Tire changer sales strategy self-adaptive generation system based on reinforcement learning

Abstract

The invention relates to the technical field of tire changer sales, in particular to a tire changer sales strategy self-adaptive generation system based on reinforcement learning, which comprises a supply and demand feature sensing module, a conversion fluctuation recognition module, a control module, a decision-making module, a decision-making module, a decision-making module, a decision-making module and a decision-making module, the state operation association module analyzes multi-dimensional attributes to establish mapping, the interaction process arrangement module arranges operations according to a state sequence, and the interaction behavior self-adaptive adjustment module positions deviation according to feedback and selects replacement operations for rearrangement to obtain a self-adaptive interaction strategy generation result. According to the method, the market state is enabled to have a structural property through multi-dimensional disassembly and transaction consistency description in the sales process, abnormal limits are converted to positions, business constraints are associated to enable positioning to have directivity, stable mapping is established between the process state and operation, and an interactive predictable basis is given; and the operation sequence is rearranged according to feedback, so that the sales behavior keeps coherent adjustment capability, and the supply and demand identification definition, the interaction stability and the strategy adaptation performance are enhanced.

Description

Technical Field

[0001] This invention relates to the field of tire changer sales technology, and in particular to an adaptive generation system for tire changer sales strategies based on reinforcement learning. Background Technology

[0002] The tire changer sales technology field encompasses aspects such as mechanical equipment sales management and market strategy optimization. The core of this field lies in improving sales efficiency and market adaptability through in-depth analysis and modeling of market demand, customer behavior, sales channels, and strategies for specific equipment products. As a type of automotive repair equipment, the tire changer sales technology field covers customer demand identification, sales process decision support, supply and demand matching mechanisms, dynamic price adjustments, and sales strategy feedback optimization. The overall technology system primarily relies on historical sales data, market feedback information, equipment performance parameters, and sales behavior paths, combined with mathematical models and algorithmic reasoning tools, to construct, evaluate, and adjust sales strategies, thereby achieving management and support for the entire sales process.

[0003] Among them, the tire-changing machine sales strategy adaptive generation system based on reinforcement learning refers to a technical solution that uses reinforcement learning methods to automatically generate and iteratively optimize sales strategies by constructing an interaction mechanism between the sales environment and the decision agent. The technical aspects covered include sales scenario modeling, sales behavior representation, state space construction, sales action decision-making, and reward and punishment mechanism definition. Specifically, by setting sales state index parameters, a strategy gradient algorithm is used to continuously adjust the sales strategy selection path in discrete time steps, and the strategy is corrected according to the set reward function in each round of interaction, thereby forming a sales strategy output oriented towards target revenue. Based on the reinforcement learning framework, historical sales data is used as training samples, and the strategy is evaluated and updated through a simulation environment to achieve adaptive generation of sales strategies in different sales scenarios.

[0004] Existing sales strategy generation methods focus on overall strategy path deduction, and the expression of sales status is mainly based on unified abstract indicators. Regional differences and local supply and demand changes are difficult to fully reflect. In the process of multi-stage interaction, conversion fluctuations are mostly attributed to result deviations rather than specific operational nodes. When the pace of customer feedback or business conditions change, the strategy response lacks process-level positioning support, and the adjustment results tend to remain at the macro level, leading to the accumulation of local process imbalances and affecting continuous sales performance and strategy consistency. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and propose an adaptive generation system for tire changing machine sales strategies based on reinforcement learning.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: an adaptive generation system for tire changing machine sales strategies based on reinforcement learning, the system comprising: The supply and demand characteristic perception module acquires business node information in sales areas, divides sales areas and assigns regional node numbers, collects transaction data, generates trend vectors, analyzes trend consistency, distinguishes supply and demand performance types, and generates supply and demand data characteristic area results. The conversion fluctuation identification module obtains the conversion efficiency of each regional node's interactive operation based on the regional node number assigned with supply and demand feature labels in the supply and demand data feature region results, extracts response delay and protocol changes, identifies the position of conversion efficiency deviation, outputs the operation position number, and generates the conversion fluctuation position result. The status operation association module extracts the customer attributes, latency, protocol changes, feedback signals, and delivery cycles associated with the corresponding operation based on the operation position number in the conversion fluctuation position result, analyzes the relationship with the interactive operation, establishes a corresponding description of the process status and interactive operation, and generates the status operation mapping result. The interactive process orchestration module obtains the actual sequence of process states during process execution based on the corresponding descriptions of process states and interactive operations in the state operation mapping results, selects and arranges interactive operation records in sequence to form a continuous progression description, and generates an interactive operation sequence result.

[0007] As a further aspect of the present invention, the supply and demand data feature area results include the level of regional transaction volume change, the type of regional price change direction, the division of regional supply and demand tension, and the characteristics of regional transaction fluctuation amplitude. The conversion fluctuation position results specifically include the conversion anomaly occurrence sequence number, the conversion efficiency decline segment, and the conversion unstable operation point. The state operation mapping results include the process state type set, the state corresponding interactive operation category, and the state matching condition elements. The interactive operation sequence results specifically include the interactive operation arrangement order, the operation connection structure form, and the description of the process advancement rhythm.

[0008] As a further embodiment of the present invention, the supply and demand feature perception module includes a regional node division submodule, a trend vector construction submodule, and a supply and demand type labeling submodule; The regional node division submodule obtains the business execution node information corresponding to each sales region during the tire changer sales process, divides the sales regions according to the business coverage and sales execution boundaries, sets regional node numbers for each sales region, collects the transaction quantity and transaction price change direction information in the region corresponding to the regional node number, and generates a node number mapping list. The trend vector construction submodule calls the transaction volume and transaction price change direction information collected in the node number mapping list, arranges the transaction volume of different time periods under the same regional node number in order, sorts the arrangement order of the transaction price change direction information, forms continuous records respectively, generates transaction volume change vector and transaction price change vector in time order, and generates a trend vector sorting list after summarizing and sorting. The supply and demand type labeling submodule identifies the trend of changes in transaction volume and price over time based on the two types of vector order information corresponding to the node numbers of each region in the trend vector sorting list. It then labels the region node numbers based on the trend performance and matches the labeled node numbers with the corresponding label information to generate supply and demand data feature region results.

