the delivery"; and the route planning agent generates a path plan based on road conditions and delivery point locations to solve the problem of "how to transport efficiently."

[0035] The sub-chains output by each intelligent entity are connected in series through business logic (such as the warehouse distribution results determine the starting point for vehicle pickup), forming a full life cycle decision chain covering warehousing, transportation, and distribution, achieving the leap from "local optimization" to "global coordination."

[0036] Through full-process path simulation, a virtual preview of multi-resource collaboration is conducted. For example, a path planning simulation based on a decision chain is carried out, which essentially converts abstract decisions into executable resource scheduling plans. For example, it simulates the order consolidation strategy when picking up and loading, the multi-vehicle collaborative route in trunk transportation, and the dispatch order of terminal distribution, while considering resource constraints such as vehicle load and storage capacity. This virtual preview can detect resource conflicts in advance (such as vehicle overload and time window overlap), generate an initial path plan, and avoid resource waste in actual execution.

[0037] Decision optimization driven by environmental feedback is achieved through dynamic weighting. An environmentally aware agent implements the initial plan, loading it with dynamic information such as real-time traffic conditions and unexpected orders. Essentially, this establishes a closed-loop "execution-feedback" optimization mechanism. For example, if a road section experiences sudden congestion during an actual delivery, the system calculates the impact of this event on route execution and adjusts the weight of the "road condition factor" in the decision chain, prioritizing subsequent optimization efforts with greater emphasis on high-impact factors. This dynamic adjustment enables the decision model to adapt to complex environments, resolving the real-time limitations inherent in traditional route planning, which relies on static assumptions.

[0038] Through the construction of a two-way collaborative decision-making network, the deep integration of knowledge graphs and dynamic optimization is implemented. Among them, the logistics strategy agent optimizes the initial plan based on weights and constructs a knowledge graph. The essence is to combine data-driven and knowledge-driven: 1. Knowledge graph construction: Extract distribution node relationships (such as warehouse-distribution point associations) and business rules (such as "cold chain goods require refrigerated trucks for transportation") from the optimized plan to form a structured logistics knowledge system to provide prior knowledge support for decision-making; 2. Two-way collaborative mechanism: "Data-driven path": Through algorithms such as reinforcement learning, the path strategy is dynamically adjusted based on real-time data; "Knowledge-driven path": Use knowledge graphs for logical reasoning to ensure that decisions comply with industry regulations (such as restrictions on dangerous goods transportation routes); 3. Network mapping implementation: Map the two-way decision logic to order management, vehicle scheduling and other systems to achieve end-to-end collaboration of "data processing-decision generation-business execution", and ultimately build an intelligent decision-making network with self-learning capabilities.

[0039] Optionally, in a specific embodiment, acquiring multi-source heterogeneous real-time business data and historical business data and constructing a standardized data set includes:

[0040] Based on real-time logistics road condition data, delivery order information data, delivery vehicle status data, and historical delivery business data, a standardized data set for logistics business is constructed.

[0041] Optionally, in a specific embodiment, the construction of a standardized logistics business data set based on real-time logistics road condition data, delivery order information data, delivery vehicle status data, and historical delivery business data includes:

[0042] Perform spatiotemporal alignment on real-time logistics traffic data, delivery order information data, delivery vehicle status data, and historical delivery business data to unify them into the same spatiotemporal coordinate system and obtain a spatiotemporal aligned original logistics data stream.

[0043] Perform wavelet transform and multi-resolution transform on the original data stream to decompose the original data stream into different frequency levels to remove the noise and obtain the effective logistics data stream;

[0044] Based on the effective logistics data flow, a standardized data set for logistics business is constructed.

[0045] In summary, in this application, a unified benchmark for cross-domain data is constructed based on the spatiotemporal alignment of multi-source data. It eliminates the spatiotemporal heterogeneity of logistics data by establishing a standardized spatiotemporal mapping system. For example, when aligning spatial coordinates, the GPS coordinates of real-time road conditions, the geocoding of order delivery points, the inertial navigation data of vehicle positions, etc. are mapped to a unified geographic coordinate system (such as WGS84) through a coordinate conversion algorithm, and the grid is divided based on the needs of logistics business (such as 10 meters × 10 meters accuracy) to make spatial data from different sources comparable. For example, the warehouse location is associated with the congested road section of real-time traffic conditions in the same spatial grid to ensure that the delivery distance can be accurately calculated when making warehouse decisions. When synchronizing the time dimension, the timestamp of the order timeliness, the collection time of the vehicle status, and the update frequency of the road condition data are uniformly calibrated, and the NTP network time protocol is used to achieve nanosecond time synchronization. For example, the "latest delivery time" of the order and the "estimated arrival time" of the vehicle are aligned to the same time scale to avoid misjudgment of delivery timeliness due to time base differences.

[0046] Different from the traditional processing method of logistics data that only performs format conversion, this solution achieves dual alignment of logistics data in the "spatial location-time series" dimension through dynamic spatiotemporal grid division and high-precision timestamp synchronization, solving the decision-making error problem caused by the "spatiotemporal misalignment" of heterogeneous data and providing a spatiotemporally consistent basic data layer for subsequent intelligent agent reasoning.

[0047] In addition, the present application uses wavelet transform multi-resolution denoising and the time-frequency decomposition characteristics of wavelet transform to filter data noise without losing key information. For example, through multi-resolution signal decomposition, the original logistics data stream (such as real-time vehicle speed series, order volume fluctuation curve) is decomposed into different frequency levels, the high-frequency components correspond to burst noise (such as GPS signal drift, sensor anomalies), and the low-frequency components correspond to real logistics characteristics (such as road condition trends during peak hours in the morning and evening, and periodic fluctuations in order volume). Through adaptive noise filtering, the noise in the high-frequency components is attenuated by the threshold function, retaining the high-frequency details that reflect the real business characteristics (such as the instantaneous peak of sudden orders), while retaining the trend characteristics in the low-frequency components. For example, when processing vehicle speed data, it can filter the GPS jump noise in the tunnel and retain the real speed changes caused by traffic lights.

[0048] Compared to traditional filtering algorithms (such as mean and Gaussian filters), which damage the dynamic characteristics of logistics data, this application achieves the dual goals of "precise noise filtering and lossless preservation of business characteristics" through multi-resolution analysis and adaptive threshold adjustment. Compared with traditional methods, this technology can retain more effective features when filtering logistics data noise, improving the input quality of subsequent decision-making models.

[0049] Finally, the construction logic of the standardized data set is used to convert heterogeneous input to structured output. After spatiotemporal alignment and wavelet denoising, the original logistics data stream is transformed into an effective data stream with spatiotemporal consistency and low noise interference, and finally constructed into a standardized data set. For example, based on data dimension integration, real-time road conditions, order information, vehicle status and other data are indexed by spatiotemporal coordinates to form a three-dimensional data structure of "spatial grid-time window-business attributes". For example, each spatiotemporal grid unit (such as the 10:00-10:15 period in a certain area) integrates multi-dimensional features such as road speed, order density, and number of available vehicles in the area. Based on business semantic annotation, data is semantically encoded based on logistics domain knowledge, such as converting business attributes such as order type (cold chain / general cargo) and vehicle attributes (load capacity / refrigeration function) into standardized labels to make the data business interpretable. It can be seen that this application constructs a standardized set that meets the requirements of computer processing and fits the logistics business logic through the integration of three-dimensional data structure and semantic annotation, providing high-quality input for multiple intelligent agents that is "temporally and spatially ordered, semantically clear, and noise controllable", fundamentally solving the problem of low efficiency in decision-making model training caused by the "chaotic format and fuzzy features" of traditional logistics data.

[0050] Optionally, in a specific embodiment, constructing a standardized logistics business data set based on the valid logistics data flow includes:

[0051] Perform multi-dimensional feature extraction on effective logistics data streams and perform dynamic local mapping and vector space alignment to obtain a set of semantic feature vectors;

[0052] Based on the constructed spatiotemporal index and semantic index, the spatiotemporal semantic tensor and logistics knowledge graph are injected into the semantic feature vector set to generate a standardized data set for logistics business.

[0053] The essential principle of this technical solution in constructing a standardized data set for logistics business is to transform logistics data into an intelligent representation system with spatiotemporal semantic understanding capabilities through the deep integration of dynamic semantic mapping and knowledge injection. Through multi-dimensional feature extraction and dynamic mapping, a semantically consistent data representation space is constructed to achieve feature decoupling and spatial unification, solving the semantic heterogeneity problem of logistics data. Specifically, through multi-dimensional feature decoupling, spatiotemporal features (such as distribution point coordinates, order time windows), business features (cargo type, vehicle load), and relationship features (warehouse-distribution point distance, historical delivery frequency) are extracted from the effective logistics data stream, decomposing the original data into semantic dimensions that can be independently represented. For example, order data is decoupled into three-dimensional features of "spatial location-time requirements-cargo attributes", so that the similarities of different orders can be quantified and compared. Through dynamic local mapping and vector alignment, the feature mapping rules are adaptively adjusted according to real-time business scenarios (such as peak / off-peak periods) based on dynamic local mapping. For example, when orders surge, the mapping weight of the "timeliness feature" is strengthened to ensure that key information is not weakened; through vector space alignment, heterogeneous features such as the "congestion index" of road data and the "urgency" of order data are mapped to the same high-dimensional vector space through a cross-modal alignment algorithm, making different types of data semantically comparable (such as mapping "heavy rain weather" and "high-timeliness orders" to similar vector representations to reflect their similar impact on delivery).

