Tobacco enterprise composite scheduling method based on multi-agent collaborative optimization

Abstract

The invention discloses a tobacco enterprise composite scheduling method based on multi-agent collaborative optimization, and the method comprises the steps: obtaining the system data of a tobacco enterprise, and obtaining a plurality of corresponding decision-making schemes based on a sales prediction agent, a production plan agent, a purchase stock agent, and an inventory management agent; performing iterative optimization on the decision scheme based on a Nash equilibrium game and a Lupinstein bargaining model, and calculating a system utility value to obtain an optimal decision scheme; according to the optimal decision-making scheme, based on an adaptive mechanism, adaptive optimization is carried out, and an adjusted decision-making scheme of each agent is obtained, so that tobacco enterprise composite scheduling of multi-agent collaborative optimization is realized; the sales prediction agent, the production plan agent, the purchase stock agent and the inventory management agent perform information interaction through an attention mechanism. Interactive cooperation among multiple agents is realized by introducing an attention communication mechanism and a game theory cooperation mechanism, and overall optimization of tobacco enterprise scheduling is promoted.

Description

Technical Field

[0001] The invention relates to a tobacco enterprise composite scheduling method based on multi-agent collaborative optimization, and belongs to the technical field of tobacco enterprise scheduling. Background Art

[0002] As a traditional manufacturing industry, the tobacco industry is facing an increasingly complex market environment and fierce competition. As tobacco companies continue to expand in scale and their product categories become increasingly diverse, how to achieve efficient collaboration and optimize resource allocation in various links such as sales, production, procurement, and inventory has become an urgent problem for tobacco companies to solve.

[0003] The traditional dispatching mode of tobacco enterprises often relies on manual experience and simple information tools, and there are problems such as information islands, decision-making lags, and waste of resources. At present, there are some dispatching optimization systems for tobacco enterprises based on intelligent algorithms, such as systems using genetic algorithms, particle swarm optimization and other methods, but most of these systems optimize single sub-problems and lack a global perspective and coordination mechanism. In addition, the tobacco industry is also facing many challenges such as inaccurate sales forecasts, frequent changes in production plans, insufficient procurement and inventory backlogs. The accuracy of sales forecasts directly affects the formulation and execution of production plans, while frequent adjustments to production plans will lead to low production efficiency and serious waste of resources. Insufficient procurement and inventory may lead to production interruptions, while inventory backlogs will occupy a lot of funds and increase the operating costs of enterprises. The root cause of these problems lies in the lack of information sharing and decision-making coordination between various business links. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a composite scheduling method for tobacco enterprises based on multi-agent collaborative optimization. By introducing attention communication mechanism and game theory collaboration mechanism, interactive collaboration among sales forecasting agents, production planning agents, purchasing and stocking agents, and inventory management agents is achieved, thereby promoting the overall optimization of tobacco enterprise scheduling.

[0005] To achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0006] The present invention discloses a tobacco enterprise composite scheduling method based on multi-agent collaborative optimization, comprising:

[0007] Obtain system data from tobacco companies;

[0008] According to the system data, multiple corresponding decision-making schemes are obtained based on the sales forecasting agent, the production planning agent, the purchasing and stocking agent, and the inventory management agent;

[0009] Iteratively optimizing the decision-making scheme based on the Nash equilibrium game and the Rubinstein bargaining model, and obtaining the optimal decision-making scheme by calculating the system utility value;

[0010] According to the optimal decision-making scheme, adaptive optimization is performed based on the adaptive mechanism to obtain the adjusted decision-making schemes of each intelligent agent, so as to realize the composite scheduling of tobacco enterprises with multi-agent collaborative optimization;

[0011] Among them, the sales forecasting agent, production planning agent, purchasing and stocking agent, and inventory management agent exchange information through the attention mechanism.

[0012] Furthermore, the system data includes historical sales data, market environment data, seasonal factor data, production status data, inventory status data, raw material market data and supplier data;

[0013] The decision-making scheme includes sales forecasting decision, production planning decision, purchase and stocking decision and inventory management decision;

[0014] The sales forecasting agent is used to obtain sales forecasting decisions based on the historical sales data, market environment data and seasonal factor data and interactive information of other agents;

[0015] The production planning agent is used to obtain the production planning decision based on the sales forecast decision, production status data, inventory status data and interactive information of other agents;

[0016] The purchasing and stocking agent is used to obtain purchasing and stocking decisions based on the production plan decision, raw material market data, supplier data, and interactive information of other agents;

[0017] The inventory management agent is used to obtain inventory management decisions based on the sales forecast decision, production plan decision, purchase stocking decision, inventory status data and interactive information of other agents;

[0018] Among them, the sales forecasting agent adopts an attention-based long short-term memory network algorithm, and the production planning agent, procurement and stocking agent, and inventory management agent adopt an algorithm that combines deep reinforcement learning and multi-objective optimization.

[0019] Furthermore, the sales forecasting agent includes:

[0020] The sales forecast input encoding layer is used to encode the sales feature input vector to obtain a sales state vector; the sales feature input vector includes historical sales data, market environment data and seasonal factor data;

[0021] The sales prediction attention layer is used to obtain a sales attention vector by calculating the attention weight according to the sales state vector.

[0022] A sales prediction decoding layer, used to obtain a sales prediction decision according to the sales attention vector;

[0023] A sales forecast utility function layer, used to calculate the sales forecast utility value according to the sales forecast decision;

[0024] The expression of the sales forecast utility value is as follows:

[0025] ;

[0026] In the formula, represents the sales forecast utility value; represents the sales forecast decision at the tth time step; represents the actual sales quantity at the tth time step; represents the first weight coefficient; represents the second weight coefficient; Indicates the time span of the forecast.

