A finance and tax feature driven supply chain collaborative pricing optimization method and system

Abstract

The application discloses a kind of finance and tax characteristics driven supply chain collaborative pricing optimization method, it is related to finance and tax characteristics driven technical field, including the following steps: step one, based on real supply chain transaction graph and historical invoice flow data, construct supply chain training environment and the state space of deep reinforcement learning;Step two, based on the historical invoice flow data, using inverse reinforcement learning or parameter estimation method, automatically learn the real response function of supplier under specific tax constraints;Step three, construct collaborative pricing model that fuses Stackelberg game and deep reinforcement learning, take core enterprise as leader, and supplier as follower;Step four, using deep reinforcement learning algorithm, the pricing strategy of the core enterprise is iterated optimization;Step five, through multi-round interaction deduction upstream and downstream interest balance point, real-time output optimal settlement strategy that takes into account tax compliance and supply chain collaboration. The intelligent level of supply chain management is improved.

Description

Technical Field

[0001] This invention relates to the field of supply chain collaborative management and operations optimization technology, specifically to a method and system for optimizing supply chain upstream and downstream collaborative pricing and settlement strategies by integrating game theory and deep reinforcement learning under the constraints of financial and tax compliance. Background Technology

[0002] With the increasing sophistication of supply chain management, collaborative pricing and settlement strategies among upstream and downstream enterprises have become crucial for enhancing the overall competitiveness of the supply chain. Existing supply chain pricing models are mostly based on classic economic models (Stackelberg game models) or traditional operations research algorithms. These methods, when constructing the state space and reward function, typically only consider basic physical indicators such as supply and demand relationships, inventory costs, and logistics costs, neglecting the complex financial and tax characteristics and capital flow efficiency between enterprises.

[0003] The existing technology has the following main shortcomings: lack of quantitative drivers for fiscal and tax characteristics: traditional models do not include fiscal and tax indicators such as tax burden costs, capital occupancy rate and performance risk into core decision variables, which may lead to tax compliance risks or capital chain breaks in the actual implementation of the generated pricing strategy.

[0004] Lack of feedback mechanism based on real transaction data: Existing models mostly rely on static assumptions or historical averages, making it difficult to dynamically capture the supplier's real reaction function in a complex financial and tax environment.

[0005] The separation of tax and financial constraints from pricing decisions: Traditional methods treat pricing and tax compliance as two separate stages, which fails to achieve joint optimization of "price-payment period-tax burden" and makes it difficult to deduce the true equilibrium point of interests between upstream and downstream companies under tax and financial constraints.

[0006] Therefore, a supply chain collaborative pricing optimization method and system driven by fiscal and tax characteristics is very necessary. Summary of the Invention

[0007] To address the shortcomings of existing methods, this paper presents a supply chain collaborative pricing optimization method and system driven by financial and tax characteristics. It internalizes financial and tax compliance requirements as optimization variables and learns supplier reaction functions through feedback from real invoice flows, thereby achieving joint dynamic optimization of "price-payment period-tax burden-risk".

[0008] To solve the above problems, the technical solution adopted by the present invention is as follows:

[0009] A supply chain collaborative pricing optimization method driven by fiscal and tax characteristics includes the following steps:

[0010] Step 1: Based on real supply chain transaction graphs and historical invoice flow data, construct the state space of the supply chain training environment and deep reinforcement learning. The state space contains multi-dimensional financial and tax features.

[0011] Step 2: Based on the historical invoice flow data, use inverse reinforcement learning or parameter estimation methods to automatically learn the supplier's true response function under specific financial and tax constraints;

[0012] Step 3: Construct a collaborative pricing model that integrates Stackelberg game theory and deep reinforcement learning, with the core enterprise as the leader and the suppliers as followers, and use a reward function with embedded fiscal and tax constraints to optimize the pricing strategy of the core enterprise. The reward function includes economic benefit items, fiscal and tax penalty items, and performance risk items.

[0013] Step 4: Use deep reinforcement learning algorithms to iteratively optimize the pricing strategy of the core enterprise;

[0014] Step 5: Through multiple rounds of interactive deduction, the equilibrium point of upstream and downstream interests is deduced, and the optimal settlement strategy that balances financial and tax compliance with supply chain collaboration is output in real time.

[0015] A supply chain collaborative pricing optimization system driven by fiscal and tax characteristics, comprising:

[0016] The data access and feature engineering module is used to connect with supply chain transaction graphs and invoice flow data, extract and structure multi-dimensional financial and tax features, and construct a state space for deep reinforcement learning.