[0054] To this end, different from the static mapping limitations of traditional feature extraction, in this application, dynamic weight adjustment and cross-modal alignment are used to achieve intelligent conversion of logistics data from "raw values" to "semantic vectors", so that the data has the ability to adaptively represent business scenarios.

[0055] In addition, spatiotemporal-semantic index construction and tensor injection are used to construct a spatiotemporal-semantic index. When constructing the spatiotemporal index, the distribution area is hierarchically encoded based on the GeoHash algorithm, and combined with the time window division (such as slicing by 15 minutes) to form a two-dimensional index structure of "spatial grid-time slice", so as to realize the spatiotemporal rapid retrieval of data (such as querying the historical distribution data of the morning rush hour in a certain area in seconds); in addition, when constructing the semantic index, an inverted index is constructed based on the logistics business terminology system (such as "cold chain transportation" and "less-than-load delivery"), and the semantic feature vector is associated with the business concept to support fuzzy queries at the semantic level (such as searching for "all high-efficiency orders requiring refrigerated trucks").

[0056] Through the injection of spatiotemporal semantic tensors and knowledge graphs, when constructing the spatiotemporal semantic tensor, the semantic feature vectors are organized into high-order tensors (such as three-dimensional tensor T [spatial grid, time slice, business attributes]) according to the "space-time-business" dimension to capture the spatiotemporal correlation between data (such as the trend of changes in order volume in a certain area for three consecutive days); when injecting the logistics knowledge graph, the entity relationships in the knowledge graph (such as "the maximum refrigerated inventory of warehouse A" and "the load limit of vehicle B") are embedded in the tensor structure as prior knowledge. For example, the "warehouse-vehicle matching degree" feature is added to the "business attribute" dimension of the tensor, so that the data naturally contains logistics business rules.

[0057] To this end, by integrating the logical reasoning ability of knowledge graphs and the spatiotemporal correlation expression ability of tensors into the data standardization process, the traditional "data set" is upgraded to a "reasonable knowledge carrier", providing input with both data integrity and logical rigor for subsequent intelligent agent decision-making.

[0058] Optionally, in a specific embodiment, the at least two domain agents perform full-process business logic reasoning based on the dynamic environment input sequence to generate a decision chain structure covering the entire life cycle of the business, including: the intelligent warehouse distribution agent, the distribution vehicle scheduling agent, and the route planning agent perform logic reasoning on the entire life cycle of the logistics order based on the dynamic environment input sequence to generate a decision chain structure covering the entire process of warehousing and outbound delivery, in-transit transportation, and terminal delivery.

[0059] The method further includes: based on the encoding unit, performing the following steps: converting the spatial information encoding in the dynamic environment input sequence into a code with a hierarchical relationship, and performing spatial association and temporal sequence features on the code with a hierarchical relationship to obtain a spatiotemporal hierarchical semantic feature tensor;

[0060] The spatiotemporal hierarchical semantic feature tensor is fused with the logistics domain knowledge graph to obtain an enhanced dynamic environment input sequence.

[0061] The intelligent warehouse distribution agent, delivery vehicle scheduling agent, and route planning agent perform logical reasoning on the entire logistics order lifecycle based on dynamic environment input sequences to generate a decision chain structure covering the entire process of warehousing and outbound delivery, in-transit transportation, and terminal delivery, including:

[0062] Based on the intelligent warehouse partitioning agent, the order information and warehouse information in the enhanced dynamic environment input sequence are converted into graph structure data to obtain the warehouse partitioning decision logic sub-chain;

[0063] Based on the delivery vehicle scheduling agent, a mathematical model with time window constraints and vehicle load balancing constraints is established based on the vehicle status, order timeliness requirements, and real-time road condition data in the enhanced dynamic environment input sequence. This model is solved using proximal policy optimization to obtain the vehicle scheduling decision logic subchain.

[0064] Based on the route planning agent, a spatiotemporal graph is constructed based on the road network, delivery point locations, and real-time traffic flow information in the enhanced dynamic environment input sequence. The graph attention mechanism is used to determine the association weights between different road sections and delivery points to generate a route planning decision logic sub-chain.

[0065] Based on the warehouse division decision logic sub-chain, vehicle scheduling decision logic sub-chain, and route planning decision logic sub-chain, a decision chain structure covering the entire process of warehousing and delivery, in-transit transportation, and terminal distribution is generated.

[0066] In summary, this application constructs a multi-agent decision chain generation mechanism driven by the fusion of spatiotemporal hierarchical semantics and knowledge graphs, and achieves collaborative optimization of the entire logistics process through three-level progressive innovation. Its exemplary technical implementation is as follows:

[0067] In the construction of spatiotemporal semantic feature tensors, intelligent digital expression of geographic space is implemented. Hierarchical spatial coding is performed to achieve three-level coding of distribution areas according to administrative divisions (provinces / cities / districts), grid accuracy (e.g., 1km×1km), and road network nodes (intersections / road sections), forming a hierarchical relationship similar to "country-city-business district" and giving spatial information semantic granularity (e.g., orders within the same business district can be quickly clustered). Spatiotemporal correlation feature extraction is performed to perform temporal analysis (e.g., changes in order density in business districts during the morning rush hour) and spatial association (e.g., road network distances between warehouses and surrounding distribution points) on the encoded spatial information, generating a three-dimensional tensor structure. The spatial dimension includes regional codes and road network topology relationships; the temporal dimension captures order time windows and periodic changes in road conditions; and the semantic dimension embeds business attributes (e.g., cold chain requirements and urgency).

[0068] To this end, this application is different from the "coordinate point" expression of traditional spatial data. Through hierarchical coding and spatiotemporal semantic fusion, the geographic space is converted into a computable knowledge unit, enabling the system to understand complex relationships such as "the impact of congestion in a certain road section on the surrounding distribution points within a certain distance, such as 3km".

[0069] Secondly, in the dynamic environment sequence link enhanced by the knowledge graph, deep injection of domain knowledge is achieved. Among them, logistics domain knowledge (such as "Warehouse A only stores daily necessities" and "Vehicle B is a refrigerated truck") is converted into entity-relationship triples through knowledge graph fusion and aligned with the spatiotemporal semantic tensor: entity alignment: mapping physical entities such as warehouses and vehicles to the spatial dimensions of the tensor; relationship embedding: converting business relationships such as "service radius" and "transportation time" into the semantic dimension weights of the tensor. When generating enhanced sequences, through knowledge graph reasoning (such as "Order C is seafood → must match a refrigerated truck"), the feature weights of each dimension in the tensor are dynamically adjusted, so that the input sequence contains both real-time data (such as road conditions) and domain knowledge constraints (such as vehicle model restrictions).

[0070] To this end, this application deeply integrates symbolic knowledge reasoning with continuous data representation, upgrading the dynamic environment sequence from "pure data-driven" to "data + knowledge dual-wheel driven", significantly improving the rationality of decision-making in complex constraint scenarios.

[0071] Furthermore, when generating a multi-agent decision chain, a full-process optimization of division of labor and collaboration is performed. Specifically, when generating the warehouse-based decision subchain, a graph structure is constructed, abstracting orders and warehouses into nodes. A heterogeneous graph is constructed based on spatial distance (road network distance), attribute matching (inventory type), and business relevance (historical delivery frequency). During path search, the graph attention mechanism calculates the association weights between nodes, prioritizing warehouse-based paths with "close distance, inventory matching, and high historical efficiency." When generating the scheduling decision subchain, constraint modeling is performed: vehicle status (location, load), order timeliness (time window), and road condition prediction (travel time) are converted into mathematical constraints. During reinforcement learning, a proximal policy optimization algorithm is used to maximize vehicle utilization and order on-time performance while satisfying hard constraints (such as load limits). When generating the route planning subchain, a spatiotemporal graph is constructed: the road network is sliced ​​into time slices (e.g., every 15 minutes) to create a dynamic graph, with nodes representing intersections and edge weights representing real-time traffic conditions. Through attention mechanism optimization, the graph attention network automatically learns the importance of road sections (e.g., main roads are weighted higher than side roads), dynamically adjusting route planning. Finally, by exploring inter-subchain dependencies (e.g., the silo results determine the vehicle's starting point), a constraint propagation path from "silo → scheduling → routing" is constructed, resolving conflicts (e.g., competition for vehicle resources) and generating a globally optimal decision chain.

[0072] To this end, this application realizes the collaborative solution of heterogeneous constraints: spatial constraints (dividing warehouses), time constraints (scheduling), and dynamic constraints (road conditions) are unified into a graph neural network framework for solution; in addition, it also achieves end-to-end decision chain generation: breaking through the limitations of traditional "module optimization", achieving seamless connection of the entire process through the constraint propagation mechanism, so that local optimal decisions automatically serve the global goals.