[0027] Furthermore, the production planning agent includes:

[0028] The production plan state encoding layer is used to encode the production feature input vector to obtain the production plan state vector of the current time step; the production feature input vector includes sales forecast decision, production status data and inventory status data;

[0029] The production plan decision network layer is used to obtain the corresponding production plan decision according to the production plan state vector of the current time step;

[0030] The production plan state transfer function layer is used to obtain the production plan state vector of the next time step according to the production plan state vector of the current time step and the corresponding production plan decision;

[0031] The production plan reward function layer is used to calculate the reward value corresponding to the production plan decision at the current time step based on the production-sales connection, production efficiency, inventory rationality and game equilibrium factor;

[0032] The production plan utility function layer is used to calculate the production plan utility value based on the production plan efficiency, production plan cost, and the degree of deviation of the production plan from the sales forecast and inventory target;

[0033] The expression of the production plan utility value is as follows:

[0034] ;

[0035] ;

[0036] ;

[0037] ;

[0038] In the formula, represents the production plan utility value; The weight coefficient representing the efficiency of production planning; It indicates the efficiency of production planning; The weight coefficient representing the production plan cost; represents the production plan cost; The weight coefficient indicating the degree of deviation; Indicates the degree of deviation of the production plan from the sales forecast and inventory target;

[0039] Indicates The production quantity of each product; Indicates Production time of the product; Indicates The production capacity occupied by the product; Indicates The unit production cost of a product; Indicates Target inventory level for each product; Indicates The time the equipment is used; Indicates The time unit cost of the equipment; Indicates The amount of energy consumed; Indicates The unit price of the energy source; Indicates the number of product types; Indicates the number of types of equipment; Indicates the number of energy types.

[0040] Furthermore, the purchasing and stocking agent includes:

[0041] The procurement stocking status encoding layer is used to encode the procurement feature input vector to obtain the procurement stocking status vector of the current time step; the procurement feature input vector includes production plan decision, raw material market data and supplier data;

[0042] A purchasing and stocking decision network layer is used to obtain a corresponding purchasing and stocking decision according to the purchasing and stocking state vector of the current time step;

[0043] The procurement stocking state transfer function layer is used to obtain the procurement stocking state vector of the next time step according to the procurement stocking state vector of the current time step and the corresponding procurement stocking decision;

[0044] The procurement and stocking reward function layer is used to calculate the reward value corresponding to the procurement and stocking decision at the current time step based on procurement cost, inventory rationality, supply stability, and attention evaluation results;

[0045] The procurement stocking utility function layer is used to calculate the procurement stocking utility value based on procurement availability, procurement stocking cost and inventory pressure;

[0046] The expression of the purchase stocking utility value is as follows:

[0047] ;

[0048] ;

[0049] ;

[0050] ;

[0051] In the formula, Indicates the purchase and stocking utility value; The weight coefficient representing the availability of procurement; Indicates purchasing availability; Indicates the weight coefficient of procurement stocking cost; Represents the purchase and stocking cost; The weight coefficient representing the inventory pressure; Indicates inventory pressure;

[0052] Indicates The demand for raw materials; Indicates The arrival rate of raw materials; Indicates The purchase volume of raw materials; Indicates Unit price of the raw materials; Indicates Unit transportation cost of the raw materials; Indicates The actual inventory of the raw materials; Indicates Safety stock of raw materials; Indicates the number of types of raw materials.

[0053] Furthermore, the inventory management agent includes:

[0054] The inventory management state encoding layer is used to encode the inventory feature input vector to obtain the inventory management state vector of the current time step; the inventory feature input vector includes sales forecast decision, production plan decision, purchase stocking decision and inventory status data;

[0055] An inventory management decision network layer, used to obtain a corresponding inventory management decision according to the inventory management state vector of the current time step;

[0056] An inventory management state transfer function layer, used to obtain the inventory management state vector of the next time step according to the inventory management state vector of the current time step and the corresponding inventory management decision;

[0057] The inventory management reward function layer is used to calculate the reward value corresponding to the inventory management decision at the current time step based on the inventory turnover rate, out-of-stock rate, slow-moving product rate, and attention evaluation results;

[0058] Inventory management utility function layer, used to calculate the inventory management utility value based on order fulfillment rate, inventory cost and out-of-stock loss;

[0059] The expression of the inventory management utility value is as follows:

[0060] ;

[0061] ;

[0062] ;

[0063] ;

[0064] In the formula, represents the inventory management utility value; The weight coefficient representing the order fulfillment rate; represents the order fulfillment rate; Indicates the weight coefficient of inventory cost; represents inventory cost; The weight coefficient representing the out-of-stock loss; Indicates loss due to shortage of stock;

[0065] Indicates The order volume of the product; Indicates Available inventory of the product; Indicates Unit inventory holding cost of a product; Indicates The number of times a product is put in and taken out of the warehouse; Indicates The single inbound and outbound cost of a product; Indicates Unit storage cost of the product; Indicates The unit price of the product; Indicates The expected loss from a stock-out of each unit of a product; Indicates the number of product types.

[0066] Furthermore, the expression of the system utility value is as follows:

[0067]

[0068] In the formula, represents the system utility value; The weight representing the sales forecast utility value; represents the sales forecast utility value; The weight representing the utility value of the production plan; represents the production plan utility value; The weight representing the utility value of purchasing and stocking; Indicates the utility value of purchasing and stocking; The weight representing the utility value of inventory management; represents the inventory management utility value;

[0069] represents the weight of decision consistency in system utility; Represents decisions based on different agents and The decision consistency evaluation function, when m or h is s, it represents the sales forecasting agent. or When the value is p, it represents the production planning agent. or When the value is b, it indicates the purchasing and stocking agent. or When the value is d, it indicates an inventory management agent;

[0070] Indicates the weight of system profit in system utility; represents the expected income of the system, Represents the expected cost of the system.

