Multi-agent based mall management method and related device

Abstract

This invention provides a multi-agent-based e-commerce platform management method, comprising: deploying product agents for products in the e-commerce platform, and using these agents to perceive internal inventory status and external market environment, generate autonomous price adjustment decisions, and execute these decisions; when the pricing or promotional actions of a product agent affect the profit of related product categories, the product agent pays virtual collaborative points to the product agents of the related product categories as compensation, and rolls back pricing or promotional actions when virtual collaborative points are insufficient; ranking each product agent based on its virtual collaborative point balance, assigning weight to the top-ranked product agents in subsequent traffic allocation, and having each product agent use its virtual collaborative points as bidding chips to compete for traffic positions, obtaining traffic positions based on the bidding results; and sorting and displaying products in the e-commerce platform according to traffic positions. This invention effectively improves the management efficiency of e-commerce platforms.

Description

Technical Field

[0001] This invention relates to the field of e-commerce management technology, and in particular to an e-commerce management method and related equipment based on multi-agent systems. Background Technology

[0002] As e-commerce platforms continue to expand their business scale and enrich their product categories, traditional e-commerce management systems are gradually becoming unable to cope with increasingly complex operational needs. Currently, mainstream e-commerce management systems generally adopt an architecture of "centralized rule engine + independent early warning module." While business modules such as pricing, inventory, and marketing are encapsulated as intelligent agents, their actual operating logic is essentially based on local optimization of hard-coded rules, lacking an effective collaboration mechanism between modules. Currently, existing e-commerce platform management methods suffer from several problems. First, in price management, pricing strategies are typically predetermined by operations staff based on cost-plus pricing or fixed gross margins, failing to respond quickly to real-time market changes. Second, in category management, different product management modules make independent decisions. For example, when a product significantly reduces its price to clear inventory, this action can severely erode the profits of related categories, and the price reduction module completely disregards this cross-category profit impact, leading to a decline in the platform's overall gross profit. Third, in traffic allocation, e-commerce platform traffic resources are usually distributed centrally by a recommendation algorithm, leaving products without the means to actively seek exposure. Long-tail products and new products struggle to gain display opportunities, and the platform's traffic monetization efficiency decreases year by year due to homogeneous competition. Therefore, there is an urgent need for a management method that endows e-commerce platform products with autonomous perception, autonomous decision-making, and collaborative game-theoretic capabilities. This would address the problems of existing e-commerce platform management methods relying on hard-coded local optimization logic for independent decision-making, lacking quantifiable collaborative compensation mechanisms for cross-category profit impacts, and decentralized competitive mechanisms for traffic allocation. These issues result in local optimal decisions failing to converge to the global optimal, damage to the platform's overall gross profit, and low traffic allocation efficiency. Summary of the Invention

[0003] This invention provides a multi-agent-based e-commerce platform management method to improve management efficiency. By deploying product agents to products within the e-commerce platform for autonomous perception and pricing decisions, a compensation and rollback mechanism for cross-category profit impact is established through virtual collaborative points. Traffic allocation is weighted and bid on based on the ranking of points held by the agents, thereby achieving decentralized pricing collaboration and autonomous traffic allocation. This solves the problems of existing e-commerce platform management methods that rely on hard-coded local optimization logic for independent decision-making, lacking a quantifiable collaborative compensation mechanism for cross-category profit impact and a decentralized competition mechanism for traffic allocation. These methods result in local optimal decisions failing to converge to the global optimum, overall platform gross profit being damaged, and low traffic allocation efficiency.

[0004] In a first aspect, embodiments of the present invention provide a multi-agent-based e-commerce management method, the method comprising: Deploy product intelligence agents for products in the mall, and use these product intelligence agents to perceive internal inventory status and external market environment, generate autonomous price adjustment decisions, and execute these autonomous price adjustment decisions. When the pricing or promotional actions of the product agent affect the profit of related product categories, the product agent pays virtual collaboration points to the product agents of related product categories as compensation. If the virtual collaboration points are insufficient, the pricing or promotional actions are rolled back. The ranking is based on the virtual collaboration points held by each product agent. The top-ranked product agents are given weight in the subsequent traffic allocation. Each product agent uses its virtual collaboration points as a bidding chip to participate in the competition for traffic positions and obtains traffic positions based on the bidding results. Products in the online store are sorted and displayed based on traffic volume.

[0005] Optionally, the step of sensing the internal inventory status and external market environment through the product intelligent agent, generating an autonomous price adjustment decision, and executing the autonomous price adjustment decision specifically includes: The product intelligent agent acquires internal status data and external market environment data. The internal status data includes the current inventory depth, historical conversion rate, and refund rate of the product. The external market environment data includes competitor prices and social media sentiment information. The market heat index is calculated based on the internal status data and the external market environment data. The price adjustment coefficient is determined by querying the built-in dynamic decision table based on the market heat index and the current inventory days. The current product price will be adjusted according to the aforementioned price adjustment factor.

[0006] Optionally, a market heat scalar is calculated based on the internal state data and the external market environment data, specifically including: Multiply the logarithm of the site search index by the sum of the social sentiment score and 1, and then subtract the product of the competitor's price difference rate and the preset coefficient. The social sentiment score is obtained by scoring social media keywords with sentiment using an open-source sentiment analysis model and mapping it to a normalized interval. The market popularity scalar value is greater than 0, indicating that the market popularity is rising, and less than 0, indicating that the market popularity is falling.

[0007] Optionally, when the pricing or promotional actions of the product agent affect the profit of related product categories, the product agent pays virtual collaboration points to the product agents of the related product categories as compensation. If the virtual collaboration points are insufficient, the pricing or promotional actions are rolled back. Specifically, this includes: Construct a category network graph, which is built according to the purchasing behavior correlation between each product, and the edge weight represents the substitution strength or complement strength; When the product agent performs a pricing or promotional discount action, the profit impact of the action on the associated categories in the category network graph is calculated. If the action causes profit damage to the related product category, the product agent pays virtual collaborative points to the product agent corresponding to the affected related product category as compensation. The amount of compensation is the product of the absolute value of the profit loss and the preset profit and loss exchange coefficient. If the virtual collaboration points held by the product agent are insufficient to pay the compensation, the pricing or promotional discount action will be rolled back.

[0008] Optional, also includes: By listening to the pricing history of each commodity agent through an arbitration agent, the price fluctuation entropy is calculated. When any product agent adjusts its price more than a preset number of times within a preset time period and the average price adjustment exceeds a preset range threshold, a cooling command is issued to the product agent through the arbitration agent. The control agent for the product will only perform the action of maintaining the current price within a preset cooling time according to the cooling command.

[0009] Optionally, the method further includes: Each product agent maintains a Q table, whose state space includes inventory level tiers, market popularity scalar tiers, and virtual collaborative points balance tiers, and whose action space includes multiple preset price adjustment ranges; After each decision is executed, the Q table is updated immediately as the sum of the actual incremental profit from the transaction and the virtual collaborative points penalty paid. The product agent adopts an ε-greedy strategy to randomly explore the action space with a preset probability.

[0010] Optionally, the method further includes: During the offline phase, the decision-making trajectories of each product agent within a preset number of days are replayed, and the regret value at each decision moment is calculated. The regret value is the Q-value difference between the currently selected action and the optimal action afterwards. Decision trajectories with regret values ​​exceeding a preset threshold are used as negative samples and are forcibly input into the Q-table of the corresponding product agent for reverse gradient updates.