[0073] Optionally, in a specific embodiment, the intelligent warehouse partitioning agent converts the order information and warehouse information in the enhanced dynamic environment input sequence into graph structure data to obtain a warehouse partitioning decision logic subchain, including:

[0074] Extract order information and warehouse information from the enhanced dynamic environment input sequence,

[0075] Entity nodes and relationship edges are constructed based on the extracted order information and warehouse information. Order abstractions correspond to order nodes, and warehouse abstractions correspond to warehouse nodes. The relationship edges include spatial relationship edges, attribute relationship edges, and business relationship edges. Spatial relationship edges are constructed by calculating the Euclidean distance, road network distance, and estimated travel time between the order destination and each warehouse. Attribute relationship edges are constructed by evaluating the matching degree between the order cargo type and the warehouse storage conditions. Business relationship edges are constructed by mining the historical association frequency between orders and warehouses based on historical delivery data.

[0076] Generate order-warehouse heterogeneous information graph based on constructed entity nodes and relationship edges;

[0077] Based on the order-warehouse heterogeneous information graph, supply and demand matching evaluation under time and space constraints is performed to generate a supply and demand matching evaluation graph under time and space constraints;

[0078] The optimal warehouse path search and warehouse path encoding are performed on the supply and demand matching evaluation graph under time and space constraints to generate a warehouse decision logic sub-chain. The warehouse decision logic sub-chain includes preconditions, decision actions, and confidence scores. Preconditions include order attributes and warehouse status representations, and decision actions include the representation of which warehouse to allocate to.

[0079] In summary, this embodiment achieves dynamic optimal decision-making for intelligent warehouse partitioning by constructing a heterogeneous knowledge graph with multi-dimensional semantic associations and combining spatiotemporal constraint reasoning with explainable decision coding. An exemplary explanation is as follows:

[0080] Through multi-dimensional modeling of heterogeneous information graphs, intelligent mapping goes beyond simple distance matching. Specifically, the semantic construction of ternary relationship edges is based on spatial relationship edges: This breaks the simplistic assumption of traditional Euclidean distance and simultaneously calculates road network distance (actual driving distance considering real-time road conditions) and estimated driving time (integrating historical traffic data), making spatial associations more accurate for real-world delivery scenarios. Based on attribute relationship edges, this approach semantically matches order product types (such as cold chain and fragile goods) with warehouse storage conditions (refrigerated warehouses and shockproof racks). For example, the attribute similarity between "seafood orders" and "refrigerated warehouses" is calculated using word vectors to construct weighted attribute associations. Based on business relationship edges, the frequency of collaboration between orders and warehouses is mined based on historical delivery data. For example, "high-frequency order-warehouse pairs" are assigned higher association weights, reflecting the habitual optimization of business processes. Furthermore, to achieve a unified representation of heterogeneous nodes, orders and warehouses are abstracted into heterogeneous graph nodes. Order nodes carry dynamic attributes such as timeliness requirements and product attributes, while warehouse nodes integrate static attributes such as inventory status and processing capacity. A graph neural network is used to enable feature interaction across node types.

[0081] To this end, this embodiment integrates and models three types of heterogeneous relationships: spatial physical distance, business attribute matching, and historical collaboration frequency. This upgrades the warehouse division decision from the "geographical proximity principle" to "multi-dimensional semantic optimization," solving the problem of low warehousing efficiency caused by "close distance but mismatched attributes" in traditional warehouse division.

[0082] Intelligent reasoning in dynamic environments is performed through supply-demand matching evaluation under spatiotemporal constraints. Specifically, based on the graph structure embedding of spatiotemporal constraints, a time window constraint is constructed: the latest order pickup time and the warehouse's operating hours are converted into time weights for graph edges. For example, if a warehouse's available processing capacity within the order's time window is insufficient, its matching priority is automatically lowered. Spatial capacity constraints are constructed: combining the warehouse's real-time inventory with order demand, the capacity status of each node is updated through a graph propagation mechanism. For example, when warehouse A is running low on inventory, the weight of its associated edges is automatically transferred to backup warehouse B. Furthermore, through the dynamic calculation of supply-demand matching, a multi-dimensional matching function is designed, integrating spatial distance scores, attribute matching scores, and time window scores. The weights of each dimension are dynamically adjusted through an attention mechanism (for example, more emphasis is placed on time scores during peak periods), generating a matching evaluation graph with spatiotemporal constraints.

[0083] To this end, this embodiment upgrades the static matching in traditional warehouse division to "dynamic reasoning with real-time environmental awareness" through the dynamic propagation of the graph structure of constraint conditions. For example, when a sudden order is received, it can automatically evaluate the remaining spatiotemporal resources of each warehouse to avoid "time default caused by blind allocation."

[0084] In this application, the transition from black-box optimization to transparent decision-making is achieved based on the generation of interpretable decision logic sub-chains. For example, when searching for the optimal path through multiple objectives, a Monte Carlo tree search algorithm is used based on a spatiotemporal matching graph, with "delivery distance + attribute matching + historical efficiency" as a comprehensive reward function to search for a distribution path that takes into account both efficiency and reliability. For example, warehouses with "moderate distance but complete attribute matching" are given priority over warehouses with "nearest but some attributes are not supported". When performing structured coding of decision logic, the basic premise is: the matching conditions of order attributes (goods type, timeliness requirements) and warehouse status (inventory, processing capacity) are encoded as logical expressions, such as "IF the order is cold chain AND the warehouse has cold storage AND inventory ≥ order quantity THEN"; when generating decision actions: the specific warehouse allocation result is clearly pointed to, and a confidence score (calculated based on historical matching accuracy) is attached, such as "allocated to warehouse C, confidence level 92%".

[0085] To this end, by converting the optimization results into a logical rule chain with confidence, we break through the black box characteristics of traditional reinforcement learning decision-making, and make the warehouse allocation decision have the full process explainability of "condition-action-reliability", which is convenient for manual verification and business adjustments.

[0086] Optionally, in a specific embodiment, the delivery vehicle scheduling agent establishes a mathematical model including time window constraints and vehicle load balancing constraints based on the vehicle status, order timeliness requirements, and real-time road condition data in the enhanced dynamic environment input sequence, and solves it using proximal strategy optimization to obtain a vehicle scheduling decision logic subchain, including:

[0087] Extract vehicle status, order timeliness requirements, and real-time road condition data from enhanced dynamic environment input sequences;

[0088] Map vehicle status data to the road network topology and calculate the spatiotemporal accessibility matrix between the vehicle and each delivery point;

[0089] Decompose the time series of order timeliness requirements to obtain deterministic trend components and random fluctuation components;

[0090] Based on real-time traffic data and historical traffic data, predict road travel time in future periods;

[0091] Based on the spatiotemporal accessibility matrix, deterministic trend components and random fluctuation components, the future road travel time is predicted to generate a spatiotemporal constraint feature tensor;

[0092] Based on the spatiotemporal constraint feature tensor, decision variables are defined, objective functions are designed, and constraints are constructed to establish a mathematical model that includes time window constraints and vehicle load balancing constraints;

[0093] Based on the designed constraint-aware action mask, actions that violate hard constraints are prohibited. A policy network and a value network are constructed to solve the mathematical model containing time window constraints and vehicle load balancing constraints using proximal policy optimization to obtain the state-action mapping relationship and decision constraint propagation path.

[0094] Based on the state-action mapping relationship and the decision constraint propagation path, a vehicle scheduling decision logic sub-chain is generated.

[0095] In summary, in this embodiment, by constructing a fusion framework of spatiotemporal constraint feature tensors and constraint-aware reinforcement learning, dynamic optimization and explainability of vehicle scheduling decisions are achieved. The specific exemplary technical implementation is as follows:

[0096] Modeling the spatiotemporal characteristics of multi-source data enables intelligent conversion from discrete data to constraint tensors. This involves mapping road network topology and calculating accessibility: Status data such as vehicle location and available time are mapped to the road network topology. The Dijkstra algorithm is then used to calculate the spatiotemporal accessibility between vehicles and delivery points at different time slices (e.g., "Vehicle V1 departs from the warehouse at 10:00 and reaches delivery point A before 10:30"). This creates a two-dimensional accessibility matrix, transcending traditional GPS coordinate planar distance calculations and incorporating the actual traffic rules of the road network. Time series decomposition of timeliness requirements is performed: Empirical Mode Decomposition (EMD) is used to decompose order timeliness windows into deterministic trends (e.g., fixed delivery times during the weekday morning rush hour) and random fluctuations (e.g., time deviations caused by sudden congestion). This allows the system to distinguish between "predictable timeliness" and "unpredictable interference," providing refined time constraints for dynamic scheduling. Generate spatiotemporal constraint feature tensors: Organize the accessibility matrix, time decomposition results, and road condition prediction data into tensors based on the three dimensions of "vehicle-order-time". For example, the tensor element T[i,j,t] represents "the accessibility and time matching degree of vehicle i for order j at time t", achieving a unified mathematical representation of multi-source constraints.

[0097] To this end, this embodiment uses a combination of road network topology mapping + time series decomposition + three-dimensional tensor construction technology to transform the "static table data" in traditional scheduling into a "constraint feature body" containing dynamic spatiotemporal associations, thereby upgrading the matching constraints between vehicles and orders from "two-dimensional planar relationships" to "three-dimensional spatiotemporal relationships."