[0071] Furthermore, the expression of the expected income of the system is as follows:

[0072] ;

[0073] In the formula, Indicates The unit price of the product; Indicates The product is in The sales decision variables at time step are given by the sales forecasting agent; Indicates In the demand scenario The product is in The actual demand for the time step; Indicates the quantity of the product; Indicates the number of demand scenarios; Indicates the number of time steps.

[0074] Furthermore, the expression of the expected cost of the system is as follows:

[0075] ;

[0076] In the formula, Indicates The product is in Unit production cost per time step; Indicates The product is in The production decision variables at the time step are given by the production planning agent;

[0077] Indicates The product is in Unit inventory cost at the time step; Indicates that the i-th product is in The expected inventory at the time step, ;

[0078] Indicates The product is in Unit out-of-stock cost at time step; Indicates The product is in The inventory decision variables at the time step are given by the inventory management agent.

[0079] Furthermore, the expression for the adaptive optimization based on the adaptive mechanism is as follows:

[0080] ;

[0081] ;

[0082] ;

[0083] ;

[0084] ;

[0085] In the formula, Indicates that after optimization The state of the agent at the time step; Indicates that after optimization Agent decision at time step; Indicates in status Take decision The expected long-term return;

[0086] Indicated in The state of the agent at the time step; Indicated in Agent decision at time step; Indicates in status Take decision The expected long-term return;

[0087] represents the learning rate; Indicates System utility value at the time step; represents the discount factor; Indicates in status Take decision The expected long-term return;

[0088] represents the profit shared by the sales forecasting agent; represents the sales forecast utility value; represents the profit shared by the production planning agent; represents the production plan utility value; Represents the profit shared by the purchasing and stocking agent; Indicates the utility value of purchasing and stocking; represents the profit shared by the inventory management agent; represents the inventory management utility value; Indicates the total profit of the system.

[0089] Compared with the prior art, the present invention has the following beneficial effects:

[0090] The composite scheduling method for tobacco enterprises based on multi-agent collaborative optimization of the present invention realizes interactive collaboration among sales forecasting agents, production planning agents, purchasing and stocking agents, and inventory management agents by introducing attention communication mechanisms and game theory collaborative mechanisms; each agent selectively pays attention to and transmits key information through the attention mechanism, and each agent can adapt to different decision-making scenarios and information needs by dynamically adjusting the attention weight; and through the Nash equilibrium game and Rubinstein bargaining model, each agent coordinates decisions based on the game theory collaborative mechanism, achieves balance and compromise of the decisions of each agent, and promotes the overall optimization of tobacco enterprise scheduling. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] Figure 1 It is a flow chart of a composite scheduling method for tobacco enterprises based on multi-agent collaborative optimization provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0092] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0093] This embodiment provides a tobacco enterprise composite scheduling method based on multi-agent collaborative optimization. Figure 1 As shown, including:

[0094] Obtain system data from tobacco companies;

[0095] According to the system data, multiple corresponding decision-making schemes are obtained based on the sales forecasting agent, production planning agent, purchasing and stocking agent, and inventory management agent;

[0096] Based on the Nash equilibrium game and Rubinstein bargaining model, the decision-making scheme is iteratively optimized, and the optimal decision-making scheme is obtained by calculating the system utility value;

[0097] According to the optimal decision-making plan, adaptive optimization is performed based on the adaptive mechanism to obtain the adjusted decision-making plans of each intelligent agent, so as to realize the composite scheduling of tobacco enterprises with multi-agent collaborative optimization;

[0098] Among them, the sales forecasting agent, production planning agent, purchasing and stocking agent, and inventory management agent exchange information through the attention mechanism.

[0099] The technical concept of the present invention is: by introducing the attention communication mechanism and the game theory coordination mechanism, the interactive collaboration among the sales forecasting agent, the production planning agent, the purchasing and stocking agent, and the inventory management agent is realized; each agent selectively pays attention to and transmits key information through the attention mechanism, and by dynamically adjusting the attention weight, each agent can adapt to different decision-making scenarios and information needs; and through the Nash equilibrium game and the Rubinstein bargaining model, each agent coordinates the decision based on the game theory coordination mechanism, realizes the balance and compromise of the decision of each agent, and promotes the overall optimization of the scheduling of tobacco enterprises.

[0100] The specific steps of the present invention are as follows:

[0101] Step 1: Obtain system data of tobacco companies.

[0102] System data includes historical sales data, market environment data, seasonal factor data, production status data, inventory status data, raw material market data and supplier data.

[0103] Step 2: Based on the system data, multiple corresponding decision-making plans are obtained based on the sales forecasting agent, production planning agent, purchasing and stocking agent, and inventory management agent.

[0104] Among them, the sales forecasting agent adopts the attention-based long short-term memory network algorithm, and the production planning agent, procurement and stocking agent, and inventory management agent adopt an algorithm that combines deep reinforcement learning and multi-objective optimization, as follows:

[0105] 2.1. Sales forecasting agent, used to obtain sales forecasting decisions based on historical sales data, market environment data, seasonal factor data and interactive information of other agents.

[0106] The sales forecasting agent includes a sales forecasting input encoding layer, a sales forecasting attention layer, a sales forecasting decoding layer, and a sales forecasting utility function layer.

[0107] 2.1.1. The sales prediction input encoding layer is used to encode the sales feature input vector using the LSTM network, and obtain the sales state vector by generating the feature representation enhanced by attention. The expression is as follows:

[0108] ;

[0109] In the formula, represents the sales state vector at the tth time step; represents the attention-based long short-term memory network function; Represents the sales feature input vector, including historical sales data, market environment data, and seasonal factor data; Represents the sales state vector at the t-1th time step.