[0009] As a further embodiment of the present invention, the conversion fluctuation identification module includes a node label extraction submodule, an interaction information collection submodule, and an offset position identification submodule; The node tag extraction submodule obtains all region node numbers that have been assigned supply and demand feature tags in the supply and demand data feature region results, filters out node numbers with clear tag identifiers from the region node number list, performs field matching between the region node numbers with tag identifiers and the interaction operation record numbers corresponding to the business execution stage, organizes the corresponding combination content between region node numbers and interaction operation record numbers, and generates a node interaction record index table. The interaction information collection submodule calls the interaction operation record number in the node interaction record index table, extracts the conversion efficiency value, contact response delay information and settlement agreement change information corresponding to the current record from each interaction operation record, organizes and collects the conversion efficiency value, contact response delay information and settlement agreement change information in the order of interaction operation record number, and establishes information grouping according to regional node number to generate node interaction feature set. The offset position identification submodule determines whether the conversion efficiency change direction between consecutive records is consistent based on the order of the conversion efficiency values ​​corresponding to each group of interactive operation records in the node interaction feature set. Then, it filters out the numbers of interactive operation records with inconsistent change directions. At the same time, it performs joint filtering processing on the filtered interactive operation record numbers with contact response delay values ​​and settlement agreement change values ​​to filter out the position numbers with conversion efficiency offset in the interaction sequence and generate conversion fluctuation position results.

[0010] As a further aspect of the present invention, the state operation association module includes an operation record extraction submodule, an attribute feature collection submodule, and a state type identification submodule; The operation record extraction submodule obtains the operation position number listed in the conversion fluctuation position result, retrieves the corresponding interaction operation record number from each operation position number, matches the operation record number with the interaction execution record in the business data source, and archives the successfully matched interaction execution records to generate an operation record number set. The attribute feature collection submodule calls each interaction execution record in the operation record number set, extracts the customer attribute feature value, contact response delay value, settlement agreement change status, quotation feedback signal category and delivery cycle time period corresponding to the interaction operation record, and completes the structured organization of the five extracted information with the operation record number as the index, and classifies and combines them according to the field to generate interaction attribute feature groups. The status type identification submodule determines the distribution and combination differences of attribute combinations in the interactive operation sequence based on five items: customer attribute feature value, contact response delay value, settlement agreement change status, quotation feedback signal category, and delivery cycle time period corresponding to each operation record number in the interactive attribute feature group. It then groups operation record numbers with similar attribute combinations into the same process status, establishes a matching structure between interactive operations and process statuses in the order of process statuses, and generates status operation mapping results.

[0011] As a further embodiment of the present invention, the interactive process orchestration module includes a state sequence acquisition submodule, an operation record selection submodule, and a sequence arrangement generation submodule; The state sequence acquisition submodule acquires the process state and interactive operation corresponding description in the state operation mapping result, collects the actual occurrence order information of each process state during process execution, organizes and archives the process state occurrence order content in chronological order, records the process state number and occurrence order index relationship, and generates a process state sequence index table. The operation record selection submodule calls the process status number of each process status in the process status sequence index table, filters the interactive operation record content associated with the process status number according to the interactive operation record number corresponding to the process status in the status operation mapping result, and collects the filtered interactive operation records by number according to the process status sequence index to generate an interactive record index list. The sequence arrangement generation submodule determines whether there are any missing positions in the index order of the interactive operation records according to the numbering order of the interactive operation records in the interactive record index list. It then arranges the consecutively numbered interactive operation records in sequence to form an interactive operation progression order description and generates an interactive operation sequence result.

[0012] As a further aspect of the present invention, the system further includes: The adaptive adjustment module for interactive behavior, based on the order of interactive operations in the result of the interactive operation sequence, obtains the conversion feedback data corresponding to the order of interactive operations during the actual execution process, locates the deviation position of the interactive operation, selects alternative interactive operation records, rearranges the order of interactive operations, and generates an adaptive interactive strategy generation result. The adaptive interaction strategy generates results including alternative interaction operation options, operation order adjustment schemes, and interaction strategy configuration parameters.

[0013] As a further embodiment of the present invention, the interactive behavior adaptive adjustment module includes a feedback data extraction submodule, an offset position positioning submodule, and a sequence rearrangement generation submodule; The feedback data extraction submodule obtains the interaction operation progress order in the interaction operation sequence result, uses each interaction operation progress number as an index, collects the corresponding conversion feedback data content in the actual execution process, and archives the conversion feedback data with the interaction operation progress number to generate an operation feedback matching list. The offset position positioning submodule determines whether there are position numbers in the advancement sequence that are inconsistent with the feedback values ​​and state descriptions based on the conversion feedback data value content corresponding to each interactive operation advancement number in the operation feedback matching list and the interactive operation description corresponding to each process state in the state operation mapping result. It then performs filtering and positioning on the position numbers to generate a set of advancement offset position numbers. The sequence rearrangement generation submodule calls each number in the set of advancement offset position numbers, obtains the interaction operation record corresponding to the number, filters the content of the associated alternative interaction operation record, replaces the abnormal position number in the original interaction operation advancement sequence with the corresponding alternative interaction operation record number, and then recombines the operation sequence to generate the adaptive interaction strategy generation result.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, the sales process is broken down into multi-dimensional expressions at the regional and operational levels. Regional supply and demand differences are characterized by the consistency of transaction changes, making the market state more structured. Conversion anomalies are limited to specific interaction sequence positions and correspond to customer conditions and business constraints, making problem identification more targeted. A stable correspondence is formed between process status and operation selection, making interaction arrangements predictable. The operation sequence is continuously rearranged based on actual feedback, enabling sales behavior to maintain consistent adjustment capabilities in different market environments, thereby enhancing the clarity of supply and demand identification, interaction stability, and strategy adaptability. Attached Figure Description