[0011] Secondly, embodiments of the present invention also provide a multi-agent-based e-commerce management device, the multi-agent-based e-commerce management device comprising: The generation module is used to deploy product intelligence agents for products in the mall, and through the product intelligence agents, perceive the internal inventory status and external market environment, generate autonomous price adjustment decisions, and execute the autonomous price adjustment decisions. The rollback module is used to compensate the product agent for its pricing or promotional actions when the pricing or promotional actions of the product agent affect the profits of related product categories. If the virtual collaborative points are insufficient, the pricing or promotional actions are rolled back. The processing module is used to rank the product agents based on the balance of virtual collaborative points held by each product agent, and to give weight to the top-ranked product agents in subsequent traffic allocation. Each product agent participates in the competition for traffic positions by using the virtual collaborative points it holds as bidding chips, and obtains traffic positions based on the bidding results. The sorting and display module is used to sort and display products in the mall based on traffic volume.

[0012] Thirdly, embodiments of the present invention provide an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps in the multi-agent-based e-commerce management method provided in embodiments of the present invention.

[0013] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the multi-agent-based e-commerce management method provided in the embodiments of the present invention.

[0014] In this embodiment of the invention, product intelligent agents are deployed for products in the online store. These agents perceive internal inventory status and the external market environment, generate autonomous price adjustment decisions, and execute these decisions. When the pricing or promotional actions of a product intelligent agent affect the profits of related product categories, the product intelligent agent pays virtual collaborative points to the product intelligent agents of the related product categories as compensation. If the virtual collaborative points are insufficient, the pricing or promotional actions are rolled back. Product intelligent agents are ranked according to their virtual collaborative point balances, and those ranked higher are given weight in subsequent traffic allocation. Each product intelligent agent uses its virtual collaborative points as bidding chips to compete for traffic positions and obtains a traffic position based on the bidding results. Products in the online store are then sorted and displayed according to their traffic positions. This invention deploys intelligent agents for products in an online marketplace to enable autonomous perception and pricing decisions. It establishes a compensation and rollback mechanism for cross-category profit impact through virtual collaborative points, and performs weighted traffic allocation and points bidding based on the ranking of the points held by the intelligent agents. This achieves decentralized pricing collaboration and autonomous traffic allocation, solving the problems of existing online marketplace management methods that use hard-coded local optimization logic for independent decision-making, lack a quantifiable collaborative compensation mechanism for cross-category profit impact and a decentralized competition mechanism for traffic allocation. This results in local optimal decisions failing to converge to the global optimum, damage to the platform's overall gross profit, and low traffic allocation efficiency. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of a multi-agent-based e-commerce management method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a multi-agent-based e-commerce management device provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] like Figure 1 As shown, Figure 1 This is a flowchart of a multi-agent-based e-commerce management method provided by an embodiment of the present invention. The multi-agent-based e-commerce management method includes the following steps: 101. Deploy product intelligence agents for products in the mall, and use these agents to perceive internal inventory status and external market environment, generate autonomous price adjustment decisions, and execute these decisions.

[0019] 102. When the pricing or promotional actions of the product agent affect the profit of related product categories, the product agent shall pay virtual collaboration points to the product agents of related product categories as compensation. If the virtual collaboration points are insufficient, the pricing or promotional actions shall be rolled back.

[0020] 103. Rank the products based on the balance of virtual collaboration points held by each product agent, and give weight to the top-ranked product agents in subsequent traffic allocation. Each product agent participates in the competition for traffic positions by using its virtual collaboration points as bidding chips, and obtains traffic positions based on the bidding results.

[0021] 104. Sort and display products in the mall according to traffic volume.

[0022] In this embodiment, the aforementioned multi-agent-based e-commerce platform management method can be applied to an e-commerce platform. This platform can be built on a server-based or distributed architecture and includes a data interface (for data uploaded by sensors or users), a knowledge database, and a knowledge database construction program. The data interface can be used to deploy product agents for goods in the e-commerce platform. These agents can perceive internal inventory status and the external market environment, generate autonomous price adjustment decisions, and execute these decisions. The knowledge database construction program can be used to construct the knowledge database, which is specifically designed to provide additional relational information for identified data entities, thereby enhancing the data recognition system's understanding of the content.

[0023] The aforementioned shopping mall can be an e-commerce transaction platform operating in a network environment, including business elements such as multiple products, product inventory, pricing strategies, traffic resources, and user access traffic.

[0024] The aforementioned product intelligence agents can be software proxy entities deployed for each product in the online store. These agents possess real-time perception capabilities for internal status data, as well as the ability to access and analyze external market environment data. The internal status data includes current product inventory depth, historical conversion rates, and refund rates; the external market environment data includes current competitor prices and sentiment reviews on social media.

[0025] The aforementioned autonomous price adjustment decision can be a price adjustment plan independently calculated and generated by the product's intelligent agent based on its perceived multi-dimensional information and according to its built-in decision-making logic. The autonomous price adjustment decision includes the adjustment magnitude and direction.

[0026] By deploying product intelligence agents on the products in the online store, and using these agents to perceive internal inventory status and the external market environment, autonomous price adjustment decisions can be generated. This allows for the capture of minute-level fluctuations in competitor prices and changes in social media sentiment, enabling real-time responses to pricing decisions.

[0027] The aforementioned related product categories can be other product categories that are related to the current product purchase behavior. Related product categories include complementary product categories, substitute product categories, etc.

[0028] The aforementioned virtual collaborative points can be virtual digital points issued within the system with a fixed total amount. Virtual collaborative points serve as a settlement medium for cross-category profit impact compensation between commodity intelligent agents, transforming business externalities that are difficult to quantify directly into measurable and settlementable digital symbols.

[0029] The aforementioned rollback can be implemented when a product agent needs to pay virtual collaborative points to a related category due to pricing or promotional actions, but the balance is insufficient. In this case, the system can cancel the pricing or promotional actions already performed by the product agent, restoring the product price to the state before the actions were performed.

[0030] It should be noted that when the pricing or promotional actions of a product agent affect the profits of related product categories, the product agent must pay virtual collaboration points as compensation to the product agents corresponding to the affected related product categories. If the virtual collaboration points held by the product agent are insufficient to pay the compensation, the pricing or promotional actions are rolled back. Through the virtual collaboration points and rollback mechanism, the profit impact of the product agent's autonomous pricing behavior on related product categories can be transformed into an internal cost that the agent needs to bear. This forces autonomous pricing behavior to automatically consider the profit security of related product categories, effectively suppressing malicious competitive behavior that damages the profits of related product categories.

[0031] Specifically, when a product agent performs a pricing or promotional action involving price reductions or discounts, it can be assessed whether this action will negatively impact the profits of other product agents in the same related product category within the marketplace. If it is determined that the pricing or promotional action has harmed the profits of the related product category, a virtual collaborative points compensation mechanism is triggered. The product agent executing the action must, according to preset redemption rules, spend the corresponding number of points from its virtual collaborative points balance to pay the affected product agents in the related product category, thus providing economic compensation for their profit loss. Before executing the payment operation, it is first checked whether the virtual collaborative points balance currently held by the product agent executing the action is sufficient to complete the payment. If the balance is sufficient, the points payment is completed, and the pricing or promotional action takes effect. If the balance is insufficient, it is determined that the product agent does not have the capacity to bear this externality compensation, and an action rollback mechanism is triggered to cancel the pricing or promotional action already executed by the product agent, restoring the product price to its state before the action was executed. The aforementioned preset redemption rules can be pre-set redemption rules. The redemption rules can be a calculation standard and parameter system that transforms the abstract profit loss caused by the intelligent agent's autonomous pricing or promotional actions on related product categories into specific virtual collaborative points payment quantities.