[0110] 2.1.2. The sales prediction attention layer is used to dynamically focus on important input information by calculating the attention weight according to the sales state vector to obtain the sales attention vector. The expression is as follows:

[0111] ;

[0112] ;

[0113] ;

[0114] In the formula, represents the sales attention vector at the t-th time step; N represents the number of sales features; Indicates The time step is The attention weight of each sales feature; Indicates The attention vector of sales features;

[0115] Indicates The time step is Attention scores for sales features; represents the weight vector; represents the hyperbolic tangent function; represents the sales status weight matrix; represents the sales state vector at the tth time step; represents the sales feature weight matrix; represents the agent interaction weight matrix; Represents the attention vector that the sales forecasting agent obtains from the interactions with other agents.

[0116] It should be noted here that the attention information vector of other agents is reflected by the combination of the output values ​​of other agents for multiple future time steps. For example, the future output value of the production planning agent The production plan output results for each time step This means that the production planning agent will The production focus and resource allocation tendency of each time step can be regarded as an implicit attention information, that is, the attention information vector transmitted by the production planning agent Similarly, the attention information vector transmitted by the purchasing agent is It shows that the purchasing and stocking agent will The purchasing and stocking strategy for time steps, the attention information vector transmitted by the inventory management agent It shows that inventory management agents will have The inventory management strategy for time steps. At this moment, the attention information vector of other agents received by the sales prediction agent is It can be expressed as .

[0117] 2.1.3. The sales prediction decoding layer is used to obtain the sales prediction decision based on the sales attention vector. The expression is as follows:

[0118] ;

[0119] In the formula, represents the sales forecast decision at the tth time step; Indicates a connection operation; represents the sales forecast decision at the t-1th time step.

[0120] The sales forecast decision also indicates the sales forecast value for a certain period in the future.

[0121] 2.1.4. The sales forecast utility function layer is used to calculate the sales forecast utility value based on the sales forecast decision. The expression is as follows:

[0122] ;

[0123] In the formula, represents the sales forecast utility value; represents the actual sales quantity at the t-th time step; represents the first weight coefficient; represents the second weight coefficient; Indicates the time span of the forecast.

[0124] The first calculation term of this formula uses the mean absolute percentage error to represent the prediction accuracy of the sales forecasting agent, and the second calculation term of this formula uses the mean square error to represent the prediction loss of the sales forecasting agent.

[0125] In addition, this method uses the back propagation algorithm to optimize the model parameters of the sales forecasting agent, and the steps are:

[0126] Forward propagation: input data into the sales forecasting agent to get the forecast output;

[0127] Calculate the loss function: Calculate the loss function of the sales forecasting agent based on the predicted output;

[0128] Back propagation: Update model parameters through the back propagation algorithm to minimize the loss function;

[0129] Repeat the above steps until the model converges or reaches the predetermined number of training rounds.

[0130] 2.2. Production planning agent is used to obtain production planning decisions based on sales forecast decisions, production status data, inventory status data and interactive information from other agents.

[0131] Specifically include:

[0132] 2.2.1. The production plan state encoding layer is used to encode the production feature input vector to obtain the production plan state vector of the current time step. The expression is as follows:

[0133] ;

[0134] In the formula, represents the production plan state vector at the tth time step; ReLU activation function. The weight matrix representing the production plan status encoding; Represents the production feature input vector, including sales forecast decisions, production status data, and inventory status data; Represents the strategy vector obtained by the production planning agent from the interaction with other agents; Indicates the offset item of the production plan status code.

[0135] It should be noted that It represents the strategy vector obtained by the production planning agent from the interaction with other agents. It refers to the output results of other agents at the current time step, that is, the tth time step. It is different from the attention information vector of the interaction with other agents. It focuses more on the short-term perspective and reflects the current decision-making behavior of other agents.

[0136] 2.2.2. The production plan decision network layer is used for the production plan state vector of the current time step to obtain the corresponding production plan decision. The expression is as follows:

[0137] ;

[0138] In the formula, represents the production planning decision at the tth time step; Indicated in the state vector Next choice decision Expected long-term returns; Represents a set of feasible production planning decisions.

[0139] The production planning decision network layer is implemented using a deep Q network (DQN).

[0140] 2.2.3. The production plan state transfer function layer is used for the production plan state vector of the current time step and the corresponding production plan decision to obtain the production plan state vector of the next time step. The expression is as follows:

[0141] ;

[0142] In the formula, represents the production plan state vector at the t+1th time step; The weight matrix representing the state transfer function; The offset term that represents the production plan state transition.

[0143] 2.2.4. The production plan reward function layer is used to calculate the reward value corresponding to the production plan decision at the current time step based on the production-sales connection, production efficiency, inventory rationality and game equilibrium factor. The expression is as follows:

[0144] ;

[0145] In the formula, represents the reward value corresponding to the production planning decision at the tth time step;

[0146] It represents the weight coefficient of production-marketing connection; Indicates the reward item for production-marketing linkage;

[0147] represents the production efficiency weight coefficient; Indicates the production efficiency reward item;

[0148] Represents the inventory rationality weight coefficient; Indicates the inventory rationality reward item;

[0149] represents the game equilibrium weight coefficient; Represents the game equilibrium reward item.

[0150] 2.2.5. The production plan utility function layer is used to calculate the production plan utility value based on the production plan efficiency, production plan cost, and the degree of deviation of the production plan from the sales forecast and inventory target. The expression is as follows:

[0151] ;

[0152] ;

[0153] ;

[0154] ;

[0155] In the formula, represents the production plan utility value; The weight coefficient representing the efficiency of production planning; It indicates the efficiency of production planning; The weight coefficient representing the production plan cost; represents the production plan cost; The weight coefficient indicating the degree of deviation; Indicates the degree of deviation of the production plan from the sales forecast and inventory target;

[0156] Indicates The production quantity of each product; Indicates Production time of the product; Indicates The production capacity occupied by the product; Indicates The unit production cost of a product; Indicates Target inventory level for each product; Indicates The time the equipment is used; Indicates The time unit cost of the equipment; Indicates The amount of energy consumed; Indicates The unit price of the energy source; Indicates the number of product types; Indicates the number of types of equipment; Indicates the number of energy types.