[0015] Figure 1 This is a system flowchart of the present invention; Figure 2This is a flowchart illustrating the acquisition process of the supply and demand feature sensing module of the present invention. Figure 3 This is a flowchart illustrating the acquisition process of the fluctuation recognition module of the present invention. Figure 4 This is a flowchart illustrating the acquisition process of the state operation association module of the present invention. Figure 5 This is a flowchart illustrating the acquisition process of the interactive flow orchestration module of the present invention. Figure 6 This is a flowchart of the interactive behavior adaptive adjustment module of the present invention. Detailed Implementation

[0016] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0017] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0018] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0019] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0020] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0021] Please see Figure 1 This invention provides a technical solution: an adaptive generation system for tire changing machine sales strategies based on reinforcement learning, the system comprising: The supply and demand characteristic perception module acquires business execution node information corresponding to each sales region during the tire changer sales process. It divides the sales regions according to the business coverage and sales execution boundaries, and assigns a regional node number to each sales region. It collects information on the transaction quantity and transaction price change direction within the region corresponding to the regional node number, forming a transaction volume trend vector and a transaction value trend vector. It analyzes and processes the consistency of the changes of the two types of trend vectors in the continuous collection sequence, distinguishes and describes the supply and demand performance type of the regional nodes, and generates supply and demand data characteristic region results. The conversion fluctuation identification module obtains the conversion efficiency data of each interactive operation within the region corresponding to the region node number based on the regional node number assigned to the supply and demand feature label in the supply and demand data feature region result. It extracts the contact response delay information and settlement agreement changes corresponding to the interactive operation, analyzes the performance of the conversion efficiency data in the process of interactive operation sequence, identifies the operation position where the conversion efficiency performance deviates in the interactive operation sequence, outputs the corresponding operation position number, and generates conversion fluctuation position result. The status operation association module retrieves the interaction operation records associated with the corresponding operation position based on the operation position number in the conversion fluctuation position result. It extracts the customer attribute characteristics, contact response delay, settlement agreement changes, quotation feedback signals and delivery cycle information corresponding to the interaction operation, analyzes the correspondence between the interaction operation and the customer attribute characteristics, contact response delay, settlement agreement changes, quotation feedback signals and delivery cycle information under different operation positions, distinguishes the interaction operation types used in different states during the process, forms a corresponding description between process states and interaction operations, and generates status operation mapping results. The interactive process orchestration module obtains the actual occurrence order of each process state during process execution based on the corresponding description of the process state and interactive operation in the state operation mapping result. Using the occurrence order of the process state as the index order, it sequentially selects the interactive operation records associated with each process state in the corresponding description, arranges the selected interactive operation records continuously according to the index order, forms a continuous interactive operation progression order description, and generates the interactive operation sequence result. The adaptive adjustment module for interactive behavior obtains the conversion feedback data corresponding to the actual execution sequence of interactive operations based on the order of interactive operations in the result of the interactive operation sequence. It compares and analyzes the correspondence between the process state and the interactive operation in the conversion feedback data and the state operation mapping result to locate the operation position in the order of interactive operations that has deviated. It selects the alternative interactive operation record associated with the operation position, rearranges the order of interactive operations, and obtains the adaptive interaction strategy generation result.

[0022] The regional results of supply and demand data characteristics include the level of regional transaction volume changes, the type of regional price change direction, the division of regional supply and demand tension, and the characteristics of regional transaction fluctuation amplitude. The results of conversion fluctuation location specifically include the sequence number of conversion anomalies, the segment of conversion efficiency decline, and the point of conversion instability. The results of state operation mapping include the set of process state types, the category of interactive operation corresponding to the state, and the elements of state matching conditions. The results of interactive operation sequence specifically include the order of interactive operations, the form of operation connection structure, and the description of process progress rhythm. The results of adaptive interaction strategy generation include alternative interactive operation options, operation order adjustment schemes, and interaction strategy configuration parameters.

[0023] Please see Figure 2 The supply and demand feature perception module includes a regional node division submodule, a trend vector construction submodule, and a supply and demand type labeling submodule. The regional node division submodule obtains the business execution node information corresponding to each sales region during the tire changer sales process, divides the sales regions according to the business coverage and sales execution boundaries, sets regional node numbers for each sales region, collects the transaction quantity and transaction price change direction information in the region corresponding to the regional node number, and generates a node number mapping list. To obtain business execution node information corresponding to each sales region during the tire changing machine sales process, in the initialization phase of constructing the state space of the reinforcement learning environment, the business execution nodes are treated as discrete anchor points in the environment. By calling the logistics and network database in the enterprise ERP system, the geographical coordinate parameters of each anchor point are extracted. As a positioning benchmark, a coverage threshold is set based on the existing maximum logistics delivery distance or after-sales service response radius of a single tire changer. For example, setting The distance is 150 kilometers. Using the spatial query function of a GIS geographic information system, the coordinates of all potential customers within the sales area to be divided are traversed. Calculate the Euclidean distance between the customer's location and each business execution node. If the calculated result Less than the coverage threshold Then, the customer's location is assigned to the coverage set of the corresponding business execution node. For customer locations located in the overlapping coverage area of ​​multiple nodes, the principle of minimizing distance is applied. After spatial clustering, each sales region is assigned a unique regional node number according to a sequence of natural numbers starting from 1, and then assigned to a unique business execution node. For example, generating the numbers "Area_001" and "Area_002" for each For the corresponding regional set, an SQL data acquisition channel is established to monitor and record the transaction quantity of each tire changer order within the region in real time from the order management system. and transaction price At the same time, retrieve the price of the previous transaction record. Perform interpolation ; like Then the price change direction information is recorded as +1, if Then it is recorded as -1, if The record is 0, which quantifies the price fluctuation state. The regional node number is used as the primary key, and its geographical range metadata and the collected initial transaction volume and price direction data are associated to build a hash mapping table and generate a node number mapping list.