[0017] The reaction function learning module is used to automatically learn the supplier's true reaction function under specific financial and tax constraints based on the invoice flow data, using inverse reinforcement learning or parameter estimation methods.

[0018] The collaborative pricing decision engine is used to build a collaborative pricing model that integrates Stackelberg game theory and deep reinforcement learning. It treats core enterprises as leaders and suppliers as followers, and generates strategies, calculates rewards, and deduces equilibrium points based on reward functions with embedded fiscal and tax constraints.

[0019] The strategy execution and feedback module is used to distribute the optimal settlement strategy to the business system and collect real financial and tax data after execution for closed-loop iteration. Detailed Implementation

[0020] The more detailed description of embodiments of the invention below is not intended to limit the scope of the claimed invention, but is merely illustrative and does not limit the description of the features and characteristics of the invention, in order to suggest the best mode for carrying out the invention and to enable those skilled in the art to practice the invention. However, it should be understood that various modifications and variations can be made without departing from the scope of the invention as defined by the appended claims. The detailed description should be considered illustrative only and not restrictive, and any such modifications and variations shall fall within the scope of the invention described herein. Furthermore, the background art is intended to illustrate the current state of research and development and significance of the technology, and is not intended to limit the invention or the scope of application of this application.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention.

[0022] The present invention will be further described below with reference to specific embodiments.

[0023] Example 1

[0024] This embodiment provides a supply chain collaborative pricing optimization method driven by fiscal and tax characteristics, including the following steps:

[0025] Step 1: Based on real supply chain transaction graphs and historical invoice flow data, construct the state space of the supply chain training environment and deep reinforcement learning. The state space contains multi-dimensional financial and tax features.

[0026] The system first connects with the enterprise's internal supply chain transaction graph and historical invoice flow data. It extracts key fields such as invoice date, amount, tax rate, and payment period from the invoice flow data, and calculates the supplier's tax burden rate, capital occupation period, and historical performance risk index to form multi-dimensional financial and tax features. These multi-dimensional financial and tax features are then integrated with basic supply and demand parameters (such as demand and inventory levels) to construct a joint state vector, which serves as the state space for deep reinforcement learning. The training environment, built upon a real supply chain transaction graph, can simulate the interactive behavior of upstream and downstream enterprises under different financial and tax constraints, providing a realistic simulation foundation for subsequent strategy optimization.

[0027] Step 2: Based on the historical invoice flow data, use inverse reinforcement learning or parameter estimation methods to automatically learn the supplier's true response function under specific financial and tax constraints.

[0028] By leveraging real transaction records contained in historical invoice stream data, the system employs inverse reinforcement learning or parameter estimation methods to automatically learn the supplier's true response function under specific financial and tax constraints. Unlike the static response function in traditional game theory based on the assumption of perfect rationality, this invention uses inverse reinforcement learning to deduce the supplier's implicit reward function and decision preferences from their historical behavior, thereby fitting a dynamic response function that more closely reflects the real business environment. The parameter estimation method estimates the probability distribution of the supplier's response to different combinations of prices and payment terms by statistically modeling the joint distribution of prices and payment terms in historical invoice data.

[0029] Step 3: Construct a collaborative pricing model that integrates Stackelberg game theory and deep reinforcement learning, with the core enterprise as the leader and the suppliers as followers. Optimize the core enterprise's pricing strategy using a reward function embedded with fiscal and tax constraints. The reward function includes economic benefit terms, fiscal and tax penalty terms, and performance risk terms.

[0030] In constructing the collaborative pricing model, this invention positions the core enterprise as the leader in a Stackelberg game, and suppliers as followers. The core enterprise first proposes its purchase price and payment terms strategy, and suppliers respond optimally based on their true reaction functions. Simultaneously, this invention introduces a deep reinforcement learning algorithm to iteratively optimize the core enterprise's strategy. In designing the reward function, this invention innovatively embeds fiscal and tax constraints: an economic benefit term measures the core enterprise's procurement cost savings and the time value of money; a fiscal and tax penalty term inflicts negative penalties when the strategy leads to an abnormal increase in the overall tax burden or exceeds the capital utilization limit; and a performance risk term discounts high-risk settlement strategies based on a performance risk index. Through the weighted combination of these three terms, the reward function guides the model to generate an equilibrium strategy that balances economic benefits, fiscal and tax compliance, and risk control.

[0031] Step 4: Use deep reinforcement learning algorithms to iteratively optimize the pricing strategy of the core enterprise.