[0098] In addition, constraint-aware reinforcement learning modeling is also carried out to achieve the coordination of hard constraint embedding and decision optimization, in which the mathematical model is constructed in multiple dimensions: decision variables: define the binary variable x[i,j] to represent "whether vehicle i is assigned to order j", and the continuous variable t[i,j] to represent the delivery start time, and construct a discrete-continuous hybrid decision space; objective function: take into account both the minimization of total delivery time and the balancing of vehicle loads, for example, avoiding overloading of a vehicle through variance minimization constraints; constraints: explicitly embed hard constraints such as the time window (order j must be delivered within [t_start, t_end]) and load balancing (total load of vehicle i ≤ rated load).

[0099] Constraint-aware action mask mechanism: A binary mask matrix is ​​designed to directly prohibit the selection of an action (such as assigning an order to a vehicle) when it violates hard constraints (such as time window overlap or load overload). This ensures that invalid solutions are not generated during the reinforcement learning process and accelerates convergence.

[0100] Perform proximal policy optimization (PPO) solution: Construct the policy network π(a|s) and the value network V(s), and prevent the policy update from being too large by truncating the objective function. For example, when the new policy improves load balancing but increases time consumption, the update step size is automatically adjusted to balance multiple objectives.

[0101] To this end, this application solves the problem of traditional reinforcement learning "easy to generate invalid solutions" in logistics scheduling by deeply integrating hard constraint preprocessing (action mask) with reinforcement learning optimization. At the same time, through multi-objective function design, the scheduling decision is upgraded from "single efficiency optimization" to "efficiency + balance" multi-dimensional optimization.

[0102] The process of generating a constraint propagation path from a black-box strategy through an interpretable decision logic subchain is carried out. This involves state-action mapping and constraint propagation recording: During the PPO optimization process, the mapping relationship between each state (e.g., vehicle load, order timeliness) and action (allocation decision) is recorded, while the constraint propagation path is tracked (e.g., "Order A's timeliness constraint → affects vehicle B's route selection → triggers load balancing adjustments") to form a decision dependency tree. Structured encoding of the logic subchain is performed: the mapping relationship and constraint path are converted into a logical chain of "preconditions - decision actions - constraint verification." For example: Precondition: "Vehicle V2's current load is ≤70% and the timeliness window for order B is 11:00-12:00"; Action: "Assign order B to V2"; Verification: "The estimated time from V2's current location to B is 25 minutes, which meets the timeliness requirement."

[0103] Therefore, this embodiment transforms the black-box strategy generated by reinforcement learning into an interpretable chain with logical causal relationships through the constraint propagation record of the decision-making process, solving the problem of traditional scheduling algorithms of "knowing the results but not the reasons behind them" and facilitating manual intervention and exception handling.

[0104] Optionally, in a specific embodiment, the route planning agent constructs a spatiotemporal graph based on the road network, delivery point locations, and real-time traffic flow information in the enhanced dynamic environment input sequence, and determines the association weights between different road segments and delivery points through a graph attention mechanism to generate a route planning decision logic subchain, including:

[0105] The road network is abstracted as an undirected weighted graph, and the attributes of each node and each edge are defined. Node attributes include coordinate location, intersection type (such as crossroads, T-junctions), and traffic control devices (traffic lights, roundabouts). Edge attributes include length, number of lanes, speed limit, and road type (highway, urban road).

[0106] The undirected weighted graph is expanded in time and space dimensions to obtain a time and space graph. The distribution point locations are mapped to the time and space graph and real-time traffic flow information is added to each edge to obtain a time and space dynamic road network graph.

[0107] Determine the static and dynamic features of the spatiotemporal dynamic road network graph and calculate the attention weight between the static and dynamic features;

[0108] Based on the attention weights between static and dynamic features, the association weights between different road sections and delivery points are calculated;

[0109] The road condition response strategy is encoded according to the association weights between different road sections and delivery points, forming a route planning decision logic sub-chain.

[0110] In summary, this embodiment and this technical solution achieve intelligent understanding of complex road conditions and dynamic path planning by constructing a fusion model of spatiotemporal dynamic road network graph and graph attention mechanism. An exemplary description is as follows:

[0111] Multidimensional modeling of spatiotemporal dynamic road network graphs is implemented to upgrade from static topology to dynamic semantics. This involves attributed abstraction of undirected weighted graphs: the road network is abstracted into an undirected graph of nodes (intersections) and edges (road segments). Node attributes include coordinates, intersection type (e.g., crossroads), and traffic control devices (traffic lights), while edge attributes encompass static features such as length, number of lanes, and speed limit, forming a structured road network topology model. This abstraction transcends the fragmented representation of traditional GPS trajectories, enabling a semantic description of the road network based on the "node function - road segment attributes" model. Spatiotemporal dimension expansion and dynamic information injection are implemented to achieve temporal dimension expansion: the road network graph is layered into time slices (e.g., 15-minute intervals) to form a spatiotemporal cube structure. Each edge carries dynamic attributes such as real-time traffic flow and congestion index in different time slices. Semantic mapping of delivery points is performed: the coordinates of delivery points are mapped to the nearest road network node, and attributes such as access time and service duration are added to deeply integrate delivery demand with the road network topology.

[0112] To this end, through three-dimensional modeling of static topology attribution + spatiotemporal dimension expansion + dynamic information injection, the "static map" in traditional navigation is transformed into a dynamic semantic network that includes "spatial structure-temporal changes-business needs", providing richer environmental representation for path planning.

[0113] In addition, the dynamic weight calculation of the graph attention mechanism is the core of adaptive road condition understanding. When decoupling static and dynamic features, static features: extract inherent attributes such as road section length, speed limit, and traffic control type of the node (such as the traffic light cycle); dynamic features: real-time traffic flow, sudden congestion events and other time-varying attributes, and extract time series features (such as the average traffic speed in the past 30 minutes) through a sliding window. When performing cross-dimensional attention weight calculation, a bidirectional attention mechanism is designed to make static and dynamic features mutually reinforce each other. Among them, static guides dynamic: the high static weight of the main road will enhance its dynamic attention during congestion (such as prioritizing the calculation of the real-time traffic impact of the main road); and then dynamically corrects static: the dynamic characteristics of sudden congestion will temporarily increase the weight of the corresponding road section, overriding its static weight (such as the weight of the branch road that is usually unobstructed will be increased when the main road is congested).

[0114] Finally, the association weights between road sections and delivery points are generated: through the multi-head attention mechanism, the road network topological distance, delivery time requirements, and real-time road conditions are comprehensively considered to calculate the association weights between any road section and delivery point. For example, "the comprehensive weight of road section A to delivery point B during the morning rush hour = 0.4 × static distance + 0.5 × dynamic congestion coefficient + 0.1 × time urgency".

[0115] To this end, this application avoids the limitations of "fixed weights or single dynamic weights" in traditional path planning, and realizes adaptive weighting of road network features through a cross-dimensional attention mechanism, so that the system can automatically focus on high-impact factors (such as congested sections during peak hours and sections that must be passed for urgent orders), solving the problem that "static rules cannot cope with dynamic changes".

[0116] The logical coding of the road condition response strategy is implemented from weight calculation to explainable decision-making. Among them, weight-driven strategy generation is used to design multi-scenario response strategies based on associated weights: Normal scenario: select the path with the "minimum sum of weights" (taking into account distance and road conditions); Emergency scenario: increase the time-related weights to prioritize on-time delivery of orders; Abnormal scenario: when a sudden increase in the weight of a congested road section is detected, the search for alternative paths is automatically triggered.

[0117] Perform structured coding of the decision logic sub-chain and convert the strategy into a "condition-action" rule chain, for example: Premise: "The dynamic congestion weight of section X is greater than the threshold and the weight difference of alternative section Y is ≤15%"; Action: "Switch to section Y"; Verification: "Expected time saving is 5 minutes".

[0118] Therefore, it can be seen that this application converts the black box calculation results of the attention mechanism into interpretable decision rules with confidence through the three-layer conversion of weight-strategy-logic chain, so that path planning is upgraded from "optimal solution output" to "interpretable intelligent decision", which facilitates manual intervention and strategy optimization.

[0119] Optionally, in a specific embodiment, the decision chain structure covering the entire process of warehousing and delivery, in-transit transportation, and terminal delivery is generated based on the warehouse division decision logic sub-chain, the vehicle scheduling decision logic sub-chain, and the route planning decision logic sub-chain, including:

[0120] Mining the dependencies between the warehouse division decision logic sub-chain, vehicle scheduling decision logic sub-chain, and route planning decision logic sub-chain to obtain the warehouse division → scheduling → route dependency link;

[0121] Conflict detection is performed on the warehouse distribution → scheduling → route dependency link to determine hard constraint conflicts and soft constraint conflicts;

[0122] Based on the conflict negotiation optimization framework, hard constraint conflicts and soft constraint conflicts are determined to encode collaborative optimization strategies;

[0123] Based on the collaborative optimization strategy coding, the warehouse division decision logic sub-chain, vehicle scheduling decision logic sub-chain, and route planning decision logic sub-chain are adjusted;

[0124] The adjusted warehouse distribution decision logic sub-chain, vehicle scheduling decision logic sub-chain, and route planning decision logic sub-chain are topologically sorted to generate a decision chain structure covering the entire process of warehousing, in-transit transportation, and terminal distribution.