[0032] The aforementioned ranking can be considered as being at the beginning of the sorting sequence when each commodity agent is sorted according to its virtual collaborative points balance from high to low.

[0033] The aforementioned traffic allocation can be described as the process of distributing user access traffic resources of an online store to various products according to certain rules. These access traffic resources include homepage display positions, category recommendation positions, and weighted search rankings.

[0034] The aforementioned weighting can be used in the subsequent traffic allocation process to assign higher allocation priority or a larger allocation share to products corresponding to top-ranked product agents than to other products.

[0035] The aforementioned traffic slots can be allocated exposure positions on the e-commerce page used to display products. They are the core carriers of e-commerce traffic, including homepage banner slots, category recommendation slots, and top slots on search results pages.

[0036] The aforementioned bidding chips can be the amount of virtual collaborative points used by a commodity agent when competing for a traffic slot, as the consideration it is willing to pay to obtain that traffic slot.

[0037] In one possible implementation, Product Agent C is a high-profit new product, and Product Agent D is a low-profit clearance item. Product Agents C and D simultaneously bid for a banner ad slot on the homepage of the digital category. Product Agent C bids 15 ST, while Product Agent D bids 20 ST to avoid failing to sell. Product Agent D wins the ad slot with the highest bid, but the final settlement price is only the second highest bid, 15 ST. Product Agent D profits because it won the traffic but paid less than its expected amount; Product Agent C, having not won but not incurring high bidding costs, has no regrets. This approach maximizes the overall efficiency of the mall's advertising resource allocation and prevents malicious price gouging.

[0038] Specifically, at the end of each set statistical period, the virtual collaboration points balance held by each product agent in the marketplace is tallied, and all product agents are ranked in descending order of their balance values. Based on the ranking results, the top-ranked product agents are given weighted preferential treatment in subsequent traffic allocation processes; that is, the products corresponding to the top-ranked product agents will receive higher allocation priority or more traffic exposure quotas in the next period's traffic allocation. Simultaneously, each product agent uses its virtual collaboration points as bidding chips to compete for traffic slots, and obtains traffic slots based on the bidding results. Through this traffic allocation weighted and points-based bidding mechanism driven by virtual collaboration point balance ranking, the marketplace's traffic allocation no longer relies on the unified decision-making of a central recommendation algorithm. Instead, it achieves self-organized optimization through decentralized points competition among product agents, causing traffic resources to spontaneously tilt towards products with higher marginal profits and lower negative ecological externalities, thereby improving the overall monetization efficiency of the marketplace's traffic and the platform's economic benefits.

[0039] The above sorting and display can be a process of arranging and presenting products on the front-end page of the e-commerce platform according to the order determined by the bidding results of each product agent in the traffic competition.

[0040] In this embodiment, the present invention can sort and display each product on the front-end page of the mall according to the traffic positions obtained by each product agent and their competition results, so that products that obtain high-value traffic positions can get more priority exposure positions.

[0041] In this embodiment of the invention, product intelligent agents are deployed for products in the online store. These agents perceive internal inventory status and the external market environment, generate autonomous price adjustment decisions, and execute these decisions. When the pricing or promotional actions of a product intelligent agent affect the profits of related product categories, the product intelligent agent pays virtual collaborative points to the product intelligent agents of the related product categories as compensation. If the virtual collaborative points are insufficient, the pricing or promotional actions are rolled back. Product intelligent agents are ranked according to their virtual collaborative point balances, and those ranked higher are given weight in subsequent traffic allocation. Each product intelligent agent uses its virtual collaborative points as bidding chips to compete for traffic positions and obtains a traffic position based on the bidding results. Products in the online store are then sorted and displayed according to their traffic positions. This invention deploys intelligent agents for products in an online marketplace to enable autonomous perception and pricing decisions. It establishes a compensation and rollback mechanism for cross-category profit impact through virtual collaborative points, and performs weighted traffic allocation and points bidding based on the ranking of the points held by the intelligent agents. This achieves decentralized pricing collaboration and autonomous traffic allocation, solving the problems of existing online marketplace management methods that use hard-coded local optimization logic for independent decision-making, lack a quantifiable collaborative compensation mechanism for cross-category profit impact and a decentralized competition mechanism for traffic allocation. This results in local optimal decisions failing to converge to the global optimum, damage to the platform's overall gross profit, and low traffic allocation efficiency.

[0042] It is understood that in the specific embodiments of this application, data such as product data, price data, and traffic data are involved. When the embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0043] Optionally, the steps for generating and executing autonomous price adjustment decisions by sensing internal inventory status and external market environment through a product intelligent agent include: Acquire internal status data and external market environment data through product intelligent agents; The market heat index is calculated based on internal status data and external market environment data. The price adjustment coefficient is determined by querying the built-in dynamic decision table based on the market heat index and the current inventory days. The current product price is adjusted according to the price adjustment factor.

[0044] In this embodiment, the aforementioned product intelligence agent can be a software agent entity deployed on each product in the online store.

[0045] The aforementioned internal status data can be data stored within the e-commerce platform's internal system that relates to the current operational status of the product itself. Internal status data includes the current product's inventory depth, historical conversion rate, and refund rate. Inventory depth indicates the sufficiency of product supply and the urgency of clearance sales; historical conversion rate indicates the product's sales appeal and pricing rationality; and refund rate indicates the product's quality performance or user satisfaction.

[0046] The aforementioned external market environment data can be data and information from outside the e-commerce system that reflects market competition and consumer opinion. External market environment data includes competitor pricing and social media sentiment information. Competitor pricing can be the current selling price of top-tier competitors that directly compete with the current product, obtained through legally publicly available data interfaces, used to characterize the intensity of market competition and serve as a price benchmark. Social media sentiment information can be quantitative scoring results obtained by analyzing the sentiment tendencies of keyword text related to the current product on social media platforms using open-source sentiment analysis models, used to characterize consumers' market reputation and public opinion surrounding the current product.

[0047] The aforementioned market popularity scalar can be a single numerical indicator used to quantitatively represent the overall popularity of the current product's market segment, calculated comprehensively based on internal status data and external market environment data. The positive or negative sign of the market popularity scalar indicates the direction of change in market popularity: a positive value indicates rising market popularity, making it appropriate to maintain or increase prices; a negative value indicates declining market popularity, requiring price reductions and promotions.

[0048] The aforementioned dynamic decision table can be a two-dimensional lookup table structure pre-stored within the product's intelligent agent. One dimension of the dynamic decision table can be the inventory days level, and the other dimension can be the market popularity scalar trend level, such as rising, stable, or falling. Each cell in the dynamic decision table stores a pre-set price adjustment coefficient value. The dynamic decision table can discretize multi-dimensional continuous perceptual inputs into a finite combination of states and map them into specific price adjustment instructions.

[0049] The aforementioned price adjustment coefficient can be a numerical coefficient output by the product agent after querying the built-in dynamic decision table based on the market popularity scalar and the current inventory days. The price adjustment coefficient can be in the range of upper and lower limits, such as -8% to +3%. The price adjustment coefficient is used to indicate the percentage adjustment that the current product price should be.