[0032] During training, deep reinforcement learning algorithms take a joint state vector as input and the core enterprise's pricing strategy (a combination of purchase price and payment terms) as output. Through continuous interaction with the training environment, they constantly collect feedback rewards and update the policy network parameters. In each iteration, the algorithm generates candidate actions based on the current policy, observes the supplier's response and the reward signal from the environment, and updates the policy using policy gradient methods or the Actor-Critic framework to maximize long-term cumulative rewards.

[0033] Step 5: Through multiple rounds of interactive deduction, the equilibrium point of upstream and downstream interests is deduced, and the optimal settlement strategy that balances financial and tax compliance with supply chain collaboration is output in real time.

[0034] The system uses multiple rounds of interactive deduction to determine the equilibrium point of interests between upstream and downstream entities. In each round, the core enterprise outputs a combination of purchase price and payment terms based on the current strategy network, and the supplier makes the optimal response based on the true reaction function. The system calculates rewards and updates its strategy based on the interaction results between the two parties. After multiple iterations, the strategy network converges to a strategy near Nash equilibrium or Stackelberg equilibrium, and finally outputs the optimal settlement strategy in real time, namely the optimal combination of purchase price and payment terms.

[0035] The effects of the above technical solution are as follows: By embedding multi-dimensional financial and tax features into the state space, the pricing strategy fully considers tax costs, capital occupation cycles, and performance risks during the generation stage, fundamentally avoiding compliance risks caused by traditional models neglecting financial and tax constraints; By automatically learning supplier reaction functions from real invoice flow data through inverse reinforcement learning, the model can accurately capture the real decision-making behavior of suppliers in complex financial and tax environments, significantly improving the feasibility and predictive accuracy of the strategy in actual business; By integrating Stackelberg game theory and deep reinforcement learning, joint dynamic optimization of price and payment terms is achieved, breaking the limitations of separate decision-making between price and payment terms in traditional methods, effectively improving the overall capital turnover efficiency of the supply chain; By introducing financial and tax penalties and performance risk items, the model can automatically avoid high-risk and non-compliant settlement strategies while pursuing economic benefits, enhancing the risk resistance of the supply chain.

[0036] Example 2

[0037] This embodiment provides a supply chain collaborative pricing optimization system driven by fiscal and tax characteristics, including:

[0038] The data access and feature engineering module is used to connect with supply chain transaction graphs and invoice flow data, extract and structure multi-dimensional financial and tax features, and construct a state space for deep reinforcement learning.

[0039] The reaction function learning module is used to automatically learn the supplier's true reaction function under specific financial and tax constraints based on the invoice flow data, using inverse reinforcement learning or parameter estimation methods.

[0040] The collaborative pricing decision engine is used to build a collaborative pricing model that integrates Stackelberg game theory and deep reinforcement learning. It treats core enterprises as leaders and suppliers as followers, and generates strategies, calculates rewards, and deduces equilibrium points based on reward functions with embedded fiscal and tax constraints.

[0041] The strategy execution and feedback module is used to distribute the optimal settlement strategy to the business system and collect real financial and tax data after execution for closed-loop iteration.

[0042] The data access and feature engineering module connects with the enterprise's internal ERP system, tax management system, and e-invoice platform through standardized data interfaces to acquire supply chain transaction data and invoice flow data in real time or near real time. This module incorporates feature engineering components to clean, transform, and extract features from raw data, calculating multi-dimensional financial and tax characteristics such as supplier tax burden rate, capital occupation cycle, and historical performance risk index. These multi-dimensional financial and tax characteristics are then fused with basic supply and demand parameters to form a joint state vector, constructing and maintaining the state space for deep reinforcement learning. The reaction function learning module incorporates an inverse reinforcement learning algorithm library and parameter estimation toolkit. It acquires historical supplier behavior data from the data access module and continuously learns and updates the supplier's true reaction function using maximum entropy inverse reinforcement learning or Bayesian parameter estimation methods. The collaborative pricing decision engine is the core decision-making component of the system. It incorporates a Stackelberg game model and deep reinforcement learning algorithms, responsible for constructing the collaborative pricing model, designing reward functions embedded with financial and tax constraints, executing strategy generation and iterative optimization, and outputting the optimal settlement strategy in real time through multi-round interactive deduction of the equilibrium point of upstream and downstream interests. The strategy execution and feedback module connects with the enterprise's ERP system, procurement management system, and financial settlement system through the API gateway. It is responsible for distributing the optimal settlement strategy output by the collaborative pricing decision engine to the business system for execution, and continuously collecting real financial and tax data after execution. The data is then fed back to the data access and feature engineering module and the reaction function learning module to achieve closed-loop iterative optimization of the system.