[0024] The trend vector construction submodule calls the collected transaction volume and transaction price change direction information in the node number mapping list, arranges the transaction volume of different time periods under the same regional node number in order, sorts the arrangement order of the transaction price change direction information, forms continuous records, generates transaction volume change vector and transaction price change vector in time order, and generates a trend vector sorting list after summarizing and sorting. The system retrieves the collected transaction quantity and price change direction information from the node number mapping list. During the reinforcement learning state observation sequence construction step, this information is used for each independent region node number in the list. Historical transaction log data is extracted using a database cursor, with a set time step. For discrete sampling units, for example, taking 5 transactions as a sampling window, the number of transactions under the same node number in the same area is stacked according to the order of the timestamps of the transactions; Construct a sequence array of transaction quantities ,in Representing the The cumulative number of transactions within the sampling window, for example, the transaction volume of "Area_001" in three consecutive time periods being 5, 8, and 12 units respectively, then the corresponding... The segment is [5,8,12]. The corresponding transaction price change direction information is extracted synchronously and arranged according to the same timestamp index order; Construct a price trend sequence array ,in For example, if the prices in all three of the above time periods show an upward trend, then... The fragment is [+1,+1,+1]. These two sequence arrays are merged and encapsulated into a feature tensor that can represent the dynamic changes in the market environment of this region, and used as a reinforcement learning agent. Based on the state input at each time step, the feature tensors of all region nodes are formatted and validated. Cold start region data with a sequence length of less than 3 time steps are removed. The remaining valid feature tensors are indexed and sorted according to the ASCII code order of the region node numbers to form a structured list containing multiple sets of time series data. This completes the preprocessing of environmental state features. After summarizing and organizing, a trend vector sorted list is generated.

[0025] The supply and demand type labeling submodule identifies the trend of changes in transaction volume and price over time based on the two types of vector order information corresponding to the node numbers of each region in the trend vector sorting list. It then labels the region node numbers based on the trend performance and matches the labeled node numbers with the label information to generate supply and demand data feature region results. Based on the two types of vector order information corresponding to the node numbers of each region in the trend vector sorting list, the transaction quantity sequence array in the list is used in the state classification and reward function auxiliary calculation stage of the reinforcement learning environment. and price trend sequence array Slope features were extracted and fitted using the least squares method. The data points in the formula are as follows: Calculate the slope of the trend line of change in trading volume. Set a baseline threshold for the trend of quantity changes. This threshold is set to 0.5 times the historical average trading volume volatility variance. For example, if the historical variance is 20, then... If the calculated result If it is determined to be in a state of "surge in demand", then If the condition is determined to be "demand contraction", it is considered "demand stability" otherwise. Simultaneously, calculations are performed. The average value of the elements Set a benchmark threshold for price trends. ,like If it is judged as "price increase", then If the condition is determined as "price decline," it is considered "price fluctuation." The quantity trend and price trend are combined for logical judgment. If the result is "demand surge" and "price rise," the node number of that region is assigned the feature label "supply shortage." If the result is "demand contraction" and "price decline," it is assigned the feature label "supply surplus." Other combinations are assigned specific intermediate state labels such as "quantity-price divergence" or "market equilibrium." The generated labels are used as the environmental state identifier (State ID) of that region and updated to the region attribute table. This allows the reinforcement learning agent to identify the current environmental pattern and determine subsequent interaction strategies, generating the supply and demand data feature region results.

[0026] Please see Figure 3 The transformation fluctuation recognition module includes a node label extraction submodule, an interaction information collection submodule, and an offset position recognition submodule; The node tag extraction submodule retrieves all region node numbers that have been assigned supply and demand feature tags from the supply and demand data feature region results. It then filters out node numbers with clear tag identifiers from the region node number list, performs field matching between the region node numbers with tag identifiers and the interaction operation record numbers corresponding to the business execution stage, organizes the corresponding combination content between region node numbers and interaction operation record numbers, and generates a node interaction record index table. Retrieve all region node numbers with assigned supply and demand feature labels from the supply and demand data feature region results, iterate through the label fields in the supply and demand data feature region results, and execute programming logic judgment operations. The system identifies all node numbers with non-empty labels, removes node numbers with a status of "pending" or insufficient data that prevents classification, and retains only valid node numbers with clear market characteristics (such as "supply shortage" or "supply surplus"). For example, retain the record with the ID "Area_001" and the tag "supply shortage", and construct a list of valid nodes. ,against For each node number in the database, a multi-table join query (JOIN) request is initiated to the business database. Using the node number as a foreign key index, the query retrieves all sales order logs in the "in progress" or "completed" state within that region, and extracts the unique serial number generated for each interaction action (such as quoting, negotiating, or signing) in these logs. For example, extract the serial numbers of 100 associated interaction records from “Area_001” to “Act_100” from “Area_001”, and then extract the serial numbers of those records. The data is stored in a chain according to the chronological order of the business transactions, forming a key-value pair structure. Perform data integrity checks on key-value pairs and remove corresponding entries. Empty nodes with a list length of 0 will have their cleaned and matched key-value pairs written into a structured index database, providing a basic index path for the state-action trajectory replay of reinforcement learning agents and generating a node interaction record index table.

[0027] The interaction information collection submodule calls the interaction operation record number in the node interaction record index table, extracts the conversion efficiency value, contact response delay information and settlement agreement change information corresponding to the current record from each interaction operation record, organizes and collects the conversion efficiency value, contact response delay information and settlement agreement change information in the order of interaction operation record number, and establishes information grouping by regional node number to generate node interaction feature set. The system retrieves the interaction record number from the node interaction record index table and iterates through each interaction record number in the index table. Access the underlying data logs of the CRM system, read the conversion funnel data recorded at the time the operation occurred, and calculate the retention rate of the current step relative to the previous step as a conversion efficiency value. (Dimensionless value, range 0-1), for example, if a quotation is sent to 100 potential clients, and only 20 clients proceed to the subsequent contract review stage, then the value of this operation is... The difference between the server timestamps from when the request was initiated to when the client first responded is synchronously calculated as contact response latency information. The unit is accurate to the minute. For example, if the quotation was issued at 10:00 and the customer responded at 14:30, then... Within minutes, extract the contract draft version number and standard template version number associated with the operation, and use the text difference comparison algorithm (Diff Algorithm) to compare the number of clause differences and the modification range of key parameters (such as price and payment cycle) between the two. The degree of variation is quantified into discrete level values. Where 0 represents no changes to the standard protocol, 1 represents only modifications to non-core terms, and 2 represents changes involving core pricing or payment terms, such as changes to payment terms in a particular interaction. The acquired three-dimensional feature data Encapsulated as a feature vector and according to the corresponding Perform grouping and aggregation to construct a structure like The feature sequence set is directly mapped to the state transition trajectory data in the reinforcement learning environment, generating a node interaction feature set.