[0157] 2.3. Purchasing and stocking agent is used to obtain purchasing and stocking decisions based on production plan decisions, raw material market data, supplier data and interactive information from other agents.

[0158] Similar to the production planning agent, the procurement and stocking agent also uses a procurement and stocking model that combines deep reinforcement learning and multi-objective optimization as its core algorithm. Its components include the procurement and stocking state encoding layer, the procurement and stocking decision network layer, the procurement and stocking state transfer function layer, the procurement and stocking reward function layer, and the procurement and stocking utility function layer. The details are as follows:

[0159] 2.3.1. The procurement stocking status encoding layer is used to encode the procurement feature input vector to obtain the procurement stocking status vector of the current time step.

[0160] The expression of the procurement stocking state vector at the current time step is as follows:

[0161] ;

[0162] In the formula, Represents the purchasing stocking state vector at the tth time step; ReLU activation function. The weight matrix representing the procurement stocking status coding; Represents the input vector of purchasing features, including production planning decisions, raw material market data, and supplier data; Represents the strategy vector obtained by the purchasing and stocking agent from the interaction with other agents; Indicates the offset item of the purchase stock status code.

[0163] The difference between the production planning agent and the production planning agent, in addition to the difference in input features, is that the production planning agent incorporates the strategy information of other agents into the state encoding in order to better understand the current production environment and make corresponding decision adjustments. This is because in the process of formulating the optimal production plan, the strategies of other agents have a direct impact on the decision of the production planning agent. The main goal of the procurement and stocking agent is to specify the optimal procurement and stocking strategy based on the production plan and market conditions. Therefore, the procurement and stocking agent incorporates the attention information vectors of other agents into the state encoding to better understand the current supply chain environment.

[0164] 2.3.2. The purchasing and stocking decision network layer is used to obtain the corresponding purchasing and stocking decision according to the purchasing and stocking state vector of the current time step;

[0165] The expression of purchasing stocking decision is as follows:

[0166] ;

[0167] In the formula, represents the purchasing and stocking decision at the tth time step; Indicated in the state vector Next choice decision The probability of Represents a set of feasible purchasing and stocking decisions.

[0168] Specifically, the procurement and stocking decision network layer is implemented using the proximal policy optimization (PPO) algorithm.

[0169] 2.3.3. The procurement and stocking state transfer function layer is used for the procurement and stocking state vector of the current time step and the corresponding procurement and stocking decision to obtain the procurement and stocking state vector of the next time step.

[0170] The expression of the purchase stocking state vector at the next time step is as follows:

[0171] ;

[0172] In the formula, Represents the purchasing stocking state vector at the t+1th time step; Indicates the use of deep neural networks for state transfer;

[0173] The procurement and stocking state transfer function layer is implemented using a deep neural network.

[0174] 2.3.4. Procurement and stocking reward function layer, which is used to calculate the reward value corresponding to the procurement and stocking decision at the current time step based on procurement cost, inventory rationality, supply stability and attention evaluation results.

[0175] The expression of reward value is as follows:

[0176] ;

[0177] In the formula, Represents the reward value corresponding to the purchase and stocking decision at the tth time step;

[0178] represents the reward function based on the weighted sum method; represents the purchase cost; Indicates inventory rationality; Indicates supply stability; Represents the attention evaluation result of purchasing and stocking.

[0179] 2.3.5. Procurement and stocking utility function layer, used to calculate the procurement and stocking utility value based on procurement availability, procurement and stocking cost and inventory pressure;

[0180] The expression of the purchasing stocking utility value is as follows:

[0181] ;

[0182] ;

[0183] ;

[0184] ;

[0185] In the formula, Indicates the purchase and stocking utility value; The weight coefficient representing the availability of procurement; Indicates purchasing availability; Indicates the weight coefficient of procurement stocking cost; Represents the purchase and stocking cost; The weight coefficient representing the inventory pressure; Indicates inventory pressure;

[0186] Indicates The demand for raw materials; Indicates The arrival rate of raw materials; Indicates The purchase volume of raw materials; Indicates Unit price of the raw materials; Indicates Unit transportation cost of the raw materials; Indicates The actual inventory of the raw materials; Indicates Safety stock of raw materials; Indicates the number of types of raw materials.

[0187] 2.4. Inventory management agent, which is used to make inventory management decisions based on sales forecast decisions, production planning decisions, purchasing and stocking decisions, inventory status data, and interactive information from other agents;

[0188] Similarly, the inventory management agent also uses an inventory management model that combines deep reinforcement learning and multi-objective optimization as its core algorithm. Its components include an inventory management state encoding layer, an inventory management decision network layer, an inventory management state transfer function layer, an inventory management reward function layer, and an inventory management utility function layer, including:

[0189] 2.4.1. Inventory management state encoding layer is used to encode the inventory feature input vector to obtain the inventory management state vector of the current time step.

[0190] The expression of the inventory management state vector at the current time step is as follows:

[0191]

[0192] In the formula, represents the inventory management state vector at the tth time step; ReLU activation function. The weight matrix representing the inventory management status encoding; Represents the inventory feature input vector, including sales forecast decisions, production planning decisions, purchase stocking decisions and inventory status data; represents the strategy vector obtained by the inventory management agent from the interaction with other agents; The bias term representing the inventory management status code;

[0193] 2.4.2. Inventory management decision network layer is used for the inventory management state vector of the current time step to obtain the corresponding inventory management decision.