[0043] The effects of the above technical solution are as follows: Through the coordinated operation of the four functional modules, the system realizes a complete closed loop from data access, feature extraction, reaction function learning, strategy generation to strategy execution and feedback, and builds a self-learning and self-evolving supply chain collaborative pricing optimization capability; through the standardized interface design of the data access and feature engineering modules, the system can flexibly connect to various enterprise information systems and data sources, and has good scalability and compatibility; through the game-reinforcement learning fusion architecture of the collaborative pricing decision engine, the system can output the optimal settlement strategy that takes into account both financial and tax compliance and supply chain collaboration while ensuring the core enterprise's profits, significantly improving the intelligence level of supply chain management and the efficiency of capital operation.

[0044] Example 3

[0045] This embodiment takes a collaborative pricing scenario between a core manufacturing enterprise and its upstream steel supplier as an example to illustrate the specific implementation of the present invention in actual business, including the numerical construction of the state space, the formulaic definition of the reward function, the network structure and hyperparameter settings of the PPO algorithm, and the complete multi-round interactive deduction process.

[0046] Step 1: Numerical Construction of the State Space

[0047] Let the state space of the core enterprise during a certain procurement cycle be a joint state vector s, with a dimension of 7, as defined below:

[0048] s = [D, I, T, C, F, R, P]

[0049] Where D represents the steel demand of core enterprises (unit: tons), I represents the current steel inventory level (unit: tons), T represents the supplier's tax burden cost rate (dimensionless, value range [0, 1]), C represents the capital occupation period (unit: days), F represents the historical performance risk index (dimensionless, value range [0, 1]), R represents the current market benchmark interest rate (dimensionless, value range [0, 1]), and P represents the actual purchase price of the previous period (unit: yuan / ton).

[0050] In this embodiment, a typical state sample is: s = [500, 120, 0.13, 45, 0.08, 0.035, 3800], that is, the demand is 500 tons, the inventory is 120 tons, the supplier's tax burden rate is 13%, the capital occupation period is 45 days, the historical performance risk index is 0.08, the market benchmark interest rate is 3.5%, and the purchase price in the previous period was 3800 yuan / ton.

[0051] The system connects to the core enterprise's ERP system and electronic invoice platform through a standardized data interface, extracts the above fields in real time, and constructs a continuous state sequence using a sliding time window (window length set to 90 days, step size of 1 day) as input for the PPO algorithm.

[0052] Step Two: Learning the Supplier's True Reaction Function

[0053] This embodiment employs the Maximum Entropy Inverse Reinforcement Learning (MaxEnt IRL) method to learn the true response function of steel suppliers from historical invoice stream data over the past 36 months.

[0054] Specifically, suppose the supplier's behavioral trajectory τ consists of a series of state-action pairs: The supplier's action 'a' is defined as a response decision to the core enterprise's proposed combination of purchase price and payment terms, including accepting, rejecting, or submitting a counter-offer.

[0055] MaxEnt IRL estimates the supplier's reward function R_supplier(s, a) by solving the following optimization problem:

[0056] maximize Σ_τ P(τ|R) · log P(τ|R)

[0057] Where P(τ|R) = (1 / Z(R)) · exp(Σ_{(s,a)∈τ} R_supplier(s, a)), and Z(R) is the partition function.

[0058] In this embodiment, the supplier's reward function is assumed to be in linear parameterized form: , where φ(s, a) is the feature vector, containing the following four items: supplier gross profit margin, cash collection speed, tax compliance, and order fulfillment rate. The weight vector w is iteratively optimized using gradient descent, with a learning rate set to 0.001 and an iteration count of 5000.

[0059] After learning, the system obtains the supplier's true response function π_supplier(a|s), which is the conditional probability distribution of the supplier taking action a given state s.

[0060] Step 3: Constructing a reward function with embedded fiscal and tax constraints

[0061] In this embodiment, the reward function R_total for the core enterprise is designed as a weighted sum of three terms:

[0062] R_total = α·R_economic + β·R_tax + γ·R_risk

[0063] Wherein, α, β, and γ are weighting coefficients, satisfying α + β + γ = 1. In this embodiment, α = 0.5, β = 0.3, and γ = 0.2 are set.