[0028] The offset position identification submodule determines whether the conversion efficiency change direction between consecutive records is consistent based on the order of the conversion efficiency values ​​corresponding to each group of interaction operation records in the node interaction feature set. Then, it filters out the numbers of the interaction operation records with inconsistent change directions. At the same time, it performs joint filtering processing on the filtered interaction operation record numbers with contact response delay values ​​and settlement agreement change values ​​to filter out the position numbers with conversion efficiency offset in the interaction sequence and generate conversion fluctuation position results. Based on the order of conversion efficiency values ​​corresponding to each group of interaction operation records in the node interaction feature set, during the reward feedback calculation and anomaly detection stages of reinforcement learning, the conversion efficiency value sequence within each group is... Perform first-order difference operations to calculate the gradient of conversion rate changes between adjacent operations. Set a tolerance threshold for conversion rate fluctuations. (Normal fluctuations within %), traverse the difference sequence, if detected If the conversion rate shows a significant negative deviation (a sharp drop), then position i is initially marked as a potential anomaly. For position i, the corresponding contact response delay value is then retrieved. Changes in settlement agreement To ensure consistency of physical dimensions, the time delay is first normalized during joint weighted evaluation. ( (This refers to the region's historical maximum time delay constant, such as 1440 minutes). Normalize the changes to the agreement. 2. Set latency penalty weights =0.4 and protocol change penalty weight =0.6; Calculate the dimensionless comprehensive drag coefficient = + If the calculated result Greater than the preset resistance threshold If the decrease in conversion rate at position i is determined to be a substantial deviation caused by high operational resistance (such as slow response or excessive protocol change), then position i and its associated operations are designated as such. It was identified as a negative reward trigger point in reinforcement learning, and a transformation fluctuation position result was generated.

[0029] Please see Figure 4 The status operation association module includes an operation record extraction submodule, an attribute feature collection submodule, and a status type identification submodule. The operation record extraction submodule obtains the operation position number listed in the conversion fluctuation position result, retrieves the corresponding interaction operation record number from each operation position number, matches the operation record number with the interaction execution record in the business data source, and archives the successfully matched interaction execution records to generate an operation record number set. Retrieve the operation location numbers listed in the conversion fluctuation location results, and identify these location numbers that indicate abnormal fluctuations. The preprocessing interface for the Experience Replay Buffer, imported into the reinforcement learning agent, is numbered for each position. Using the database primary key ID in the original business logs, a precise retrieval request is initiated to the central business database to locate and extract the complete interaction execution record data packet associated with that location. The extracted data packet contains a unique identifier for that interaction. The operation type code (e.g., Code 101 for telephone follow-up, Code 202 for on-site demonstration), as well as the specific timestamp and executor ID of the operation execution, are used. For example, for location number "Pos_42", a record with ID "Int_9981" is extracted, showing the operation as "send revised contract". The extracted unique identifier is used... As a verification key, it is compared with the original logs backed up in the business data source for field consistency. It verifies whether the operation status field in the record is "closed" or "effective," removes invalid records with a status of "draft" or "undo," instantiates all interaction execution record objects that pass the consistency check, and then... The numerical values ​​are rearranged into an ordered list of records, which directly corresponds to the state-action pair sample set in reinforcement learning, generating an operation record number set.

[0030] The attribute feature collection submodule calls each interaction execution record in the operation record number set, extracts the customer attribute feature value, contact response delay value, settlement agreement change status, quotation feedback signal category and delivery cycle time period corresponding to the interaction operation record, and completes the structured organization of the five extracted information with the operation record number as the index, and classifies and combines them by field to generate interaction attribute feature groups. For each interaction execution record in the call operation record number set, during the feature engineering phase of constructing the reinforcement learning multidimensional state space, the JSON or XML format metadata in the record is parsed to extract customer attribute feature values. Specifically, this includes customer size level (1-micro, 2-medium, 3-large) and historical credit score (0-100 points). For example, extracting a customer as "Level 2, Credit 85" and extracting the contact response delay value. That is, the time difference (in minutes) between sending the previous message and receiving the response to this record, to extract the change status of the settlement agreement. (0 - None, 1 - Yes), Extract quotation feedback signal category By calling the NLP (Natural Language Processing) sentiment analysis interface, the feedback text is mapped to discrete categories (1-positive / acceptable, 0-neutral / neutral, -1-negative / rejected). For example, a customer's reply "the price is too high" is mapped to -1. The delivery cycle time period is then extracted. This refers to the promised number of days from shipment to delivery, such as "7 days." These five characteristic data points should be included. Numerical processing is performed according to a unified normalization standard, for example, customer size levels are normalized to... The latency value is scaled by dividing it by the maximum baseline value of 1440, and the five processed components are combined into a feature vector. and numbered by operation record The vector is stored in the Feature Store as a hash key, and used as the state input dimension of the Q-Table in subsequent reinforcement learning algorithms or as the input layer data of the deep neural network to generate interactive attribute feature groups.