[0194] The expression of inventory management decision is as follows:

[0195] ;

[0196] In the formula, represents the inventory management decision at the tth time step; Indicated in the state vector Next choice decision The probability of represents the set of feasible inventory management decisions;

[0197] The inventory management decision network layer is also implemented using the Proximal Policy Optimization (PPO) algorithm

[0198] 2.4.3. Inventory management state transfer function layer, which is used for the inventory management state vector of the current time step and the corresponding inventory management decision to obtain the inventory management state vector of the next time step.

[0199] The expression of the inventory management state vector at the next time step is as follows:

[0200] ;

[0201] In the formula, represents the inventory management state vector at the t+1th time step; Represents the use of deep neural networks for state transfer.

[0202] 2.4.4. Inventory management reward function layer, which is used to calculate the reward value corresponding to the inventory management decision at the current time step based on the inventory turnover rate, out-of-stock rate, slow-moving product rate and attention evaluation results.

[0203] The expression of reward value is as follows:

[0204] ;

[0205] In the formula, represents the reward value corresponding to the inventory management decision at the tth time step;

[0206] represents the reward function based on the weighted sum method; represents the inventory turnover ratio; represents the out-of-stock rate; Indicates the rate of slow-moving goods; Indicates the results of the attention assessment for inventory management;

[0207] 2.4.5. Inventory management utility function layer, used to calculate the inventory management utility value based on order fulfillment rate, inventory cost and out-of-stock loss.

[0208] The expression of inventory management utility value is as follows:

[0209] ;

[0210] ;

[0211] ;

[0212] ;

[0213] In the formula, represents the inventory management utility value; The weight coefficient representing the order fulfillment rate; represents the order fulfillment rate; Indicates the weight coefficient of inventory cost; represents inventory cost; The weight coefficient representing the out-of-stock loss; Indicates loss due to shortage of stock;

[0214] Indicates The order volume of the product; Indicates Available inventory of the product; Indicates Unit inventory holding cost of a product; Indicates The number of times a product is put in and taken out of the warehouse; Indicates The single inbound and outbound cost of a product; Indicates Unit storage cost of the product; Indicates The unit price of the product; Indicates The expected loss from a stock-out of each unit of a product; Indicates the number of product types.

[0215] In order to coordinate and balance the decisions of different agents to maximize the overall utility, a multi-agent game mechanism is introduced for the multi-agent collaborative decision-making of sales forecasting agents, production planning agents, procurement and stocking agents, and inventory management agents.

[0216] Step 3: Iterate and optimize the decision-making plan based on the Nash equilibrium game and Rubinstein bargaining model, and obtain the optimal decision-making plan by calculating the system utility value.

[0217] A set of candidate decision plans from the above four intelligent agents is obtained, and by calculating the system utility value, a consensus is reached on the decision based on the Nash equilibrium game and the Rubinstein bargaining model to obtain the optimal decision plan.

[0218] The expression of system utility value is as follows:

[0219]

[0220] In the formula, represents the system utility value; The weight representing the sales forecast utility value; represents the sales forecast utility value; The weight representing the utility value of the production plan; represents the production plan utility value; The weight representing the utility value of purchasing and stocking; Indicates the purchase and stocking utility value; The weight representing the utility value of inventory management; represents the inventory management utility value;

[0221] represents the weight of decision consistency in system utility; Represents decisions based on different agents and The decision consistency evaluation function, when m or h is s, it represents the sales forecasting agent. or When the value is p, it represents the production planning agent. or When the value is b, it indicates the purchasing and stocking agent. or When the value is d, it indicates an inventory management agent;

[0222] Indicates the weight of system profit in system utility; represents the expected income of the system, Represents the expected cost of the system.

[0223] The expression of the system's expected income is as follows:

[0224] ;

[0225] In the formula, Indicates The unit price of the product; Indicates The product is in The sales decision variables at time step are given by the sales forecasting agent; Indicates In the demand scenario The product is in The actual demand for the time step; Indicates the quantity of the product; Indicates the number of demand scenarios; Indicates the number of time steps.

[0226] The expected cost of the system is expressed as follows:

[0227] ;

[0228] In the formula, Indicates The product is in Unit production cost per time step; Indicates The product is in The production decision variables at the time step are given by the production planning agent;

[0229] Indicates The product is in Unit inventory cost at the time step; Indicates that the i-th product is in The expected inventory at the time step, ;

[0230] Indicates The product is in Unit out-of-stock cost at time step; Indicates The product is in The inventory decision variables at the time step are given by the inventory management agent.

[0231] Step 4: According to the optimal decision plan, adaptive optimization is performed based on the adaptive mechanism to obtain the adjusted decision plans of each intelligent agent to achieve composite scheduling of tobacco enterprises with multi-agent collaborative optimization.

[0232] The input of this step is the optimal decision-making plan from step three, the utility value and state variables of each agent from step two, the expected system income and the expected system cost. The output is the adjusted and optimized agent decision-making strategy and profit sharing plan.

[0233] The expression for adaptive optimization based on the adaptive mechanism is as follows:

[0234] ;

[0235] ;

[0236] ;

[0237] ;

[0238] ;

[0239] In the formula, Indicates that after optimization The state of the agent at the time step; Indicates that after optimization Agent decision at time step; Indicates in status Take decision The expected long-term return;

[0240] Indicated in The state of the agent at the time step; Indicated in Agent decision at time step; Indicates in status Take decision The expected long-term return;

[0241] represents the learning rate; Indicates The system utility value at the time step, which is fed back to the adaptive optimization mechanism as an immediate reward; represents the discount factor; Indicates in status Take decision The expected long-term return;

[0242] represents the profit shared by the sales forecasting agent; represents the sales forecast utility value; represents the profit shared by the production planning agent; represents the production plan utility value; Represents the profit shared by the purchasing and stocking agent; Indicates the utility value of purchasing and stocking; represents the profit shared by the inventory management agent; represents the inventory management utility value; Indicates the total profit of the system.