[0064] The economic benefit term R_economic is defined as:

[0065] R_economic = (P_base - P_t) · Q_t + (P_t · Q_t · r · (T_base - C_t)) / 365

[0066] Where P_base is the benchmark purchase price (set to 4000 yuan / ton in this embodiment), P_t is the actual purchase price in the t-th iteration, Q_t is the purchase quantity, r is the market benchmark interest rate, T_base is the benchmark payment period (set to 60 days in this embodiment), and C_t is the actual payment period in the t-th iteration (unit: days). The first term represents the savings in procurement costs, and the second term represents the time value of money saved due to the shortened payment period.

[0067] The tax penalty item R_tax is defined as follows:

[0068]

[0069] Where T_actual is the actual comprehensive tax burden rate after the strategy is executed, T_threshold is the tax burden rate warning threshold (set to 15% in this embodiment), F_rate is the capital occupancy rate, and F_threshold is the upper limit of the capital occupancy rate (set to 30% in this embodiment). These are the penalty coefficients (set to 10000 and 5000 respectively in this embodiment). When the actual tax burden rate or capital occupancy rate exceeds the threshold, a corresponding negative penalty is imposed.

[0070] The performance risk item R_risk is defined as:

[0071]

[0072] Where F is the supplier's historical performance risk index, and δ is the risk discount factor (set to 0.1 in this embodiment). When the supplier's performance risk index is high, this item imposes a greater negative penalty on high-amount, high-risk settlement strategies.

[0073] Step 4: Iterative optimization of strategy based on PPO algorithm

[0074] This embodiment uses the Proximal Policy Optimization (PPO) algorithm to iteratively optimize the pricing strategy of the core enterprise.

[0075] The policy network π_θ(a|s) and value network V_φ(s) of the PPO algorithm both employ fully connected neural network structures. The specific network parameter settings are as follows:

[0076] The policy network π_θ has 7 input neurons (corresponding to the 7 dimensions of the state vector s), two hidden layers with 64 neurons each, using the ReLU activation function, and two output neurons (outputting the purchase price and payment period respectively). The output is normalized to a reasonable range using the Tanh activation function.

[0077] Value network V_φ: 7 neurons in the input layer, 2 hidden layers with 64 neurons each, ReLU activation function, and 1 neuron in the output layer (output state value estimation).

[0078] The key hyperparameter settings for the PPO algorithm are as follows: learning rate (policy network) is... The learning rate (value network) is The discount factor γ is 0.99, the GAE parameter λ is 0.95, the pruning parameter ε is 0.2, the number of sampling trajectories per batch is 2048, the number of optimization iterations is 10, and the small batch size is 64.

[0079] In each training iteration, the system performs the following steps:

[0080] (1) Generate candidate actions a_t = [P_t, C_t] based on the current policy network π_θ, which is the combination of purchase price and payment period;

[0081] (2) Input a_t into the training environment, and the environment generates the supplier's response based on the supplier's true response function π_supplier(a|s);

[0082] (3) Based on the interaction results between the two parties, calculate the immediate reward r_t using the above reward function R_total;

[0083] (4) Store the experience samples (s_t, a_t, r_t, s_{t+1}) into the experience pool;

[0084] (5) When the number of samples in the experience pool reaches 2048, calculate the advantage function A_t = δ_t + (γλ)δ_{t+1} + ... + (γλ)^{nt}δ_n, where δ_t = r_t + γV_φ(s_{t+1}) - V_φ(s_t);

[0085] (6) Update the policy network parameters θ using the clipped surrogate objective:

[0086] L^{CLIP}(θ) = Ê_t [min(r_t(θ)·Â_t, clip(r_t(θ), 1-ε, 1+ε)·Â_t)]

[0087] Where r_t(θ) = π_θ(a_t|s_t) / π_{θ_old}(a_t|s_t);

[0088] (7) Update the network parameters φ using the mean square error loss:

[0089] L^{VF}(φ) = Ê_t [(V_φ(s_t) - R_t)^2]

[0090] (8) Repeat steps (1) to (7) until the cumulative reward of the policy network converges or reaches the preset maximum number of training rounds (set to 1 million steps in this embodiment).

[0091] Step 5: Multi-round interactive simulation and output of the optimal settlement strategy

[0092] After the policy network is trained, the system enters a multi-round interactive simulation phase to verify and output the optimal settlement strategy.