[0031] The status type identification submodule determines the distribution and combination differences of attribute combinations in the interactive operation sequence based on five items: customer attribute feature value, contact response delay value, settlement agreement change status, quotation feedback signal category and delivery cycle time period corresponding to each operation record number in the interactive attribute feature group. It also groups operation record numbers with similar attribute combination content into the same process status, establishes a matching structure between interactive operations and process status in the order of process status, and generates status operation mapping results. Based on the five items corresponding to the operation record numbers in the interaction attribute feature group, the K-Means clustering algorithm is used to process the feature vectors during the state space discretization and clustering stage of reinforcement learning. Perform unsupervised classification and set the number of clusters. The estimated number of process states (e.g.) (corresponding to initial contact, needs confirmation, solution negotiation, business contract signing, and after-sales delivery, respectively), calculate the feature vector of each record and the cluster centers. Euclidean distance The records are assigned to the state category represented by the cluster center with the smallest distance. In the process, for example, if a record's feature vector is closest to the "solution negotiation" state center, then that record is marked as being in the "solution negotiation" state. This applies to each category of states after classification. Statistically analyze the top 3 most frequently occurring interaction types in this state. Construct a state-action probability distribution table. For example, in the "Solution Negotiation" state, the "Send Revised Quote" operation accounts for 60%, and the "Initiate Video Conference" operation accounts for 30%. Establish a mapping relationship M: →{opi}, this mapping relationship constitutes the initial policy network skeleton of the reinforcement learning agent, which clarifies what interaction operation (i.e., action) should be taken under what combination of customer attributes and feedback signals (i.e., state), and generates the state-action mapping result.

[0032] Please see Figure 5 The interaction flow orchestration module includes a state sequence acquisition submodule, an operation record selection submodule, and a sequence arrangement generation submodule; The state sequence acquisition submodule acquires the corresponding descriptions of process states and interactive operations in the state operation mapping results, collects the actual occurrence order information of each process state during process execution, organizes and archives the occurrence order of process states in chronological order, records the relationship between process state number and occurrence order index, and generates a process state sequence index table. The process obtains the corresponding descriptions of process states and interactive operations from the state-operation mapping results. During the policy playback and trajectory construction phase of reinforcement learning (RL), the state transition logic of the Markov Decision Process (MDP) defined in the mapping description is parsed. Log probes or event listeners are deployed to the business process engine to capture trigger signals of sales personnel's operation interface jumps and business node changes in the CRM system in real time. For example, if the "click the generate contract button" signal is captured, the state is determined to change from "solution negotiation" to "business signing". The logical timestamp of each state change is recorded. (Construct a sequence of state trajectories, in units of operation steps, such as Step1, Step2) ,in Representing the The process state number of each step (e.g., State_03) is traversed. Each state node in the system is assigned an auto-incrementing index number. ( ), and verify adjacent states. and Check whether the transitions between them conform to the preset business logic diagram (e.g., it cannot jump directly from "Initial Contact" to "After-sales Delivery"). If an illegal jump is detected, mark the sequence as abnormal and prompt for manual verification. Then, assign a valid status number to the sequence. Its corresponding index number Store in a key-value database and construct a structure like The ordered dictionary structure provides a standard time step index for Temporal Difference Learning in subsequent reinforcement learning algorithms, generating a process state sequence index table.

[0033] The operation record selection submodule calls the process status number of each process status in the process status sequence index table. Based on the interaction operation record number corresponding to the process status in the status operation mapping result, it filters the interaction operation record content associated with the process status number. The filtered interaction operation records are numbered and collected according to the process status sequence index to generate an interaction record index list. The process state sequence index table calls the process state numbers. During the action selection and execution simulation phase of reinforcement learning, the agent follows the order in the index table. Read the current state sequentially Query the Q-Table or policy network generated by the preceding module. Retrieve in this state The interaction number with the highest expected return (Q-Value) For example, in the state "State_02 (Requirement Confirmation)," the operation with the highest Q value is "Action_50 (Send Selection Guide)." If there are multiple optimal operations with similar Q values, then the one with the highest Q value is chosen. - Use a greedy strategy to make the selection; That is to A random exploratory operation is selected with a probability of [a certain value]. The optimal operation is selected based on probability, and the selected operation number is assigned. Compared with the current index Bind them together to form state-action pairs. It also extracts the standard execution script (such as a script template or document attachment ID) corresponding to the operation number from the historical operation database, appends the extracted script content summary to the binding record, and finally executes it according to... The ascending order will select all Stack the strategies to build a list of strategies to be executed and generate an index of interaction records.

[0034] The sequence arrangement generation submodule determines whether there are any missing positions in the index order of the interactive operation records based on the numbering order of the interactive operation records in the interactive record index list. It then arranges the consecutively numbered interactive operation records in sequence to form a description of the interactive operation progression order and generates the interactive operation sequence result. Based on the order of interaction operation record numbers in the interaction record index list, the action sequences in the list are traversed during the trajectory integrity verification and policy output phase of reinforcement learning. Check for logical breakpoints or gaps in the sequence, such as detecting... Check if the corresponding operation record is missing at the current location. If it is missing, call the preceding state. and subsequent states The missing intermediate operations can be filled in using an interpolation prediction model or by invoking a rule-based default policy. For example, if an operation is missing between "Send Quote" and "Sign Contract," it will automatically be filled in with "Payment Reminder" to ensure the continuity of the sequence. Perform logical coherence smoothing to eliminate conflicts between adjacent operations (such as the incorrect order of "shipping goods" before "receiving payment"), and streamline the continuous interactive operations. Encapsulate the process according to the standard business execution flow format, and attach the expected immediate reward value for each operation. and state transition probability The predicted information is used to form a standardized set of instructions that can be directly parsed by the sales execution system. This serves as the optimal strategy path output by the reinforcement learning agent for the current sales case, generating an interactive operation sequence result.

[0035] Please see Figure 6 The interactive behavior adaptive adjustment module includes a feedback data extraction submodule, an offset position positioning submodule, and a sequence rearrangement generation submodule; The feedback data extraction submodule obtains the interaction operation progress order in the interaction operation sequence result, uses each interaction operation progress number as an index, collects the corresponding conversion feedback data content in the actual execution process, and archives the conversion feedback data with the interaction operation progress number to generate an operation feedback matching list. To obtain the sequence of interactive operations in the results, during the reward signal feedback and policy evaluation phase of reinforcement learning (RL), the system deploys real-time monitoring probes to the business execution terminal to monitor the sequence. Each operation node in It tracks the execution results in the real physical world, captures the instantaneous feedback signals after each operation, and specifically collects two types of core values: one is the action execution status code. (e.g., 200 - Success, 408 - Timeout, 500 - Rejection), and secondly, the incremental business metrics triggered by this action. For example, if the data shows that after the action "send quotation" is performed, the customer clicks to view the file within 30 minutes, then... This is recorded as "Attention +1". If the customer rejects the application outright, then... This is denoted as "Intention Level - 1". The collected raw feedback data is standardized into the scalar reward value required for reinforcement learning. ; The calculation formula is ,in This is a state mapping function (outputs 0 or 1). This is a normalization function (outputting from -1 to 1). To meet Preset weighting coefficients, for example If the operation is successful and produces positive indicators, then ,otherwise The calculated With the corresponding operation number and time step Perform triple binding And store it in the historical feedback database, constructing a database in the form of... The mapping table serves as the basis for subsequent calculation of the strategy value function. and action value function Provide real-world sample support to generate an operational feedback matching list.