[0050] In one possible implementation, for example, at 10:05 AM, a competing platform lowers the price of a certain brand of Bluetooth headphones to 199 yuan, and related price comparison discussions appear on Weibo. The product agent detects the competitor's price change, indicating an increase in the competitor's price difference rate; the related Weibo sentiment score is +0.6, indicating an increase in popularity. The market popularity scalar calculation score is -2.3, triggering a price reduction. The product agent checks the dynamic decision table and, based on the current inventory level DOI=15 days, autonomously determines the price adjustment coefficient PAC=-5%; at 10:08 AM, the price of the certain brand of Bluetooth headphones in the online store is adjusted to 199 yuan and synchronized on the homepage.

[0051] It should be noted that by acquiring internal state data and external market environment data through the product intelligent agent, a comprehensive real-time perception of the product's operational status and market competition situation is achieved. By calculating a market heat scalar based on internal state data and external market environment data, and then determining the price adjustment coefficient by querying a built-in dynamic decision table based on the market heat scalar and the current inventory days, the high-dimensional continuous perceptual input is dimensionality-reduced and mapped into discrete, structured pricing instructions, realizing the automatic generation and interpretable output of pricing rules. By adjusting the current product price according to the price adjustment coefficient, a fully automated closed loop from perception to decision-making to execution is achieved, without central approval or manual intervention, significantly improving the efficiency of pricing decision-making. This invention extends the agent's responsiveness from internal event-driven to external market environment-driven, and introduces multi-dimensional unstructured data to participate in structured pricing decisions, breaking the limitations of traditional pricing that relies on static cost-plus pricing or single competitor price comparisons.

[0052] Optionally, the steps for calculating the market heat scalar based on internal status data and external market environment data include: Multiply the logarithm of the site search index by the sum of the social sentiment score and 1, and then subtract the product of the competitor's price difference rate and the preset coefficient. Among them, the social sentiment score is obtained by scoring social media keywords with sentiment using an open-source sentiment analysis model and mapping it to a normalized interval. A market popularity scalar greater than 0 indicates an increase in market popularity, while a value less than 0 indicates a decrease in market popularity.

[0053] In this embodiment, the aforementioned in-site search index can be a real-time frequency or popularity statistic of searches performed by users within the online store by entering the current product name or related keywords through the in-site search engine. The in-site search index can be generated in real time by the online store's backend search log system, and can be the number of searches in the past hour, the number of searches in the past 24 hours, or a normalized search popularity score, etc.

[0054] The aforementioned social sentiment score can be a quantitative score mapped to a normalized interval of [-1, 1] after analyzing the sentiment tendency of keyword texts related to the current product on social media platforms using an open-source sentiment analysis model. A value close to 1 indicates that social media discussions are predominantly positive (e.g., positive reviews, recommendations), a value close to -1 indicates predominantly negative (e.g., negative reviews, complaints), and a value close to 0 indicates neutral sentiment or extremely low discussion activity. The aforementioned open-source sentiment analysis model can be a sentiment analysis model built based on deep learning or machine learning, such as a sentiment classifier fine-tuned based on BERT. The aforementioned BERT-fine-tuned sentiment classifier achieves sentiment discrimination by adding a classification layer on top of a pre-trained model, using the final hidden state labeled [CLS] as the semantic representation of the entire sentence, and training end-to-end with a small amount of labeled data. The aforementioned open-source sentiment analysis model can analyze multimodal data such as facial expressions, voice tone, text semantics, or physiological signals, mapping human emotional states to emotions such as joy, anger, sadness, or continuous dimensions. The essence of the open-source sentiment analysis model is the pattern recognition and reasoning process from observed features to emotional labels.

[0055] The aforementioned competitor price difference rate can be the relative deviation ratio between the current selling price of the current product and the current selling price of top-tier competitors in the market. It can be calculated by subtracting the current product's selling price from the competitor's current selling price, and then dividing the result by the current product's selling price.

[0056] The aforementioned preset coefficient, denoted as λ, is a fixed constant pre-set by the e-commerce platform operators based on factors such as the platform's gross profit margin strategy, market competition intensity, and product category characteristics. This preset coefficient is used to control the penalty imposed by competitor price difference rates on market popularity metrics.

[0057] Specifically, the formula for calculating the market popularity scalar is as follows:

[0058] in, Indicates market activity level; MHS represents the preset coefficient; MHS>0 indicates that market activity is rising, and it is appropriate to raise prices or maintain them; MHS<0 indicates that market activity is declining, and it is necessary to lower prices to promote sales.

[0059] It should be noted that by taking the logarithm of the site search index and multiplying it by the sum of the social sentiment score and 1, a quantitative model of the nonlinear synergistic amplification or inhibition effect between site search demand and social opinion is achieved, enabling positive signals to reflect the true market heat after the coupling of the two. Subtracting the product of the competitor's price difference rate and the preset coefficient introduces a negative hedging mechanism for competitive pressure on market heat, and the preset coefficient enables flexible adjustment of competitive sensitivity. When the market heat scalar is greater than 0, it indicates that the combined impact of positive signals such as site search heat and social sentiment tendencies outweighs the negative pressure of competitor price differences, and the overall market heat is on the rise. The product agent should tend to adopt a strategy of raising or maintaining prices. When the market heat scalar is less than 0, it indicates that negative signals such as competitor price reduction pressure or negative social sentiment dominate, and the market heat is on the decline. The product agent should tend to adopt a strategy of price reduction and promotion. The market heat scalar provides an accurate and directly indexable quantitative dimension for subsequent decision table queries, ensuring the logical consistency and reproducibility of the entire autonomous pricing decision-making chain.

[0060] Optionally, when the pricing or promotional actions of the product agent affect the profit of related product categories, the product agent pays virtual collaboration points to the product agents of related product categories as compensation. The steps for rolling back pricing or promotional actions when virtual collaboration points are insufficient include: Construct a category network diagram; When the product agent performs a pricing or promotional discount action, calculate the profit impact of the action on the related categories in the category network graph; If an action causes profit damage to a related product category, the product agent pays virtual collaboration points to the product agent corresponding to the affected related product category as compensation. The amount of compensation is the product of the absolute value of the profit loss and the preset profit and loss exchange coefficient. If the virtual collaboration points held by the product agent are insufficient to cover the compensation, the pricing or promotional discount action will be rolled back.

[0061] In this embodiment, the category network graph is constructed according to the purchasing behavior correlations between various products, and the edge weights represent the substitution strength or complementarity strength. The nodes of the category network graph are the categories corresponding to each product, and the edges connect two related categories. The substitution strength is used to quantify the strength of the substitution relationship between two categories. The larger the absolute value of the substitution strength, the stronger the consumer's perception of these two categories as an either-or choice when making a purchase. That is, a price reduction of one product will significantly attract consumers who would otherwise purchase the other product, resulting in a loss of sales and profits for the latter. The complementarity strength is used to quantify the strength of the complementarity relationship between two categories. The greater the complementarity strength, the more likely consumers are to purchase one product while simultaneously purchasing the other. That is, a price reduction or promotion of one product will promote the sales of the other product through a cross-purchasing effect, resulting in a positive profit gain.

[0062] The aforementioned profit impact can be the change in profit caused by pricing or promotional discount actions performed by the product agent on related categories in the category network graph. The change in profit can be calculated by combining the price change magnitude, historical sales data of related categories, and edge weights in the category network graph.