[0093] Let the initial strategy output of the core enterprise be... = [3850, 50], meaning the purchase price is 3850 yuan / ton, and the payment term is 50 days. The system distributes this strategy to the training environment, and the supplier responds based on the learned real response function π_supplier(a|s). Assuming the supplier's response is to accept the offer, the system calculates the reward for this round. .

[0094] Subsequently, the system fine-tunes the core enterprise's strategy according to the strategy update rules of the PPO algorithm, generating new candidate strategies. After 500 rounds of interactive deduction, the policy network converged to an equilibrium policy. = [3820, 42], meaning the optimal purchase price is 3820 yuan / ton and the optimal payment period is 42 days.

[0095] Under this equilibrium strategy, the core enterprise's overall economic benefits are: procurement cost savings of (4000-3820)×500=90,000 yuan; and the time value of money benefit of 3820×500×0.035×(60-42) / 365≈3,291 yuan. The supplier's tax burden rate remains at 13%, the capital occupation period is shortened from 45 days to 42 days, the overall tax burden rate does not exceed the warning threshold of 15%, the capital occupation rate does not exceed the upper limit of 30%, and the historical performance risk index is 0.08, which falls within the low-risk range.

[0096] The system outputs the optimal settlement strategy in real time and distributes it to the core enterprise's procurement management system and financial settlement system for execution via the API gateway.

[0097] Step Six: Closed-Loop Iterative Optimization

[0098] In the next procurement cycle after the optimal settlement strategy is implemented, the system continuously collects real execution feedback data through the data interface, including actual purchase price, actual payment period, actual supplier response time, actual tax burden changes, and actual capital occupation.

[0099] Suppose that actual execution data shows that due to the market interest rate rising to 4.0%, the supplier's funding costs have increased, causing a shift in its true reaction function. The system incorporates this new true feedback data into the historical invoice stream dataset, triggering the incremental update mechanism of the reaction function learning module to retrain the supplier's true reaction function π_supplier'(a|s).

[0100] Simultaneously, the system uses the state vector s' = [520, 100, 0.135, 42, 0.07, 0.040, 3820] from the actual execution data as the initial state for a new round of PPO training, and re-executes steps one through five to generate the optimal settlement strategy adapted to the new market environment. .

[0101] Example 4

[0102] This embodiment takes system-level data flow and module interaction as an example to explain in detail the interface definition, data format and complete data flow process between various functional modules.

[0103] Data Access and Feature Engineering Module

[0104] This module interfaces with the enterprise's internal ERP system, tax management system, and e-invoice platform via RESTful APIs. The main interface definitions are as follows:

[0105] (1) Invoice data retrieval interface: GET / api / v1 / invoices?start_date={start}&end_date={end}&supplier_id={id}

[0106] The request parameters include the start date, end date, and supplier ID. The returned data is in JSON format and contains the following fields: invoice_id (unique invoice identifier), issue_date (invoice date), amount (invoice amount), tax_rate (applicable tax rate), tax_amount (tax amount), payment_term (payment period in days), supplier_id (supplier ID), and material_code (material code).

[0107] (2) Supply chain transaction graph acquisition interface: GET / api / v1 / supply-chain / graph?core_enterprise_id={id}

[0108] The returned data is graph-structured, using a node-edge format. Nodes represent enterprise entities, and edges represent transaction relationships. Edge attributes include transaction frequency, cumulative transaction amount, and average payment period.

[0109] The feature engineering component preprocesses and extracts features from the raw data. The main processing flow is as follows:

[0110] First, the invoice data is cleaned, removing fields with a missing rate exceeding 5%, and filling in fields with a missing rate within 5% using linear interpolation. Second, the supplier's tax burden rate is calculated: T = Σ(tax_amount_i) / Σ(amount_i), where the summation covers all invoice records for that supplier within the sliding window period. Third, the capital occupation period is calculated: C = Σ(payment_term_i × amount_i) / Σ(amount_i), which is the weighted average payment period weighted by transaction amount. Finally, the historical performance risk index is calculated. ,in These are weighting coefficients (set to 0.4, 0.3, and 0.3 respectively in this embodiment).

[0111] After processing, the module pushes the joint state vector s to the collaborative pricing decision engine in real time in the form of an Apache Kafka message queue. The message format is Protocol Buffers serialization format to reduce network transmission overhead.

[0112] Reaction function learning module:

[0113] This module has two built-in learning modes: inverse reinforcement learning mode (suitable for scenarios with sufficient data and complete behavioral trajectories) and parameter estimation mode (suitable for scenarios with limited data and the need for rapid modeling). The module receives state-behavior sequence data from the data access module via the POST interface / api / v1 / reaction-function / learn and returns the learned reaction function model file (ONNX format).