[0036] The offset position positioning submodule determines whether there are position numbers in the advancement sequence that are inconsistent with the feedback values ​​and status descriptions based on the conversion feedback data value corresponding to each interactive operation advancement number in the operation feedback matching list and the interactive operation description corresponding to each process status in the status operation mapping result. It then performs filtering and positioning on the position numbers to generate a set of advancement offset position numbers. Based on the conversion feedback data value corresponding to each interaction action progress number in the operation feedback matching list, during the reinforcement learning time difference error (TD Error) calculation and anomaly detection stage, the reward value sequence in the list is traversed. The expected Q-value predicted by the current policy network is called. Calculate each time step Time difference error ,in Set an error tolerance threshold for the discount factor (e.g., 0.9). If the calculated absolute error This indicates actual feedback. With the agent's expectations There is a significant bias, for example, the expectation that the customer will "accept the offer" (high Q value), but the actual feedback is "rejection" (…). (negative), leading to A surge indicates that the location is being determined. If a strategy failure or environmental mutation occurs, record all conditions that meet the requirements. Conditional time step index And extract the corresponding operation number. and deviation value Construct an exception set For each outlier in the set, further analysis is performed. The direction of the sign, if This indicates performance falling short of expectations (and requiring punishment). If the performance is better than expected (and therefore deserves a reward), only the position numbers with negative deviations (i.e., performance that does not meet the standard) are retained as the objects to be optimized, and a set of propulsion offset position numbers is generated.

[0037] The sequence rearrangement generation submodule calls each number in the set of advancement offset position numbers, obtains the interaction operation record corresponding to the number, filters the content of the associated alternative interaction operation record, replaces the abnormal position number in the original interaction operation advancement sequence with the corresponding alternative interaction operation record number, and then recombines the operation sequence to generate the adaptive interaction strategy generation result. In the policy update and re-planning phase of reinforcement learning, each number in the set of offset position numbers is called upon to apply the algorithm to each position marked as an anomaly. and its original operation Query the current Q-Table or policy network and remove the original operations that caused the negative bias. In the remaining set of candidate actions Re-search for the alternative action with the largest Q value: For example, if the original action "telephone follow-up" causes customer resentment, and the system finds that the second-best action "sending SMS care" has a Q value that is second only to the original action and has not been tried, then "sending SMS care" is selected as the alternative. If there are multiple alternative actions; Introducing the Upper Confidence Bound (UCB) algorithm To make exploratory choices, among which To explore constants, To determine the number of state visits, and to balance utilization and exploration, select alternative actions. Inserted into the original sequence Position, and based on the new state transition probability. Re-predict and generate subsequent action sequences It completes the local repair and reconstruction of the entire interaction trajectory, and finally outputs the optimal strategy path after dynamic correction, generating the adaptive interaction strategy generation result.

[0038] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An adaptive generation system for tire changing machine sales strategies based on reinforcement learning, characterized in that, The system includes: The supply and demand characteristic perception module acquires business node information in sales areas, divides sales areas and assigns regional node numbers, collects transaction data, generates trend vectors, analyzes trend consistency, distinguishes supply and demand performance types, and generates supply and demand data characteristic area results. The conversion fluctuation identification module obtains the conversion efficiency of each regional node's interactive operation based on the regional node number assigned with supply and demand feature labels in the supply and demand data feature region results, extracts response delay and protocol changes, identifies the position of conversion efficiency deviation, outputs the operation position number, and generates the conversion fluctuation position result. The status operation association module extracts the customer attributes, latency, protocol changes, feedback signals, and delivery cycles associated with the corresponding operation based on the operation position number in the conversion fluctuation position result, analyzes the relationship with the interactive operation, establishes a corresponding description of the process status and interactive operation, and generates the status operation mapping result. The interactive process orchestration module obtains the actual sequence of process states during process execution based on the corresponding descriptions of process states and interactive operations in the state operation mapping results, selects and arranges interactive operation records in sequence to form a continuous progression description, and generates an interactive operation sequence result.

2. The adaptive generation system for tire-changing machine sales strategy based on reinforcement learning according to claim 1, characterized in that: The supply and demand data feature area results include the level of regional transaction volume change, the type of regional price change direction, the division of regional supply and demand tension, and the characteristics of regional transaction fluctuation. The conversion fluctuation location results specifically include the conversion anomaly occurrence sequence number, the conversion efficiency decline segment, and the conversion unstable operation point. The state operation mapping results include the process state type set, the state corresponding interactive operation category, and the state matching condition elements. The interactive operation sequence results specifically include the interactive operation arrangement order, the operation connection structure form, and the description of the process advancement rhythm.

3. The adaptive generation system for tire-changing machine sales strategy based on reinforcement learning according to claim 1, characterized in that: The supply and demand feature perception module includes a regional node division submodule, a trend vector construction submodule, and a supply and demand type labeling submodule. The regional node division submodule obtains the business execution node information corresponding to each sales region during the tire changer sales process, divides the sales regions according to the business coverage and sales execution boundaries, sets regional node numbers for each sales region, collects the transaction quantity and transaction price change direction information in the region corresponding to the regional node number, and generates a node number mapping list. The trend vector construction submodule calls the transaction volume and transaction price change direction information collected in the node number mapping list, arranges the transaction volume of different time periods under the same regional node number in order, sorts the arrangement order of the transaction price change direction information, forms continuous records respectively, generates transaction volume change vector and transaction price change vector in time order, and generates a trend vector sorting list after summarizing and sorting. The supply and demand type labeling submodule identifies the trend of changes in transaction volume and price over time based on the two types of vector order information corresponding to the node numbers of each region in the trend vector sorting list. It then labels the region node numbers based on the trend performance and matches the labeled node numbers with the corresponding label information to generate supply and demand data feature region results.