[0243] In summary, the composite scheduling method for tobacco enterprises based on multi-agent collaborative optimization realizes interactive collaboration among four agents by introducing attention communication mechanism and game theory collaborative mechanism. Agents selectively focus on and transmit key information through the attention mechanism. By dynamically adjusting the attention weight, agents can adapt to different decision-making scenarios and information needs. Through the game module, agents coordinate decisions based on the game theory collaborative mechanism to achieve balance and compromise of the decisions of each agent and promote overall optimization. The present invention can realize dynamic simulation and auxiliary decision-making of business objectives, better sales target management, closer production and marketing connection, more flexible production plan, more refined auxiliary material stocking strategy, and more reasonable specifications and inventory distribution.

[0244] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0245] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0246] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0247] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0248] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A composite scheduling method for tobacco enterprises based on multi-agent collaborative optimization, characterized in that: include: Obtain system data from tobacco companies; According to the system data, multiple corresponding decision-making schemes are obtained based on the sales forecasting agent, the production planning agent, the purchasing and stocking agent, and the inventory management agent; Iteratively optimizing the decision-making scheme based on the Nash equilibrium game and the Rubinstein bargaining model, and obtaining the optimal decision-making scheme by calculating the system utility value; According to the optimal decision-making scheme, adaptive optimization is performed based on the adaptive mechanism to obtain the adjusted decision-making schemes of each intelligent agent, so as to realize the composite scheduling of tobacco enterprises with multi-agent collaborative optimization; Among them, the sales forecasting agent, production planning agent, purchasing and stocking agent, and inventory management agent exchange information through the attention mechanism.

2. The tobacco enterprise composite scheduling method based on multi-agent collaborative optimization according to claim 1 is characterized in that: The system data includes historical sales data, market environment data, seasonal factor data, production status data, inventory status data, raw material market data and supplier data; The decision-making scheme includes sales forecasting decision, production planning decision, purchase and stocking decision and inventory management decision; The sales forecasting agent is used to obtain sales forecasting decisions based on the historical sales data, market environment data and seasonal factor data and interactive information of other agents; The production planning agent is used to obtain the production planning decision based on the sales forecast decision, production status data, inventory status data and interactive information of other agents; The purchasing and stocking agent is used to obtain purchasing and stocking decisions based on the production plan decision, raw material market data, supplier data, and interactive information of other agents; The inventory management agent is used to obtain inventory management decisions based on the sales forecast decision, production plan decision, purchase stocking decision, inventory status data and interactive information of other agents; Among them, the sales forecasting agent adopts an attention-based long short-term memory network algorithm, and the production planning agent, procurement and stocking agent, and inventory management agent adopt an algorithm that combines deep reinforcement learning and multi-objective optimization.

3. The tobacco enterprise composite scheduling method based on multi-agent collaborative optimization according to claim 2 is characterized in that: The sales forecasting agent includes: The sales forecast input encoding layer is used to encode the sales feature input vector to obtain a sales state vector; the sales feature input vector includes historical sales data, market environment data and seasonal factor data; The sales prediction attention layer is used to obtain a sales attention vector by calculating the attention weight according to the sales state vector. A sales prediction decoding layer, used to obtain a sales prediction decision according to the sales attention vector; A sales forecast utility function layer, used to calculate the sales forecast utility value according to the sales forecast decision; The sales forecast utility value is expressed as follows: ； In the formula, represents the sales forecast utility value; represents the sales forecast decision at the tth time step; represents the actual sales quantity at the t-th time step; represents the first weight coefficient; represents the second weight coefficient; Indicates the time span of the forecast.

4. The tobacco enterprise composite scheduling method based on multi-agent collaborative optimization according to claim 2 is characterized in that: The production planning agent comprises: The production plan state encoding layer is used to encode the production feature input vector to obtain the production plan state vector of the current time step; the production feature input vector includes sales forecast decision, production status data and inventory status data; The production plan decision network layer is used to obtain the corresponding production plan decision according to the production plan state vector of the current time step; The production plan state transfer function layer is used to obtain the production plan state vector of the next time step according to the production plan state vector of the current time step and the corresponding production plan decision; The production plan reward function layer is used to calculate the reward value corresponding to the production plan decision at the current time step based on the production-sales connection, production efficiency, inventory rationality and game equilibrium factor; The production plan utility function layer is used to calculate the production plan utility value based on the production plan efficiency, production plan cost, and the degree of deviation of the production plan from the sales forecast and inventory target; The expression of the production plan utility value is as follows: ； ； ； ； In the formula, represents the production plan utility value; The weight coefficient representing the efficiency of production planning; It indicates the efficiency of production planning; The weight coefficient representing the production plan cost; represents the production plan cost; The weight coefficient indicating the degree of deviation; Indicates the degree of deviation of the production plan from the sales forecast and inventory target; Indicates The production quantity of each product; Indicates Production time of the product; Indicates The production capacity occupied by the product; Indicates The unit production cost of a product; Indicates Target inventory level for each product; Indicates The time the equipment is used; Indicates The time unit cost of the equipment; Indicates The amount of energy consumed; Indicates The unit price of the energy source; Indicates the number of product types; Indicates the number of types of equipment; Indicates the number of energy types.