[0063] The absolute value of the aforementioned profit loss is used to quantify the specific amount of profit damage caused to related product categories by the pricing or promotional discount actions performed by the product agent. This amount serves as the base for calculating the amount of virtual collaborative points compensation.

[0064] The aforementioned preset profit and loss conversion factor can be a pre-set profit and loss conversion factor. This factor is used to convert the actual amount of profit loss into the amount of virtual collaborative points paid out. A larger profit and loss conversion factor means that the same amount of profit loss requires the payment of more points, resulting in stronger compensation.

[0065] The aforementioned virtual collaborative points balance can be the current virtual collaborative points stock held by each product agent, which can be used to pay for externality compensation or participate in traffic competition. The virtual collaborative points balance can be allocated according to preset rules when the product agent is initialized, and dynamically updated each time points are paid, traffic is consumed in a bidding process, or periodic settlement is completed.

[0066] The aforementioned rollback can be performed when the virtual collaboration points held by the product agent are insufficient to pay compensation, thereby canceling the pricing or promotion actions already executed by the product agent and restoring the product price and promotion status to the state before the action was executed.

[0067] Understandably, when a product agent performs pricing or promotional discount actions, an externality check is required. The product agent calculates the profit impact of the pricing or promotional discount action on related categories in the category network graph. If the profit impact is less than 0, it means the interests of neighbors have been harmed, and the product agent must pay virtual cooperation points ST=| to the affected neighbor agents. |×ρ, where, For profit impact, ρ is the profit-loss exchange factor. This payment is automatically completed by the atomic settlement protocol between agents to prevent default. If the virtual collaborative points held by the product agent are insufficient to pay the externality compensation, its action is forcibly rolled back, and the decision is added to the historical forbidden set, prohibiting recent re-attempts. The above externality verification can be a pre-check operation performed when the product agent executes a pricing or promotional discount action, before the pricing or promotional discount action officially takes effect. Externality verification is performed by finding the current product node and its neighboring nodes in the category network graph, calculating the expected profit impact of the action on the category represented by the neighboring nodes, to determine whether the action generates a positive or negative externality, and determining the compensation amount corresponding to a negative externality.

[0068] In one possible implementation, for example, at the end of each natural day, the virtual collaboration points balance held by each product agent is counted, and the top 10% of product agents with the highest balance will receive additional weighting in the traffic auction the next day, thereby incentivizing product agents to proactively consider the overall ecosystem health when making autonomous decisions.

[0069] In another possible implementation, for example, Product Agent A clears out its inventory of SLR cameras by voluntarily reducing prices by 15%, resulting in a gross profit decrease of 50 yuan per unit. This voluntary price reduction will severely impact the associated purchase rate of its strongly complementary Product Agent B, memory cards. Before Product Agent A executes its action, the system queries the category network diagram and finds that the complementarity strength coefficient between Product Agent A and Product Agent B is 0.8. It is estimated that the associated sales of Product Agent B will decrease by 20%, equivalent to a profit loss of 40 yuan. Product Agent A needs to pay Product Agent B virtual collaboration points ST = 40 × 0.5 = 20. Since Product Agent A has sufficient balance, the action takes effect. After receiving the ST compensation, Product Agent B proactively lowers its own memory card price, further stimulating bundled sales. The total sales volume of the camera + memory card package significantly increases compared to the individual price reduction model, and the platform's total profit is not affected.

[0070] It should be noted that by constructing a category network graph and setting edge weights based on the correlation of purchasing behavior between products, a structured relational topology is provided for the quantitative calculation of cross-category profit impact. This transforms the originally abstract and difficult-to-measure business relationships between categories into computable and traversable graph structure data. By calculating the profit impact of pricing or promotional discount actions on related categories in the category network graph when product agents perform pricing or promotional discount actions, a quantitative boundary is established between positive and negative externalities, providing a clear judgment basis and accurate monetary base for triggering points compensation. Through virtual collaborative points compensation, actions are allowed to take effect when points are sufficient but compensation costs are borne, and rollback is forcibly executed when the points balance is insufficient, fundamentally curbing short-sighted pricing behavior that damages the profits of related categories.

[0071] Optionally, the method also includes: By listening to the pricing history of each commodity agent through an arbitration agent, the price fluctuation entropy is calculated. When any product agent adjusts its price more than a preset number of times within a preset time period and the average price adjustment exceeds a preset range threshold, a cooling command is issued to the product agent through the arbitration agent. The intelligent agent controlling the goods will only perform actions to maintain the current price within a preset cooling period, based on the cooling command.

[0072] In this embodiment, the aforementioned arbitration agent may be responsible for monitoring, arbitrating, and intervening in the pricing behavior of other product agents. The arbitration agent does not directly participate in the pricing or sales decisions of the product, but rather acts as a regulator, assessing the compliance of the product agents' behavior and the stability of their strategies, and issuing binding instructions when necessary.

[0073] The aforementioned pricing history can be a complete sequence of data continuously recorded by the arbitration agent, including the timestamps, adjustment directions, and adjustment magnitudes of every price adjustment action performed by each commodity agent over a past period.

[0074] The aforementioned price fluctuation entropy can be a statistical indicator calculated by the arbitration agent through monitoring the pricing history of various commodity agents. It is used to quantify the degree of disorder and predictability of the price adjustment behavior of commodity agents. Price fluctuation entropy reflects the disorder and intensity of the agent's price adjustment behavior. The higher the entropy value, the more frequent the price adjustment behavior, the more irregular the magnitude, and the more difficult it is to predict its next move. When the entropy value exceeds the normal range, it indicates that the agent has perceptual failures, algorithm degradation, or malicious probing behavior.

[0075] The aforementioned preset time period can be a pre-set time period, such as the past 1 hour, the past 30 minutes, etc.

[0076] The aforementioned preset threshold number can be a pre-defined threshold. The threshold number is used to limit the upper limit of the number of price adjustments that the product agent is allowed to perform within a preset time period.

[0077] The aforementioned preset amplitude threshold can be a pre-set amplitude threshold, which is used to limit the upper limit of the average price adjustment amplitude executed by the commodity intelligent agent within a preset time period.

[0078] The aforementioned cooling-off order can be a mandatory constraint command issued by the arbitration agent to a product agent determined to be abnormal. The cooling-off order requires the target product agent to restrict its pricing decision-making behavior during the order's effective period, allowing it only to maintain the current price and prohibiting any price adjustment operations.

[0079] The aforementioned preset cooling-off period can be a pre-set cooling-off period, such as 30 minutes, 40 minutes, etc. During the cooling-off period, the commodity agent bound by the cooling command can only perform the action of maintaining the current price and cannot perform any price adjustment action.

[0080] The aforementioned action of maintaining the current price can be that, after receiving a cooling instruction, the product agent keeps the current product price unchanged and does not perform any upward or downward adjustments.

[0081] In one possible implementation, the arbitration agent monitors the pricing history of all product agents and calculates the price fluctuation entropy. If product agent A adjusts its price more than 5 times within one hour with an average magnitude greater than 5%, the arbitration agent determines that product agent A is maliciously testing the market or has an algorithmic malfunction. It then issues a cooling-off command to product agent A, forcing it to only perform the action of maintaining the current price for the next 30 minutes, thus ensuring the basic stability of the mall's pricing system.