4. The adaptive generation system for tire-changing machine sales strategy based on reinforcement learning according to claim 1, characterized in that: The conversion fluctuation identification module includes a node label extraction submodule, an interaction information collection submodule, and an offset position identification submodule; The node tag extraction submodule obtains all region node numbers that have been assigned supply and demand feature tags in the supply and demand data feature region results, filters out node numbers with clear tag identifiers from the region node number list, performs field matching between the region node numbers with tag identifiers and the interaction operation record numbers corresponding to the business execution stage, organizes the corresponding combination content between region node numbers and interaction operation record numbers, and generates a node interaction record index table. The interaction information collection submodule calls the interaction operation record number in the node interaction record index table, extracts the conversion efficiency value, contact response delay information and settlement agreement change information corresponding to the current record from each interaction operation record, organizes and collects the conversion efficiency value, contact response delay information and settlement agreement change information in the order of interaction operation record number, and establishes information grouping according to regional node number to generate node interaction feature set. The offset position identification submodule determines whether the conversion efficiency change direction between consecutive records is consistent based on the order of the conversion efficiency values ​​corresponding to each group of interactive operation records in the node interaction feature set. Then, it filters out the numbers of interactive operation records with inconsistent change directions. At the same time, it performs joint filtering processing on the filtered interactive operation record numbers with contact response delay values ​​and settlement agreement change values ​​to filter out the position numbers with conversion efficiency offset in the interaction sequence and generate conversion fluctuation position results.

5. The adaptive generation system for tire-changing machine sales strategy based on reinforcement learning according to claim 1, characterized in that: The state operation association module includes an operation record extraction submodule, an attribute feature collection submodule, and a state type identification submodule; The operation record extraction submodule obtains the operation position number listed in the conversion fluctuation position result, retrieves the corresponding interaction operation record number from each operation position number, matches the operation record number with the interaction execution record in the business data source, and archives the successfully matched interaction execution records to generate an operation record number set. The attribute feature collection submodule calls each interaction execution record in the operation record number set, extracts the customer attribute feature value, contact response delay value, settlement agreement change status, quotation feedback signal category and delivery cycle time period corresponding to the interaction operation record, and completes the structured organization of the five extracted information with the operation record number as the index, and classifies and combines them according to the field to generate interaction attribute feature groups. The status type identification submodule determines the distribution and combination differences of attribute combinations in the interactive operation sequence based on five items: customer attribute feature value, contact response delay value, settlement agreement change status, quotation feedback signal category, and delivery cycle time period corresponding to each operation record number in the interactive attribute feature group. It then groups operation record numbers with similar attribute combinations into the same process status, establishes a matching structure between interactive operations and process statuses in the order of process statuses, and generates status operation mapping results.

6. The adaptive generation system for tire-changing machine sales strategy based on reinforcement learning according to claim 1, characterized in that: The interaction process orchestration module includes a state order acquisition submodule, an operation record selection submodule, and a sequence arrangement generation submodule. The state sequence acquisition submodule acquires the process state and interactive operation corresponding description in the state operation mapping result, collects the actual occurrence order information of each process state during process execution, organizes and archives the process state occurrence order content in chronological order, records the process state number and occurrence order index relationship, and generates a process state sequence index table. The operation record selection submodule calls the process status number of each process status in the process status sequence index table, filters the interactive operation record content associated with the process status number according to the interactive operation record number corresponding to the process status in the status operation mapping result, and collects the filtered interactive operation records by number according to the process status sequence index to generate an interactive record index list. The sequence arrangement generation submodule determines whether there are any missing positions in the index order of the interactive operation records according to the numbering order of the interactive operation records in the interactive record index list. It then arranges the consecutively numbered interactive operation records in sequence to form an interactive operation progression order description and generates an interactive operation sequence result.

7. The adaptive generation system for tire-changing machine sales strategy based on reinforcement learning according to claim 1, characterized in that: The system also includes: The adaptive adjustment module for interactive behavior, based on the order of interactive operations in the result of the interactive operation sequence, obtains the conversion feedback data corresponding to the order of interactive operations during the actual execution process, locates the deviation position of the interactive operation, selects alternative interactive operation records, rearranges the order of interactive operations, and generates an adaptive interactive strategy generation result. The adaptive interaction strategy generates results including alternative interaction operation options, operation order adjustment schemes, and interaction strategy configuration parameters.

8. The adaptive generation system for tire-changing machine sales strategy based on reinforcement learning according to claim 7, characterized in that: The interactive behavior adaptive adjustment module includes a feedback data extraction submodule, an offset position positioning submodule, and a sequence rearrangement generation submodule. The feedback data extraction submodule obtains the interaction operation progress order in the interaction operation sequence result, uses each interaction operation progress number as an index, collects the corresponding conversion feedback data content in the actual execution process, and archives the conversion feedback data with the interaction operation progress number to generate an operation feedback matching list. The offset position positioning submodule determines whether there are position numbers in the advancement sequence that are inconsistent with the feedback values ​​and state descriptions based on the conversion feedback data value content corresponding to each interactive operation advancement number in the operation feedback matching list and the interactive operation description corresponding to each process state in the state operation mapping result. It then performs filtering and positioning on the position numbers to generate a set of advancement offset position numbers. The sequence rearrangement generation submodule calls each number in the set of advancement offset position numbers, obtains the interaction operation record corresponding to the number, filters the content of the associated alternative interaction operation record, replaces the abnormal position number in the original interaction operation advancement sequence with the corresponding alternative interaction operation record number, and then recombines the operation sequence to generate the adaptive interaction strategy generation result.

Images