[0114] In this embodiment, since the historical data includes complete invoice records from the past 36 months (approximately 120,000 records), the system defaults to enabling inverse reinforcement learning mode. The module uses the PyTorch framework to implement the MaxEnt IRL algorithm, supporting GPU acceleration (in this embodiment, an NVIDIA RTX 3090 graphics card is used, with a single training round taking approximately 15 minutes). After learning is complete, the reaction function model is exported in ONNX format and deployed in real-time to the inference service of the collaborative pricing decision engine via the gRPC interface.

[0115] Collaborative pricing decision engine

[0116] This module is the core decision-making component of the system. It adopts a microservice architecture and mainly includes three sub-services: game model service, PPO training service, and strategy inference service.

[0117] The game model service is responsible for constructing the environment model of the Stackelberg game and defining the interaction rules between the core firm (leader) and suppliers (followers). This service receives action query requests from the strategy deduction service via the gRPC interface and returns the supplier's response actions.

[0118] The PPO training service is responsible for executing the training process of deep reinforcement learning algorithms. This service uses the Ray distributed training framework and supports multi-machine, multi-GPU parallel training. In this embodiment, the training cluster consists of four servers, each configured with eight NVIDIA A100 GPUs, and the total training time is approximately six hours. After training, the policy network model is saved in TensorFlowSavedModel format and version controlled through the Model Version Management Service (MLflow).

[0119] The strategy simulation service is responsible for executing multi-round interactive simulations and outputting the optimal settlement strategy. This service loads the trained PPO strategy network and the supplier reaction function model exported from the reaction function learning module. Through Monte Carlo simulation or deterministic simulation, it simulates the multi-round interaction process between upstream and downstream enterprises and outputs an equilibrium strategy. The simulation results are returned in JSON format, containing the fields: optimal_price (optimal purchase price), optimal_term (optimal payment term), expected_reward (expected cumulative reward), convergence_round (number of convergence rounds), and equilibrium_type (equilibrium type, such as Stackelberg equilibrium or Nash equilibrium).

[0120] Strategy Execution and Feedback Module

[0121] This module connects to the enterprise's business systems through an API gateway, and is responsible for the distribution and execution of policies and the collection and feedback of execution data.

[0122] Strategy delivery interface: POST / api / v1 / strategy / execute

[0123] The request body contains the following fields: strategy_id (unique strategy identifier), core_enterprise_id (core enterprise ID), supplier_id (supplier ID), optimal_price (optimal purchase price), optimal_term (optimal payment term), and effective_date (effective date).

[0124] This interface distributes the optimal settlement strategy to the enterprise's procurement management system. The procurement management system automatically generates purchase orders based on the strategy and pushes them to the supplier. After the supplier confirms the order, the financial settlement system automatically arranges a payment plan based on the optimal payment terms.

[0125] Execute feedback interface: POST / api / v1 / strategy / feedback

[0126] This interface is called periodically by the business system (set to 2 AM daily in this embodiment) to send the execution results data from the previous day back to the system. The returned data includes the following fields: actual_price (actual transaction price), actual_term (actual payment period), supplier_response (supplier response type, such as accept / reject / counter-quote), actual_tax_rate (actual overall tax rate), actual_fund_occupancy_rate (actual fund occupancy rate), and fulfillment_status (fulfillment status, such as normal / overdue / cancelled).

[0127] The strategy execution and feedback module writes the collected real execution data into the time series database (InfluxDB is used in this embodiment), and triggers the data update process of the data access and feature engineering module and the incremental training process of the reaction function learning module to realize the closed-loop iterative optimization of the system.

[0128] Through the aforementioned closed-loop iterative optimization mechanism, the system achieves continuous self-learning and adaptive optimization of the supply chain collaborative pricing strategy, ensuring that the strategy always remains synchronized with the real business environment.

[0129] The present invention has been described in detail above with reference to specific exemplary embodiments. However, it should be understood that various modifications and variations can be made without departing from the scope of the invention as defined by the appended claims. The detailed description should be considered illustrative only and not restrictive, and any such modifications and variations shall fall within the scope of the invention described herein. Furthermore, the background art is intended to illustrate the current state of development and significance of the technology and is not intended to limit the present invention or the scope of application of the present application.