[0082] In another possible implementation, for example, product agent B, having captured incorrect competitor prices, lowers its price five times consecutively within 10 minutes, with a cumulative price reduction of 25%. The arbitration agent calculates the price fluctuation entropy, finds that the entropy value exceeds a safe threshold, and immediately issues a cooling-off command to product agent B. Product agent B is locked in read-only mode for the next 30 minutes, and the price decline stops. After manual intervention to investigate and repair, the cooling-off period is lifted. The mall's pricing system does not become chaotic due to the failure of a single agent.

[0083] It should be noted that by continuously monitoring the pricing history of each product's intelligent agent and calculating the price fluctuation entropy through the arbitration intelligent agent, real-time statistical monitoring of cumulative abnormal behavior such as frequent and repeated price adjustments is achieved. When the number of price adjustments by any product's intelligent agent within a preset time period exceeds a preset threshold and the average price adjustment exceeds a preset threshold, a cooling-off command is issued to that product's intelligent agent through the arbitration intelligent agent. This achieves more accurate identification of abnormal behavior and effectively avoids false alarms caused by judgment based on a single threshold. By issuing cooling-off commands through the arbitration intelligent agent and controlling the product's intelligent agent to only perform the action of maintaining the current price within a preset cooling-off period, the overall stability of the mall's pricing system is effectively maintained.

[0084] Optionally, the method also includes: Each product agent maintains a Q table; After each decision is executed, the Q table is updated immediately based on the sum of the actual incremental profit from the transaction and the virtual collaborative points penalty paid. The product agent adopts an ε-greedy strategy to randomly explore the action space with a preset probability.

[0085] In this embodiment, the Q table can be a two-dimensional or multi-dimensional tabular data structure maintained internally by each product agent. The Q table is used to store the Q value under different state-action combinations, i.e., the expected cumulative reward.

[0086] The state space of the aforementioned Q table includes inventory level tiers, market popularity scalar tiers, and virtual collaborative points balance tiers. The action space includes multiple preset price adjustment ranges. The inventory level tiers can be divided into multiple discrete tiers based on the current inventory days of the product, such as high inventory, medium inventory, and low inventory. The market popularity scalar tiers can be divided into multiple discrete tiers based on the currently calculated market popularity scalar value range, such as increasing popularity, stable popularity, and decreasing popularity. The virtual collaborative points balance tiers can be divided into multiple discrete tiers based on the current virtual collaborative points balance held by the product agent, such as sufficient balance, normal balance, and insufficient balance. The preset price adjustment ranges can be pre-set price adjustment ranges, such as a 3% price reduction, a 5% price reduction, maintaining the current price, or a 1% price increase. Each action corresponds to a price adjustment coefficient value.

[0087] The aforementioned actual increase in profit from transactions can be the actual new sales profit generated.

[0088] The virtual collaborative points penalty for the aforementioned payment can be the number of points paid out.

[0089] The aforementioned immediate reward can be an immediate feedback value for the product agent after each decision is made, based on the sum of the actual incremental profit from sales and the virtual collaborative points penalty paid. This immediate reward reflects both the positive effect of the pricing action on increased sales profit and the negative effect of the points penalty incurred due to damage to the profits of related product categories. This comprehensive feedback signal, encompassing both positive and negative aspects, allows the agent to balance its own profits with the overall health of the ecosystem during learning.

[0090] The aforementioned ε-greedy strategy can be considered a balance between exploration and exploitation strategy employed by a commodity agent when using a Q-table for action selection. This ε-greedy strategy is used in reinforcement learning to explore randomly with probability ε and exploit randomly with probability 1. The core action selection mechanism of ε-chooses the current optimal action to balance exploration and exploitation. Under the ε-greedy policy, the agent selects the action with the largest Q value in the current Q-table with a probability of 1-ε. That is, it uses existing optimal experience to randomly select an action in the action space with a probability of ε, which is to explore new unknown strategies.

[0091] The aforementioned preset probabilities are preset exploration probability values, such as 0.1. By employing an ε-greedy strategy to randomly explore the action space with preset probabilities, a dynamic balance mechanism is established between utilizing existing experience and exploring unknown strategies. This allows the agent to gradually converge to the optimal strategy while continuously discovering better price adjustment ranges outside the coverage of the decision table.

[0092] Understandably, if the agent always chooses the action with the highest current Q-value, it will adhere to initial experience or early-formed strategy paths, failing to discover potentially better but unverified price adjustment ranges. For example, an 8% price reduction not included in the decision table might be highly effective in certain states. Conversely, if the agent always chooses actions randomly, it cannot accumulate stable experience, and the strategy will fail to converge. The ε-greedy strategy, by utilizing learned experience to ensure a lower bound on decision quality while simultaneously exploring randomly with a small probability, continuously supplements the Q-table with new state-action reward information. This allows the agent's strategy to slowly and continuously evolve towards a better direction over long-term operation, effectively mitigating the risk of significant fluctuations in decision quality due to overexploration.

[0093] Optionally, the method also includes: During the offline phase, the decision-making trajectories of each product agent within a preset number of days are replayed, and the regret value at each decision moment is calculated. Decision trajectories with regret values ​​exceeding a preset threshold are used as negative samples and are forcibly input into the Q-table of the corresponding product agent for reverse gradient updates.

[0094] In this embodiment, the aforementioned offline phase can be a time window during off-peak operating periods when the e-commerce system pauses or reduces the pressure of real-time pricing decisions, allowing for the replay analysis of historical data and optimization of model parameters. The offline phase and the online phase are separate in time and do not affect each other.

[0095] The preset number of days can be a pre-set number of days, such as the past 3 days, the past 5 days, the past 7 days, etc.

[0096] The aforementioned decision trajectory can be a complete sequence record of the state experienced by each product agent at each decision moment, the action selected, the immediate reward obtained, and the transition to the next state during a preset number of days of online operation. The decision trajectory can be stored in the system log in the form of a four-tuple sequence, with each record including complete information on the current state → action performed → reward obtained → next state.

[0097] The optimal action mentioned above could be the action with the largest Q value from the next state row of the Q-table of the commodity agent when replaying each decision record, for the next state in that record, under the state snapshot of the Q-table at that time.

[0098] The regret value mentioned above is the difference in Q-value between the currently chosen action and the optimal action afterwards.

[0099] The aforementioned preset threshold can be a pre-defined threshold. This threshold is used to filter decision trajectories with high regret. Only when the regret value of a decision record exceeds the preset threshold is the record considered a negative sample with significant learning value and forced to be input into the Q-table for backward gradient updates.

[0100] The aforementioned negative samples can be decision trajectory records that are identified as having a regret value exceeding a preset threshold during offline playback.

[0101] The aforementioned backward gradient update can be an operation that forcibly updates the Q-value corresponding to the actual action selected in a decision record identified as a negative sample, in a direction that reduces the regret value. Specifically, for the Q-value of the actual action selected in a negative sample, an additional update step is used to move the Q-value of the actual action selected towards the Q-value of the subsequent optimal action, thereby reducing the gap between the two.

[0102] It should be noted that by replaying the decision trajectories of each product agent within a preset number of days during the offline phase and calculating the regret value at each decision moment, the failure experience contained in historical erroneous decisions is upgraded from single use to batch reuse, significantly improving the learning and utilization rate of historical operational data. By using decision trajectories with regret values ​​exceeding a preset threshold as negative samples and using the difference between the actual selected action and the Q value of the subsequent optimal action as the quantitative screening criterion, accurate identification and efficient screening of negative samples are achieved, avoiding the computational burden caused by indiscriminately updating all historical records. By forcibly inputting the screened negative samples into the Q-table of the corresponding product agent for reverse gradient update, the Q value of the wrong action selected at that time in the Q-table moves towards the Q value of the subsequent optimal action, realizing the systematic and centralized correction of historical erroneous experience and accelerating the convergence process of the Q-table from suboptimal strategy to optimal strategy.