[0125] In summary, the technical solution of this embodiment integrates the independent warehouse division, scheduling, and route planning decision-making sub-chains into a globally optimal full-process decision-making chain by building a closed-loop mechanism for sub-chain dependency mining and conflict collaborative optimization. The technical example is explained as follows:

[0126] Subchain dependency mining is conducted to build business logic-driven decision chains. This involves identifying causal dependencies: By analyzing logistics business processes, the hard causal relationships between the warehouse distribution, scheduling, and route planning subchains are determined. For example, "the warehouse distribution results determine the starting point for vehicle pickup, which in turn affects scheduling and route planning," forming the basic dependency chain of "warehouse distribution → scheduling → route." Data association analysis is performed: cross-data items from each subchain's decision results are extracted (such as the warehouse coordinates determined by the warehouse distribution and the vehicle starting point coordinates in the scheduling). The correlation of the dependency chain is strengthened through data mapping relationships. For example, "whether the location of warehouse A in the warehouse distribution is within the service radius of the scheduled vehicle" is verified.

[0127] To this end, this embodiment overcomes the limitations of independent optimization of each link in traditional logistics decision-making. Through business causal analysis + data association verification, it builds a cross-link dependency chain, so that local decisions automatically serve the global goal and solves the coordination failure problem of "optimal warehouse distribution but inefficient scheduling".

[0128] Conflict detection and classification are performed to achieve intelligent identification of multi-dimensional constraints, including hard constraint conflict detection: defining hard rules that cannot be violated (such as the upper limit of vehicle load and the latest delivery time of orders), such as checking "whether the total load of scheduled vehicles exceeds the rated load" and "whether the estimated arrival time of route planning exceeds the order time window." Soft constraint conflict detection: identifying negotiable flexible goals (such as minimizing transportation costs and minimizing vehicle idle rate), such as analyzing whether "selecting a distant warehouse for sub-warehouses leads to increased transportation costs" exceeds the preset threshold. Conflict propagation analysis: tracking the scope of conflict impact by relying on links. For example, "insufficient inventory in sub-warehouse warehouse A" will trigger a chain reaction of "scheduled vehicles need to detour to pick up goods → increased route planning time → order time limit breach."

[0129] To this end, this embodiment upgrades the "single-point conflict detection" in traditional logistics to "full-link chain conflict identification" through hard and soft constraint classification + conflict propagation tracking, thereby discovering potential contradictions hidden in each sub-chain in advance and avoiding systemic risks during decision-making execution.

[0130] Collaborative optimization strategy encoding is performed to implement a constraint-driven decision-making and adjustment mechanism. This includes designing a conflict negotiation framework: Hard constraints are prioritized: For hard conflicts such as overload and time limit violations, subchain re-optimization (such as re-allocation of warehouses or adjustment of shift schedules) is immediately triggered. Soft constraint negotiation: For soft conflicts such as rising costs and high idle rates, a multi-objective optimization algorithm (such as a genetic algorithm) is used to find a balance. For example, between "increasing warehouse costs by 10%" and "reducing transportation costs by 15%," the solution with the best overall benefits is chosen. Strategy encoding and execution are performed: Optimization rules are converted into executable logic code, such as "IF warehouse allocation results in a scheduling time limit conflict THEN trigger warehouse allocation reselection and update of shift schedules," ensuring that conflict resolution solutions are automatically integrated into each subchain.

[0131] To this end, this embodiment combines hierarchical processing of hard and soft constraints with strategy logic coding to build an intelligent response mechanism of "conflict detection-rule matching-automatic adjustment", enabling the system to "prioritize the resolution of key contradictions and then optimize secondary goals" like logistics experts, thereby improving the rationality and efficiency of decision-making.

[0132] Perform topological sorting and decision chain generation to achieve time-series global decision integration. This involves topological sorting of dependencies: Based on the dependency chain of "warehouse allocation → shift scheduling → route," use the Kahn algorithm to topologically sort the adjusted subchains to ensure the order of decision execution complies with the business process (e.g., warehouse allocation first, shift scheduling, and finally route planning). Build a full-process decision chain: connect the sorted subchains in series according to business timeline. Each subchain serves as a link in the decision chain, retaining its preconditions, decision actions, and confidence scores, forming a complete decision logic chain from warehouse delivery to final delivery.

[0133] This embodiment uses a topological sorting algorithm to transform decentralized sub-chains into time-series decision chains, giving full-process decisions a clear logical context and execution sequence. This overcomes the problem of "parallel processing of various links leading to time sequence confusion" in traditional logistics decision-making and achieves true full-process collaboration.

[0134] Optionally, in a specific embodiment, the full-process path planning simulation is implemented based on the decision chain structure to generate an initial path plan including multi-resource collaborative scheduling, including: based on the decision chain structure covering the entire process of warehousing and outbound delivery, a full-process path planning simulation from collection and loading, trunk line transportation to terminal delivery is implemented, and an initial distribution path plan including multi-vehicle collaborative distribution and order consolidated transportation is preliminarily determined.

[0135] In summary, the technical solution of this embodiment transforms abstract decision logic into an executable multi-resource collaborative scheduling solution by constructing a full-process dynamic simulation framework based on the decision chain. The exemplary technical implementation is described as follows:

[0136] A full-process dynamic simulation driven by a decision chain is implemented from logical reasoning to execution rehearsal. Specifically, the decision chain is mapped to the execution process: the warehouse distribution, scheduling, and route planning decision sub-chains are mapped to the control nodes of the simulation process according to the business time sequence (warehouse outbound → in-transit transportation → terminal delivery). For example, the warehouse location determined by the warehouse distribution decision is directly used as the starting point for collection and loading, and the vehicle scheduling plan generated by the scheduling decision drives the allocation of vehicles in the trunk transportation stage. Dynamic deduction is performed in the time and space dimensions: Time dimension: The time consumption of each link is simulated in minute-level time slices, such as the time consumption of collection and loading, the travel time of the trunk transportation section, and the service time of the terminal delivery, to ensure that the simulation time is synchronized with the actual business; Spatial dimension: Based on the road network topology and the location of the distribution point, the vehicle movement trajectory and the spatial transfer process of the goods are simulated. For example, the route of a multi-vehicle collaborative fleet is simulated in the trunk transportation stage.

[0137] To this end, this embodiment overcomes the limitations of traditional path planning's "single-point optimization" and, through bidirectional mapping of the decision chain and the execution process, achieves a seamless transition from the logical reasoning of "warehouse distribution - scheduling - routing" to the physical execution of "loading - transportation - delivery." This allows the simulation process to reflect both the logical constraints of the decision chain and the spatiotemporal dynamics of real business.

[0138] A simulation mechanism for multi-resource collaborative scheduling is used to achieve globally optimized resource allocation. Specifically, when unified modeling of multi-resource states is performed, vehicle resources: simulate vehicle load, fuel consumption, real-time location and other states to support multi-vehicle collaborative scheduling (such as multiple vehicles traveling along a convoy route in trunk transportation); order resources: group orders according to cargo type and timeliness requirements to support order merging and transportation (such as multiple orders in the same area are delivered by the same vehicle); warehousing resources: simulate warehouse outbound efficiency and inventory changes to ensure that the collection and loading links match the warehousing operation capacity.

[0139] Embed collaborative scheduling strategies: Multi-vehicle collaboration: Based on the vehicle grouping results of the scheduling decision sub-chain, simulate the path collaboration of multiple vehicles in trunk transportation (such as vehicle A and vehicle B exchanging goods at an intermediate intersection); Order merging: Based on the distribution area division of the route planning sub-chain, merge and load orders on the same route to reduce the number of vehicle stops.

[0140] To this end, this embodiment upgrades the simple matching of "single vehicle-single order" in traditional path planning to global resource collaboration of "multiple vehicles-multiple orders-warehouses" through unified modeling of multiple resource states + dynamic embedding of collaborative strategies, solving the configuration inefficiency problem caused by resource fragmentation in traditional simulation.

[0141] Generate the logic of the initial path plan: achieve a balance between constraint satisfaction and efficiency, and prioritize hard constraints: strictly follow the hard constraints in the decision chain (such as vehicle load limits and order time windows) during the simulation process. For example, when order mergers cause vehicle overload, automatically split the order to other vehicles. Perform multi-objective optimization balance: on the basis of satisfying hard constraints, optimize soft goals (such as the shortest total driving distance and the lowest vehicle empty rate). For example, use the simulated annealing algorithm to search for the optimal route combination for collaborative distribution of multiple vehicles to maximize the connection efficiency between trunk transportation and terminal delivery. Perform structured output of the plan: convert the simulation results into a structured plan that includes "vehicle route-order allocation-time node", such as each vehicle's driving route map, loading order list, and expected arrival time stamp of each node.