[0130] More specifically, although exemplary embodiments of the invention have been described herein, the invention is not limited to these embodiments, but includes any and all embodiments modified, omitted, such as combinations between various embodiments, adaptive changes, and / or substitutions, as would be apparent to those skilled in the art from the foregoing detailed description. The limitations in the claims are to be interpreted broadly as used in the language of the claims and are not limited to the examples described in the foregoing detailed description or during the implementation of this application, which should be considered non-exclusive. Any step listed in any method or process claim may be performed in any order and is not limited to the order set forth in the claims. Therefore, the scope of the invention should be determined solely by the appended claims and their legal equivalents, and not by the description and examples given above.

Claims

1. A supply chain collaborative pricing optimization method driven by fiscal and tax characteristics, characterized in that, Includes the following steps: Step 1: Based on real supply chain transaction graphs and historical invoice flow data, construct the state space of the supply chain training environment and deep reinforcement learning. The state space contains multi-dimensional financial and tax features. Step 2: Based on the historical invoice flow data, use inverse reinforcement learning or parameter estimation methods to automatically learn the supplier's true response function under specific financial and tax constraints; Step 3: Construct a collaborative pricing model that integrates Stackelberg game theory and deep reinforcement learning, with the core enterprise as the leader and the suppliers as followers, and use a reward function with embedded fiscal and tax constraints to optimize the pricing strategy of the core enterprise. The reward function includes economic benefit items, fiscal and tax penalty items, and performance risk items. Step 4: Use deep reinforcement learning algorithms to iteratively optimize the pricing strategy of the core enterprise; Step 5: Through multiple rounds of interactive deduction, the equilibrium point of upstream and downstream interests is deduced, and the optimal settlement strategy that balances financial and tax compliance with supply chain collaboration is output in real time.

2. The supply chain collaborative pricing optimization method driven by fiscal and tax characteristics according to claim 1, characterized in that, The multidimensional financial and tax characteristics include the supplier's tax burden cost rate, capital occupation period, and historical performance risk index.

3. The supply chain collaborative pricing optimization method driven by fiscal and tax characteristics according to claim 1 or 2, characterized in that, The state space also includes basic supply and demand parameters, including demand and inventory levels. The multidimensional fiscal and tax features are integrated with the basic supply and demand parameters to form a joint state vector.

4. The supply chain collaborative pricing optimization method driven by fiscal and tax characteristics according to claim 1, characterized in that, The historical invoice stream data includes invoice date, amount, tax rate, and payment period information.

5. The supply chain collaborative pricing optimization method driven by fiscal and tax characteristics according to claim 1, characterized in that, The deep reinforcement learning algorithm includes the proximal policy optimization algorithm or the soft actor-critic algorithm.

6. The supply chain collaborative pricing optimization method driven by fiscal and tax characteristics according to claim 1, characterized in that, The economic benefits include cost savings in procurement and the time value of money for core enterprises; the tax penalties are used to impose negative penalties when the strategy leads to an abnormal increase in the overall tax burden or exceeds the limit of capital occupancy; and the performance risk is used to discount the risk of high-risk settlement strategies based on the performance risk index.

7. The supply chain collaborative pricing optimization method driven by fiscal and tax characteristics according to claim 1, characterized in that, The optimal settlement strategy includes the optimal combination of purchase price and payment term.

8. The supply chain collaborative pricing optimization method driven by fiscal and tax characteristics according to claim 1, characterized in that, After step five, the process also includes: distributing the optimal settlement strategy to the business system and collecting real financial and tax data after execution for closed-loop iterative optimization.

9. A supply chain collaborative pricing optimization system driven by fiscal and tax characteristics, characterized in that, include: The data access and feature engineering module is used to connect with supply chain transaction graphs and invoice flow data, extract and structure multi-dimensional financial and tax features, and construct a state space for deep reinforcement learning. The reaction function learning module is used to automatically learn the supplier's true reaction function under specific financial and tax constraints based on the invoice flow data, using inverse reinforcement learning or parameter estimation methods. The collaborative pricing decision engine is used to build a collaborative pricing model that integrates Stackelberg game theory and deep reinforcement learning. It treats core enterprises as leaders and suppliers as followers, and generates strategies, calculates rewards, and deduces equilibrium points based on reward functions with embedded fiscal and tax constraints. The strategy execution and feedback module is used to distribute the optimal settlement strategy to the business system and collect real financial and tax data after execution for closed-loop iteration.

10. The supply chain collaborative pricing optimization system driven by fiscal and tax characteristics according to claim 9, characterized in that, The deep reinforcement learning algorithm includes the proximal policy optimization algorithm or the soft actor-critic algorithm.