[0103] like Figure 2 As shown, this embodiment of the invention provides a multi-agent-based e-commerce management device, which includes: The generation module 201 is used to deploy product intelligent agents for products in the mall, and through the product intelligent agents, perceive the internal inventory status and the external market environment, generate autonomous price adjustment decisions, and execute the autonomous price adjustment decisions. The rollback module 202 is used to pay virtual collaborative points to the product intelligent agent of the related product category as compensation when the pricing or promotional actions of the product intelligent agent affect the profit of the related product category. If the virtual collaborative points are insufficient, the pricing or promotional actions are rolled back. The processing module 203 is used to rank the products based on the balance of virtual collaborative points held by each product agent, and to give weight to the top-ranked product agents in the subsequent traffic allocation. Each product agent participates in the competition for traffic positions by using the virtual collaborative points it holds as bidding chips, and obtains traffic positions based on the bidding results. The sorting and display module 204 is used to sort and display products in the mall according to traffic position.

[0104] Optionally, the generation module 201 is further configured to acquire internal state data and external market environment data through the product agent. The internal state data includes the current inventory depth, historical conversion rate, and refund rate of the product. The external market environment data includes competitor prices and social media sentiment information. Based on the internal state data and the external market environment data, a market popularity scalar is calculated. Based on the market popularity scalar and the current inventory days, a built-in dynamic decision table is queried to determine the price adjustment coefficient. The current product price is adjusted according to the price adjustment coefficient.

[0105] Optionally, the generation module 201 is further configured to take the logarithm of the site search index, multiply it by the sum of the social sentiment score and 1, and then subtract the product of the competitor price difference rate and the preset coefficient; wherein, the social sentiment score is obtained by scoring the social media keywords with sentiment through an open-source sentiment analysis model and mapping it to a normalized interval, and the market popularity scalar is greater than 0 to indicate that the market popularity is rising, and less than 0 to indicate that the market popularity is falling.

[0106] Optionally, the rollback module 202 is further configured to construct a category network graph, which is constructed according to the purchasing behavior correlation between each product, and the edge weight represents the substitution strength or complementarity strength; when the product agent performs a pricing or promotional discount action, the profit impact of the action on the related categories in the category network graph is calculated; if the action causes profit damage to the related categories, the product agent pays virtual collaborative points to the product agent corresponding to the affected related categories as compensation, and the compensation amount is the product of the absolute value of the profit loss and the preset profit and loss exchange coefficient; if the virtual collaborative points held by the product agent are insufficient to pay the compensation, the pricing or promotional discount action is rolled back.

[0107] Optionally, the device is further configured to monitor the pricing history of each commodity intelligence agent through an arbitration agent and calculate the price fluctuation entropy; when any commodity intelligence agent adjusts its price more than a preset number of times within a preset time period and the average price adjustment exceeds a preset magnitude threshold, the arbitration agent issues a cooling command to the commodity intelligence agent; and controls the commodity intelligence agent to perform only the action of maintaining the current price within a preset cooling period according to the cooling command.

[0108] Optionally, the device is also used to maintain a Q-table for each product agent. The state space of the Q-table includes inventory level tiers, market popularity scalar tiers, and virtual collaboration points balance tiers. The action space includes multiple preset price adjustment ranges. After each decision is executed, the Q-table is updated with the sum of the actual incremental profit from transactions and the virtual collaboration points penalty paid as an immediate reward. The product agent adopts an ε-greedy strategy to randomly explore the action space with a preset probability.

[0109] Optionally, the device is also used to replay the decision trajectories of each product agent within a preset number of days during the offline phase, calculate the regret value at each decision moment, wherein the regret value is the Q-value difference between the currently selected action and the optimal action afterwards; and take the decision trajectory with the regret value exceeding a preset threshold as a negative sample, and force it to be input into the Q-table of the corresponding product agent for reverse gradient update.

[0110] It should be noted that the multi-agent-based e-commerce management device provided in this embodiment of the invention can be applied to computers, servers and other devices that can perform multi-agent-based e-commerce management methods.

[0111] The multi-agent-based e-commerce management device provided in this embodiment of the invention can implement all the processes of the multi-agent-based e-commerce management method in the above-described method embodiments, and can achieve the same beneficial effects. To avoid repetition, further details are omitted here.

[0112] See Figure 3 , Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention, such as... Figure 3 As shown, it includes: a memory 302, a processor 301, and a computer program for a multi-agent-based e-commerce management method stored in the memory 302 and executable on the processor 301, wherein: The processor 301 is used to call the computer program stored in the memory 302 and perform the following steps: Deploy product intelligence agents for products in the mall, and use these product intelligence agents to perceive internal inventory status and external market environment, generate autonomous price adjustment decisions, and execute these autonomous price adjustment decisions. When the pricing or promotional actions of the product agent affect the profit of related product categories, the product agent pays virtual collaboration points to the product agents of related product categories as compensation. If the virtual collaboration points are insufficient, the pricing or promotional actions are rolled back. The ranking is based on the virtual collaboration points held by each product agent. The top-ranked product agents are given weight in the subsequent traffic allocation. Each product agent uses its virtual collaboration points as a bidding chip to participate in the competition for traffic positions and obtains traffic positions based on the bidding results. Products in the online store are sorted and displayed based on traffic volume.

[0113] Optionally, the process executed by processor 301, which involves perceiving the internal inventory status and external market environment through the product intelligence agent, generating an autonomous price adjustment decision, and executing the autonomous price adjustment decision, specifically includes: The product intelligent agent acquires internal status data and external market environment data. The internal status data includes the current inventory depth, historical conversion rate, and refund rate of the product. The external market environment data includes competitor prices and social media sentiment information. The market heat index is calculated based on the internal status data and the external market environment data. The price adjustment coefficient is determined by querying the built-in dynamic decision table based on the market heat index and the current inventory days. The current product price will be adjusted according to the aforementioned price adjustment factor.

[0114] Optionally, the processor 301 performs the calculation of a market heat scalar based on the internal state data and the external market environment data, specifically including: Multiply the logarithm of the site search index by the sum of the social sentiment score and 1, and then subtract the product of the competitor's price difference rate and the preset coefficient. The social sentiment score is obtained by scoring social media keywords with sentiment using an open-source sentiment analysis model and mapping it to a normalized interval. The market popularity scalar value is greater than 0, indicating that the market popularity is rising, and less than 0, indicating that the market popularity is falling.

[0115] Optionally, the processor 301 executes the following: when the pricing or promotional actions of the product agent affect the profit of related product categories, the product agent pays virtual collaboration points to the product agents of related product categories as compensation; if the virtual collaboration points are insufficient, the pricing or promotional actions are rolled back. Specifically, this includes: Construct a category network graph, which is built according to the purchasing behavior correlation between each product, and the edge weight represents the substitution strength or complement strength; When the product agent performs a pricing or promotional discount action, the profit impact of the action on the associated categories in the category network graph is calculated. If the action causes profit damage to the related product category, the product agent pays virtual collaborative points to the product agent corresponding to the affected related product category as compensation. The amount of compensation is the product of the absolute value of the profit loss and the preset profit and loss exchange coefficient. If the virtual collaboration points held by the product agent are insufficient to pay the compensation, the pricing or promotional discount action will be rolled back.