[0142] To this end, this embodiment adopts a three-layer mechanism of hard constraint pre-verification + multi-objective optimization + structured output to ensure that the generated initial plan not only meets the logical constraints of the decision chain but also has the feasibility of actual implementation, breaking through the defect of "theoretical optimality but practical impracticality" in traditional simulation.

[0143] Optionally, in a specific embodiment, the environment-aware intelligent agent executes the initial delivery path plan in an actual logistics delivery scenario, and loads real-time environmental change information to determine the weight of the logistics decision chain, including:

[0144] Extracting features from real-time environmental data to obtain environmental feature data;

[0145] The initial delivery path plan is put into actual logistics distribution scenarios for execution and the interference factors of environmental characteristic data on the initial delivery path plan are dynamically analyzed;

[0146] Calculate path execution deviation data based on the interference factor;

[0147] Based on the constructed factor importance judgment matrix and the path execution deviation data, the logistics decision chain weight is calculated.

[0148] In summary, the technical solution of this embodiment realizes adaptive optimization of the logistics decision chain to the dynamic environment by constructing a closed-loop feedback mechanism of "environmental perception-interference analysis-deviation calculation-weight calibration". An exemplary description is as follows:

[0149] Intelligent extraction of real-time environmental features enables semantic transformation of multi-source dynamic data. This multi-source data fusion feature extraction involves acquiring real-time road condition data (congestion index, accident location) from the traffic monitoring system, extracting emergency order information (urgency, cargo type) from the order system, and collecting vehicle status (location, fuel consumption) from onboard sensors. A sliding window algorithm is then used to extract temporal features (such as the average speed over the past 15 minutes) and spatial features (such as the geographic coordinates of the congested road section). Interference feature semantic encoding is performed, converting raw data into environmental feature vectors. For example, encoding "congestion on a certain road section" as [location, congestion level, expected duration] makes environmental changes business-interpretable.

[0150] To this end, this embodiment converts fragmented environmental information into structured feature vectors through simultaneous multi-source data collection and semantic encoding, breaking through the limitations of "scattered data and ambiguous semantics" in traditional perception and providing accurate input for interference analysis.

[0151] Quantitative analysis of dynamic interference factors is performed to achieve real-time assessment of environmental impacts. Interference factor modeling: Interference factors are defined as the degree of influence of environmental characteristics on the initial path plan. For example: Road condition interference factor = actual travel time / expected travel time - 1 (reflecting the delay multiple caused by congestion); Order interference factor = number of sudden orders / initial number of orders (reflecting the impact of order fluctuations on the delivery plan).

[0152] Perform multi-dimensional interference mapping: associate interference factors with key nodes of the route plan (such as road sections and delivery points). For example, the congestion interference factor of a certain road section directly affects the travel time prediction of the corresponding road section.

[0153] To this end, this embodiment achieves the transformation of environmental changes from "qualitative description" to "quantitative impact" through quantitative modeling and multi-dimensional mapping of interference factors, enabling the system to accurately identify "which environmental factors have the greatest impact on path execution."

[0154] Dynamic calculation of route execution deviations allows for tracing discrepancies between planned and actual delivery. Deviations are calculated in the following time and space dimensions: Time deviation compares the initial plan's estimated arrival time with the actual arrival time, calculating the time delay at each delivery node; and spatial deviation analyzes the degree of deviation between the vehicle's actual route and the planned route, such as the additional distance driven by detours due to congestion. Multi-objective deviation integration is performed: Timeliness deviation, cost deviation (increased fuel consumption), and service quality deviation (order timeouts) are combined to generate a multidimensional deviation vector that comprehensively reflects the impact of environmental interference.

[0155] To this end, this application breaks through the limitation of "single time dimension" in traditional deviation analysis through spatiotemporal multi-objective deviation integration, so that deviation data can fully reflect the impact of environmental changes on the entire logistics process.

[0156] Intelligent calibration of decision chain weights enables deviation-based priority redistribution. This involves constructing a factor importance judgment matrix: Based on logistics domain knowledge, a pairwise comparison matrix of influencing factors is established. For example, the importance of "road conditions" to route execution is 1.5 times that of "order factors," forming a hierarchical weight judgment framework. Data-driven iterative weight calculations are performed: Path deviation data is substituted into the judgment matrix, and the dynamic weight of each decision chain factor is calculated through weighted summation. For example, if congestion causes 40% of the deviation during a certain period, the weight of "road conditions" is automatically increased.

[0157] To this end, through the knowledge-driven judgment matrix and data-driven weight iterative fusion mechanism, the weight calibration not only incorporates the experience of logistics experts, but can also be dynamically adjusted according to real-time deviations, solving the problem of traditional weight settings being "statically fixed or purely data-driven and prone to overfitting."

[0158] Optionally, in a specific embodiment, the logistics strategy agent optimizes the initial distribution path plan based on the logistics decision chain weight to obtain a logistics business knowledge graph to construct a logistics two-path collaborative decision network, including:

[0159] Map the logistics decision chain weights to each link of the initial distribution path plan, optimize the initial distribution path plan based on the logistics decision chain weights, and obtain the weight-optimized distribution path plan;

[0160] Extract knowledge and construct graph elements from the weight-optimized delivery path plan. Structuralize the extracted knowledge to obtain a logistics business knowledge graph. The delivery nodes are abstracted as entities in the knowledge graph, and the relationships between nodes are abstracted as edges in the graph. Entities and edges are assigned corresponding attributes.

[0161] Determine the state space and action space reward functions based on the logistics business knowledge graph;

[0162] Generate a two-way logistics thinking chain path based on the state space and action space reward functions;

[0163] Disassemble the logistics dual-path thinking chain path to obtain order allocation strategy, vehicle scheduling rules, and real-time exception response logic;

[0164] The order allocation strategy, vehicle scheduling rules, and real-time exception response logic are mapped to the order management system, vehicle scheduling system, and real-time monitoring system respectively to build a logistics dual-path collaborative decision-making network.

[0165] In summary, the technical solution of this embodiment achieves the upgrade of logistics decision-making from data-driven to knowledge-integrated by building an intelligent closed loop of "weight-oriented optimization - knowledge graph construction - dual-path thinking chain generation - system mapping collaboration". An exemplary description is as follows:

[0166] Perform weight-oriented dynamic path optimization to achieve data-driven decision priority calibration. This includes implementing a cross-link weight mapping mechanism: decomposing logistics decision chain weights (such as road condition impact weights and time efficiency weights) by granularity based on path links (warehousing and outbound delivery, trunk line transportation, and terminal delivery). For example, mapping a "road condition weight of 0.4" to the travel time calculation module for each section of the path allows high-weight factors to prioritize the direction of path optimization. Perform multi-objective optimization reallocation: Reconstruct the optimization objective function based on weights. For example, automatically increasing the "time efficiency weight" ratio during peak hours allows the path plan to prioritize meeting the on-time delivery requirements of orders, rather than simply pursuing the shortest transportation distance.

[0167] To this end, this application overcomes the limitations of the "fixed objective function" in traditional path optimization and realizes "adaptive adjustment of decision priorities in different scenarios" through dynamic weight cross-link mapping. For example, in heavy rain weather, the "road condition weight" is automatically increased to avoid high-risk sections.

[0168] The construction of a logistics business knowledge graph enables the upgrade from path data to domain knowledge. This involves knowledge extraction and graph modeling: Entity abstraction: abstracting distribution nodes (warehouses, distribution points) into knowledge graph entities, assigning attributes such as coordinates and capacity; relationship construction: abstracting the distribution paths between nodes into relationship edges, including attributes such as distance and estimated time, while also mining implicit relationships (such as the business association of "a certain warehouse often delivers orders to a certain area"). Structured knowledge processing: storing entity-relationship-attribute triples in a graph database to form a reasonable logistics domain knowledge network. For example, the rule "cold chain warehouse → refrigerated trucks must be used" can be directly retrieved through graph relationships.

[0169] To this end, this embodiment proposes a path data-driven knowledge graph construction method to transform the "fragmented path records" in traditional logistics into "structured domain knowledge", enabling the system to automatically extract reusable decision rules from historical optimization results.

[0170] Generate a two-way thinking chain to achieve collaborative reasoning between knowledge and data. The state space and reward function are defined as follows: define the state space based on the knowledge graph (e.g., "warehouse A inventory status + vehicle B location + order C timeliness"); design a multi-dimensional reward function (timeliness achievement reward, cost-saving reward), and convert the business rules in the knowledge graph (e.g., "cold chain delay penalty") into reward / penalty factors. Construct a two-way thinking chain: a data-driven path: learn the mapping from state to action through a reinforcement learning algorithm (e.g., PPO) to generate an optimal decision chain based on historical data; a knowledge-driven path: use the knowledge graph for logical reasoning (e.g., "Order D is for fresh produce → must match a refrigerated truck → retrieve available refrigerated truck resources") to generate an explainable decision logic chain.

[0171] To this end, this embodiment achieves deep collaboration between knowledge graphs and reinforcement learning. The data-driven path solves adaptive decision-making in complex scenarios, and the knowledge-driven path ensures that decisions comply with business rules, forming a dual guarantee mechanism of "efficiency + reliability".

[0172] We conduct strategy decomposition and system mapping to achieve the transition from decision-making logic to business execution. Specifically, we deconstruct the thought chain strategy: breaking down the two-way thought chain into executable business policies. For example, we implement the order allocation strategy: "Prioritize high-priority orders to the nearest warehouse with sufficient inventory"; and the real-time exception response logic: "Detect road congestion → Trigger a knowledge graph to retrieve alternative routes and then use reinforcement learning to evaluate the efficiency of the new route." We also conduct system-level mapping and integration: embedding policies into business systems through APIs: the order management system executes the allocation policy; the vehicle dispatch system analyzes the dispatch rules; and the real-time monitoring system integrates the exception response logic, forming a closed decision-execution loop.

[0173] To this end, this embodiment scheme implements a semantic mapping method based on decision logic to business system, converting abstract thought chains into execution rules that can be understood by each system, breaking through the bottleneck of "data interoperability but logical separation" between traditional intelligent algorithms and business systems.

[0174] The following is an illustrative explanation of the technical steps mentioned in the above solution from the perspective of providing algorithm implementation.

[0175] 1. Spatiotemporal Alignment and Wavelet Denoising

[0176] 1. Unified transformation of space-time coordinates

[0177] The transformation formula for mapping spatial coordinates from different sources (such as GPS, inertial navigation) to a unified geographic coordinate system (WGS84) is:

[0178] P WGS84 =R·P local +T

[0179] P local : Coordinate vector in the original coordinate system (such as the coordinates of a certain order delivery point). P WGS84 : Coordinate vector converted to the WGS84 coordinate system. R: Rotation matrix, used to correct for angular deviations between different coordinate systems (such as the difference in reference between GPS and inertial navigation). T: Translation vector, used to eliminate offsets in the coordinate origin (such as the offsets between different map services).

[0180] 2. Time synchronization calibration

[0181] Use the NTP protocol to achieve time stamp unification:

[0182] t sync =t raw +Δt offset +Δt delay

[0183] t raw : Original timestamp (such as order timeliness record, vehicle status collection time). Δt offset : Device clock offset (calibrated by NTP server). Δt delay : Network transmission delay (estimated based on round-trip time RTT).

[0184] 3. Wavelet transform multi-resolution denoising

[0185] Perform discrete wavelet decomposition on the original logistics data flow x(t):

[0186]

[0187] ψ j,k (t): Wavelet basis function, corresponding to high-frequency noise components (such as GPS signal jumps).J,k (t): Scale basis function, corresponding to low-frequency real signal (such as road condition trend). j,k : Detail coefficient, through the threshold function Denoising. J,k : Approximation coefficient, retaining the trend characteristics of the original signal. λ: Adaptive threshold (e.g., based on noise standard deviation estimation).

[0188] 2. Spatiotemporal Semantic Tensor Modeling of Dynamic Environment Sequences

[0189] 1. Space-time grid coding and tensor construction

[0190] The spatial encoding based on the GeoHash algorithm and the time slice are combined into a three-dimensional tensor:

[0191] T i,j,t =f(road_speed i,j,t ,order_density i,j,t ,vehicle_num i,j,t )

[0192] i,j: Spatial grid index (e.g. 10m×10m grid number). t: Time window index (e.g. 15-minute time slice). road_speed i,j,t : The average speed of grid (i, j) at time t. order_density i,j,t : The order density of grid (i,j) at time t. vehicle_num i,j,t : The number of available vehicles in grid (i, j) at time t.

[0193] 2. Tensor Enhancement for Knowledge Graph Injection

[0194] The knowledge graph triple (e s ,r,e o )Embedding tensor:

[0195] e s ,e o : Knowledge graph entity (e.g., “warehouse A”, “refrigerated truck B”). r: Entity relationship (e.g., “service radius”, “load limit”). KG_Embedding: Embedding function that maps triples to vectors (e.g., TransE model).

[0196] 3. Mathematical Model for Multi-Agent Decision Chain Generation

[0197] 1. Heterogeneous Graph Optimization for Warehouse Splitting

[0198] Calculation of node association weights in the order-warehouse heterogeneous graph:

[0199] αo,w =σ(W1·feat o +W2 feat w +W3·d o,w )

[0200] o: order node, w: warehouse node. feat o : Order feature vector (time requirement, cargo type). w : Warehouse feature vector (inventory status, storage conditions). d o,w : The road network distance between the order and the warehouse (considering real-time traffic conditions). W1, W2, W3: Trainable weight matrices. σ: Activation function (such as Softmax).

[0201] 2. PPO Optimization Model for Vehicle Scheduling

[0202] The objective function of proximal policy optimization is:

[0203]

[0204] θ: Policy network parameters, θ old : Old policy parameters. t :Status (vehicle load, order time window). a t : Action (order allocation decision). t (θ): Probability ratio of new and old strategies. Advantage function estimation (such as GAE algorithm calculation). ∈: truncation parameter (such as 0.2), limiting the policy update range.

[0205] 3. Graph Attention Mechanism for Route Planning

[0206] The association weights between road sections and delivery points in the space-time graph:

[0207] e i,j =LeakyReLU(a T ·[W·h i ‖W·h j ])

[0208]

[0209] h i ,h j : Feature vector of node i, j (coordinates, road conditions, etc.). W: Feature transformation matrix, a T : attention weight vector. The set of neighbor nodes of node i. i,j : Unnormalized attention score, α i,j : Normalized weights.

[0210] 4. Decision Chain Weight Calibration and Two-Path Collaborative Modeling

[0211] 1. Logistics decision chain weight calculation

[0212] Weight iteration based on factor importance judgment matrix:

[0213]

[0214] M i,j : Judgment matrix elements (such as the importance ratio of "road condition factor" to "order factor"). δ i,j,k : The contribution of the kth factor to the path deviation (i, j). w k : The dynamic weight of the kth factor (such as road condition weight, timeliness weight). K: The total number of factors, n: The number of deviation samples.

[0215] 2. Reward Function for Dual-Path Collaborative Decision-Making

[0216] Integrating data-driven and knowledge-driven reward design:

[0217] R(s,a)=R data (s,a)+λ·R kg (s,a)

[0218] R data (s,a)=γ1·time_bonus+γ2·cost_saving

[0219]

[0220] R data : Data-driven rewards (timeliness, cost savings). kg : Knowledge-driven reward (score for compliance with business rules). λ, γ1, γ2: Weight coefficients. Rules: The set of business rules in the knowledge graph (e.g., cold chain transportation rules). Indicator function, 1 if the rule is satisfied, 0 otherwise.

[0221] Logistics two-way collaborative decision network, represented as a five-tuple:

[0222] Data-driven path module, which implements dynamic strategy optimization based on reinforcement learning. The knowledge-driven path module implements rule reasoning based on the knowledge graph. The dual-path fusion module is responsible for the dynamic adjustment of policy weights. The decision chain generation module integrates the dual outputs to form a decision chain. The system mapping module breaks down the decision chain into business strategies.

[0223] Data-driven path module

[0224] 1. Strategy Generation Model

[0225] The data-driven strategy is represented by a neural network trained by the PPO algorithm:

[0226]

[0227] Encoder d (s): State encoder, which maps the logistics state (such as vehicle location, order set) into a feature vector. d : Policy network weight matrix, by optimizing the objective function Iterative updates.

[0228] 2. Dynamic Optimization Goals

[0229] Optimization objective function of data-driven path (proximal strategy optimization):

[0230]

[0231] The advantage function Using the generalized advantage estimate (GAE):

[0232]

[0233] γ: discount factor, λ: GAE decay coefficient. V(s): state value function, estimated by the value network.

[0234] Knowledge-driven path module Mathematical modeling

[0235] 1. Knowledge Graph Reasoning Framework

[0236] The knowledge-driven strategy is based on the graph triple G = {(e i ,r ij ,e j )}generate:

[0237]

[0238] The matching function: Match(s,a,e s ,r,e o )=Sim(s,e s )·Rel(r)·Eff(a,e o )

[0239] Parameter Description:

[0240] Sim(s,e s ): state and entity es The semantic similarity of the relationship (such as the matching degree between the order type and the warehouse entity). Rel(r): The weight of the relationship r (such as the priority of the "must use" relationship is higher than the "optional" relationship). Eff(a,e o ): action a on entity e o The impact of efficiency (such as the coverage efficiency of distribution actions on distribution points).

[0241] 2. Rule-constrained decision correction

[0242] The hard constraints on decision-making in the knowledge-driven path are implemented through the penalty function:

[0243]

[0244] A set of hard rules (such as load limits and time windows). Action a is 1 if it violates rule r, otherwise it is 0. r : The penalty intensity of rule r (such as timeout penalty coefficient).

[0245] Dual-channel fusion module

[0246] 1. Dynamic Fusion Strategy

[0247] Dynamic weighted fusion function of the two-way strategy:

[0248]

[0249] The fusion coefficient λ(s) is dynamically adjusted by the complexity of the environment:

[0250]

[0251] Complexity(s): Environmental complexity feature vector (such as order fluctuation amplitude, road condition change rate). λ : fusion coefficient weight, σ: Sigmoid activation function.