[0116] Optionally, processor 301 may also perform the following: By listening to the pricing history of each commodity agent through an arbitration agent, the price fluctuation entropy is calculated. When any product agent adjusts its price more than a preset number of times within a preset time period and the average price adjustment exceeds a preset range threshold, a cooling command is issued to the product agent through the arbitration agent. The control agent for the product will only perform the action of maintaining the current price within a preset cooling time according to the cooling command.

[0117] Optionally, the method executed by processor 301 further includes: Each product agent maintains a Q table, whose state space includes inventory level tiers, market popularity scalar tiers, and virtual collaborative points balance tiers, and whose action space includes multiple preset price adjustment ranges; After each decision is executed, the Q table is updated immediately as the sum of the actual incremental profit from the transaction and the virtual collaborative points penalty paid. The product agent adopts an ε-greedy strategy to randomly explore the action space with a preset probability.

[0118] Optionally, the method executed by processor 301 further includes: During the offline phase, the decision-making trajectories of each product agent within a preset number of days are replayed, and the regret value at each decision moment is calculated. The regret value is the Q-value difference between the currently selected action and the optimal action afterwards. Decision trajectories with regret values ​​exceeding a preset threshold are used as negative samples and are forcibly input into the Q-table of the corresponding product agent for reverse gradient updates.

[0119] It should be noted that the electronic device provided in the embodiments of the present invention can be applied to computers, servers and other devices that can perform e-commerce management methods based on multi-agent systems.

[0120] The electronic device provided in this embodiment of the invention can implement all the processes of the multi-agent-based e-commerce management method described in the above-described method embodiments, and can achieve the same beneficial effects. To avoid repetition, further details are omitted here.

[0121] This invention also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the various processes of the multi-agent-based e-commerce management method provided in this invention and achieves the same technical effect. To avoid repetition, it will not be described again here.

[0122] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The computer-readable storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0123] The above description discloses only preferred embodiments of the present invention and should not be construed as limiting the scope of the present invention. Therefore, equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A multi-agent-based e-commerce management method, characterized in that, The method includes the following steps: Deploy product intelligence agents for products in the mall, and use these product intelligence agents to perceive internal inventory status and external market environment, generate autonomous price adjustment decisions, and execute these autonomous price adjustment decisions. When the pricing or promotional actions of the product agent affect the profit of related product categories, the product agent pays virtual collaboration points to the product agents of related product categories as compensation. If the virtual collaboration points are insufficient, the pricing or promotional actions are rolled back. The ranking is based on the virtual collaboration points held by each product agent. The top-ranked product agents are given weight in the subsequent traffic allocation. Each product agent uses its virtual collaboration points as a bidding chip to participate in the competition for traffic positions and obtains traffic positions based on the bidding results. Products in the online store are sorted and displayed based on traffic volume.

2. The e-commerce management method based on multi-agent systems as described in claim 1, characterized in that, The process of sensing internal inventory status and external market environment through the product intelligence agent, generating autonomous price adjustment decisions, and executing these decisions specifically includes: The product intelligent agent acquires internal status data and external market environment data. The internal status data includes the current inventory depth, historical conversion rate, and refund rate of the product. The external market environment data includes competitor prices and social media sentiment information. The market heat index is calculated based on the internal status data and the external market environment data. The price adjustment coefficient is determined by querying the built-in dynamic decision table based on the market heat index and the current inventory days. The current product price will be adjusted according to the aforementioned price adjustment factor.

3. The e-commerce management method based on multi-agent technology as described in claim 2, characterized in that, The market heat scalar is calculated based on the internal status data and the external market environment data, specifically including: Multiply the logarithm of the site search index by the sum of the social sentiment score and 1, and then subtract the product of the competitor's price difference rate and the preset coefficient. The social sentiment score is obtained by scoring social media keywords with sentiment using an open-source sentiment analysis model and mapping it to a normalized interval. The market popularity scalar value is greater than 0, indicating that the market popularity is rising, and less than 0, indicating that the market popularity is falling.

4. The multi-agent-based e-commerce management method as described in any one of claims 1 to 3, characterized in that, When the pricing or promotional actions of the product agent affect the profits of related product categories, the product agent pays virtual collaboration points to the product agents of the related product categories as compensation. If the virtual collaboration points are insufficient, the pricing or promotional actions are rolled back. Specifically, this includes: Construct a category network graph, which is built according to the purchasing behavior correlation between each product, and the edge weight represents the substitution strength or complement strength; When the product agent performs a pricing or promotional discount action, the profit impact of the action on the associated categories in the category network graph is calculated. If the action causes profit damage to the related product category, the product agent pays virtual collaborative points to the product agent corresponding to the affected related product category as compensation. The amount of compensation is the product of the absolute value of the profit loss and the preset profit and loss exchange coefficient. If the virtual collaboration points held by the product agent are insufficient to pay the compensation, the pricing or promotional discount action will be rolled back.

5. The multi-agent-based e-commerce management method as described in any one of claims 1 to 3, characterized in that, Also includes: By listening to the pricing history of each commodity agent through an arbitration agent, the price fluctuation entropy is calculated. When any product agent adjusts its price more than a preset number of times within a preset time period and the average price adjustment exceeds a preset range threshold, a cooling command is issued to the product agent through the arbitration agent. The control agent for the product will only perform the action of maintaining the current price within a preset cooling time according to the cooling command.

6. The multi-agent-based e-commerce management method as described in any one of claims 1 to 3, characterized in that, The method further includes: Each product agent maintains a Q table, whose state space includes inventory level tiers, market popularity scalar tiers, and virtual collaborative points balance tiers, and whose action space includes multiple preset price adjustment ranges; After each decision is executed, the Q table is updated with the sum of the actual incremental profit from the transaction and the virtual collaborative points penalty paid as an immediate reward. The product agent adopts an ε-greedy strategy to randomly explore the action space with a preset probability.

7. The multi-agent-based e-commerce management method as described in claim 6, characterized in that, The method further includes: During the offline phase, the decision-making trajectories of each product agent within a preset number of days are replayed, and the regret value at each decision moment is calculated. The regret value is the Q-value difference between the currently selected action and the optimal action afterwards. Decision trajectories with regret values ​​exceeding a preset threshold are used as negative samples and are forcibly input into the Q-table of the corresponding product agent for reverse gradient updates.

8. A multi-agent-based e-commerce management device, characterized in that, The device includes: The generation module is used to deploy product intelligence agents for products in the mall, and through the product intelligence agents, perceive the internal inventory status and external market environment, generate autonomous price adjustment decisions, and execute the autonomous price adjustment decisions. The rollback module is used to compensate the product agent for its pricing or promotional actions when the pricing or promotional actions of the product agent affect the profits of related product categories. If the virtual collaborative points are insufficient, the pricing or promotional actions are rolled back. The processing module is used to rank the product agents based on the balance of virtual collaborative points held by each product agent, and to give weight to the top-ranked product agents in subsequent traffic allocation. Each product agent participates in the competition for traffic positions by using the virtual collaborative points it holds as bidding chips, and obtains traffic positions based on the bidding results. The sorting and display module is used to sort and display products in the mall based on traffic volume.

9. An electronic device, characterized in that, include: The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the multi-agent-based e-commerce management method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the multi-agent-based e-commerce management method as described in any one of claims 1 to 7.

Images