[0252] 2. Reward Function Fusion

[0253] The joint reward function of two-way collaboration:

[0254]

[0255] Decision chain generation and system mapping

[0256] 1. Generation of dual-path thinking chains

[0257] The decision chain generation function transforms the two-way strategy into a temporal decision sequence:

[0258] Chain(s0)= <a0,a1,...,aT >,a t =π(a t |s t )

[0259] The state transition satisfies:

[0260] s t+1 =Transition(s t ,a t ,∈ t ),∈ t ~p(∈|s t ,a t )

[0261] ∈ t :Environmental random disturbances (such as sudden congestion), subject to the conditional probability distribution p(∈|s t ,a t ).

[0262] 2. Business system mapping

[0263] Mapping function from decision chain to business system:

[0264]

[0265] Map(·): Mapping function, which extracts corresponding strategies (such as order allocation and vehicle scheduling) according to the needs of the business module.

[0266] Mathematical closed loop of network operation

[0267] The iterative optimization closed loop of the two-way collaborative decision-making network can be expressed as:

[0268]

[0269] The update function:

[0270]

[0271] Parameters: k: number of iterations, {Chain}: set of historical decision chains. PPO_Update: updates the data-driven module based on the PPO algorithm. KG_Update: updates the knowledge graph based on decision chain feedback (e.g., adding new rules).

[0272] like Figure 2 As shown, the embodiment of the present application provides a device for constructing a dual-path collaborative decision-making network for multi-agent collaborative path optimization, which includes:

[0273] A first program unit is configured to acquire real-time business data and historical business data from multiple sources and heterogeneous structures and construct a standardized data set;

[0274] a second program unit configured to convert the standardized data set into a dynamic environment input sequence;

[0275] A third program unit is configured to enable at least two domain agents to perform full-process business logic reasoning based on the dynamic environment input sequence to generate a decision chain structure covering the entire life cycle of the business;

[0276] a fourth program unit configured to implement a full-process path planning simulation based on the decision chain structure to generate an initial path plan including multi-resource coordinated scheduling;

[0277] A fifth program unit is configured to enable the environment-aware intelligent agent to execute the initial delivery path plan in an actual logistics delivery scenario, and load real-time environmental change information to determine the weight of the logistics decision chain;

[0278] The fifth program unit is configured to be used by the logistics strategy intelligent agent to optimize the initial distribution path plan based on the logistics decision chain weight to obtain a logistics business knowledge graph to build a logistics two-way collaborative decision network.

[0279] Figure 3 This is a schematic diagram of the structure of an electronic device according to this embodiment; the electronic device may include:

[0280] One or more processors 301;

[0281] The computer readable medium 302 may be configured to store one or more programs,

[0282] When the one or more programs are executed by the one or more processors, the one or more processors implement the method described in the above embodiments.

[0283] Figure 4 The hardware structure of the electronic device of this embodiment is as follows; Figure 4 As shown, the hardware structure of the electronic device may include: a processor 401, a communication interface 402, a computer-readable medium 403 and a communication bus 404;

[0284] The processor 401, the communication interface 402, and the computer-readable medium 403 communicate with each other via a communication bus 404;

[0285] Optionally, the communication interface 402 may be an interface of a communication module, such as an interface of a GSM module;

[0286] The processor 401 may be specifically configured to implement the method provided in the above embodiment.

[0287] The processor 401 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, or a discrete hardware component. It can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor.

[0288] The computer-readable medium 403 may be, but is not limited to, a random access memory (RAM), a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), etc.

[0289] above Figure 2-Figure 4 For an exemplary explanation of the scheme, please refer to the above Figure 1 .

[0290] The above description is merely a preferred embodiment of the present application and an illustration of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to the technical solutions formed by the specific combination of the above-mentioned technical features, but also encompasses other technical solutions formed by any combination of the above-mentioned technical features or their equivalents without departing from the above-mentioned inventive concept. For example, a technical solution formed by replacing the above-mentioned features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A method for constructing a dual-path collaborative decision-making network for multi-agent collaborative path optimization, characterized in that: include: Acquire multi-source heterogeneous real-time business data and historical business data to build standardized data sets; converting the standardized data set into a dynamic environment input sequence; At least two domain agents perform full-process business logic reasoning based on the dynamic environment input sequence to generate a decision chain structure covering the entire life cycle of the business; Implementing full-process path planning simulation based on the decision chain structure to generate an initial path plan including multi-resource collaborative scheduling; The environment-aware agent executes the initial delivery path plan in the actual logistics delivery scenario, loading real-time environmental change information to determine the weight of the logistics decision chain; The logistics strategy agent optimizes the initial distribution path plan based on the logistics decision chain weight to obtain the logistics business knowledge graph to build a logistics two-way collaborative decision network.

2. The method according to claim 1, characterized in that The acquisition of multi-source heterogeneous real-time business data and historical business data to build a standardized data set includes: Based on real-time logistics road condition data, delivery order information data, delivery vehicle status data, and historical delivery business data, a standardized data set for logistics business is constructed.

3. The method according to claim 2, characterized in that The standardized data set for logistics business is constructed based on real-time logistics road condition data, delivery order information data, delivery vehicle status data, and historical delivery business data, including: Perform spatiotemporal alignment on real-time logistics traffic data, delivery order information data, delivery vehicle status data, and historical delivery business data to unify them into the same spatiotemporal coordinate system and obtain a spatiotemporal aligned original logistics data stream. Perform wavelet transform and multi-resolution transform on the original data stream to decompose the original data stream into different frequency levels to remove the noise and obtain the effective logistics data stream; Based on the effective logistics data flow, a standardized data set for logistics business is constructed.

4. The method according to claim 3, characterized in that The construction of a standardized logistics business data set based on the effective logistics data flow includes: Perform multi-dimensional feature extraction on effective logistics data streams and perform dynamic local mapping and vector space alignment to obtain a set of semantic feature vectors; Based on the constructed spatiotemporal index and semantic index, the spatiotemporal semantic tensor and logistics knowledge graph are injected into the semantic feature vector set to generate a standardized data set for logistics business.

5. The method according to claim 1, wherein The at least two domain agents perform full-process business logic reasoning based on the dynamic environment input sequence to generate a decision chain structure covering the entire life cycle of the business, including: the intelligent warehouse distribution agent, the distribution vehicle scheduling agent, and the route planning agent perform logical reasoning on the entire life cycle of the logistics order based on the dynamic environment input sequence to generate a decision chain structure covering the entire process of warehousing and outbound delivery, in-transit transportation, and terminal delivery.

6. The method according to claim 5, characterized in that Also includes: Based on the coding unit, perform the following steps: The spatial information encoding in the dynamic environment input sequence is converted into a hierarchical encoding, and the hierarchical encoding is spatially associated and temporally characterized to obtain a spatiotemporal hierarchical semantic feature tensor; The spatiotemporal hierarchical semantic feature tensor is fused with the logistics domain knowledge graph to obtain an enhanced dynamic environment input sequence.

7. The method according to claim 6, characterized in that The intelligent warehouse distribution agent, delivery vehicle scheduling agent, and route planning agent perform logical reasoning on the entire logistics order lifecycle based on dynamic environment input sequences to generate a decision chain structure covering the entire process of warehousing and outbound delivery, in-transit transportation, and terminal delivery, including: Based on the intelligent warehouse partitioning agent, the order information and warehouse information in the enhanced dynamic environment input sequence are converted into graph structure data to obtain the warehouse partitioning decision logic sub-chain; Based on the delivery vehicle scheduling agent, a mathematical model with time window constraints and vehicle load balancing constraints is established based on the vehicle status, order timeliness requirements, and real-time road condition data in the enhanced dynamic environment input sequence. This model is solved using proximal policy optimization to obtain the vehicle scheduling decision logic subchain. Based on the route planning agent, a spatiotemporal graph is constructed based on the road network, delivery point locations, and real-time traffic flow information in the enhanced dynamic environment input sequence. The graph attention mechanism is used to determine the association weights between different road sections and delivery points to generate a route planning decision logic sub-chain. Based on the warehouse division decision logic sub-chain, vehicle scheduling decision logic sub-chain, and route planning decision logic sub-chain, a decision chain structure covering the entire process of warehousing and delivery, in-transit transportation, and terminal distribution is generated.

8. The method according to claim 1, characterized in that The full-process path planning simulation is implemented based on the decision chain structure to generate an initial path plan including multi-resource collaborative scheduling, including: based on the decision chain structure covering the entire process of warehousing and outbound delivery, a full-process path planning simulation is implemented from pickup and loading, trunk line transportation to terminal delivery, and an initial distribution path plan including multi-vehicle collaborative distribution and order consolidated transportation is preliminarily determined.

9. The method according to claim 1, characterized in that The environment-aware agent executes the initial delivery path plan in the actual logistics delivery scenario, loading real-time environmental change information to determine the weight of the logistics decision chain, including: Extracting features from real-time environmental data to obtain environmental feature data; The initial delivery path plan is put into actual logistics distribution scenarios for execution and the interference factors of environmental characteristic data on the initial delivery path plan are dynamically analyzed; Calculate path execution deviation data based on the interference factor; Based on the constructed factor importance judgment matrix and the path execution deviation data, the logistics decision chain weight is calculated.

Images