5. The tobacco enterprise composite scheduling method based on multi-agent collaborative optimization according to claim 2 is characterized in that: The purchasing and stocking agent includes: The procurement stocking status encoding layer is used to encode the procurement feature input vector to obtain the procurement stocking status vector of the current time step; the procurement feature input vector includes production plan decision, raw material market data and supplier data; A purchasing and stocking decision network layer is used to obtain a corresponding purchasing and stocking decision according to the purchasing and stocking state vector of the current time step; The procurement stocking state transfer function layer is used to obtain the procurement stocking state vector of the next time step according to the procurement stocking state vector of the current time step and the corresponding procurement stocking decision; The procurement and stocking reward function layer is used to calculate the reward value corresponding to the procurement and stocking decision at the current time step based on procurement cost, inventory rationality, supply stability, and attention evaluation results; The procurement stocking utility function layer is used to calculate the procurement stocking utility value based on procurement availability, procurement stocking cost and inventory pressure; The expression of the purchase stocking utility value is as follows: ； ； ； ; In the formula, Indicates the purchase and stocking utility value; The weight coefficient representing the availability of procurement; Indicates purchasing availability; Indicates the weight coefficient of procurement stocking cost; Represents the purchase and stocking cost; The weight coefficient representing the inventory pressure; Indicates inventory pressure; Indicates The demand for raw materials; Indicates The arrival rate of raw materials; Indicates The purchase volume of raw materials; Indicates Unit price of the raw materials; Indicates Unit transportation cost of the raw materials; Indicates The actual inventory of the raw materials; Indicates Safety stock of raw materials; Indicates the number of types of raw materials.

6. The tobacco enterprise composite scheduling method based on multi-agent collaborative optimization according to claim 2 is characterized in that: The inventory management agent includes: The inventory management state encoding layer is used to encode the inventory feature input vector to obtain the inventory management state vector of the current time step; the inventory feature input vector includes sales forecast decision, production plan decision, purchase stocking decision and inventory status data; An inventory management decision network layer, used to obtain a corresponding inventory management decision according to the inventory management state vector of the current time step; An inventory management state transfer function layer, used to obtain the inventory management state vector of the next time step according to the inventory management state vector of the current time step and the corresponding inventory management decision; The inventory management reward function layer is used to calculate the reward value corresponding to the inventory management decision at the current time step based on the inventory turnover rate, out-of-stock rate, slow-moving product rate, and attention evaluation results; Inventory management utility function layer, used to calculate the inventory management utility value based on order fulfillment rate, inventory cost and out-of-stock loss; The expression of the inventory management utility value is as follows: ； ； ； ； In the formula, represents the inventory management utility value; The weight coefficient representing the order fulfillment rate; represents the order fulfillment rate; Indicates the weight coefficient of inventory cost; represents inventory cost; The weight coefficient representing the out-of-stock loss; Indicates loss due to shortage of stock; Indicates The order volume of the product; Indicates The available inventory of the product; Indicates Unit inventory holding cost of a product; Indicates The number of times a product is put in and taken out of the warehouse; Indicates The single inbound and outbound cost of a product; Indicates Unit storage cost of the product; Indicates The unit price of the product; Indicates The expected loss from a stock-out of each unit of a product; Indicates the number of product types.

7. The tobacco enterprise composite scheduling method based on multi-agent collaborative optimization according to claim 1 is characterized in that: The expression of the system utility value is as follows: ； In the formula, represents the system utility value; The weight representing the sales forecast utility value; represents the sales forecast utility value; The weight representing the utility value of the production plan; represents the production plan utility value; The weight representing the utility value of purchasing and stocking; Indicates the utility value of purchasing and stocking; The weight representing the utility value of inventory management; represents the inventory management utility value; represents the weight of decision consistency in system utility; Represents decisions based on different agents and The decision consistency evaluation function, when m or h is s, it represents the sales forecasting agent. or When the value is p, it represents the production planning agent. or When the value is b, it indicates the purchasing and stocking agent. or When the value is d, it indicates an inventory management agent; Indicates the weight of system profit in system utility; represents the expected income of the system, Represents the expected cost of the system.

8. The tobacco enterprise composite scheduling method based on multi-agent collaborative optimization according to claim 7 is characterized in that: The expression of the expected income of the system is as follows: ； In the formula, Indicates The unit price of the product; Indicates The product is in The sales decision variables at time step are given by the sales forecasting agent; Indicates In the demand scenario The product is in The actual demand for the time step; Indicates the quantity of the product; Indicates the number of demand scenarios; Indicates the number of time steps.

9. The tobacco enterprise composite scheduling method based on multi-agent collaborative optimization according to claim 8 is characterized in that: The expression of the expected cost of the system is as follows: ； In the formula, Indicates The product is in Unit production cost per time step; Indicates The product is in The production decision variables at the time step are given by the production planning agent; Indicates The product is in Unit inventory cost at the time step; Indicates that the i-th product is in The expected inventory at time step, ; Indicates The product is in Unit out-of-stock cost at time step; Indicates The product is in The inventory decision variables at the time step are given by the inventory management agent.

10. The tobacco enterprise composite scheduling method based on multi-agent collaborative optimization according to claim 1 is characterized in that: The expression for adaptive optimization based on the adaptive mechanism is as follows: ； ； ； ； ； In the formula, Indicates that after optimization The state of the agent at the time step; Indicates that after optimization Agent decision at time step; Indicates in status Take decision The expected long-term return; Indicated in The state of the agent at the time step; Indicated in Agent decision at time step; Indicates in status Take decision The expected long-term return; represents the learning rate; Indicates System utility value at the time step; represents the discount factor; Indicates in status Take decision The expected long-term return; represents the profit shared by the sales forecasting agent; represents the sales forecast utility value; represents the profit shared by the production planning agent; represents the production plan utility value; Represents the profit shared by the purchasing and stocking agent; Indicates the purchase and stocking utility value; represents the profit shared by the inventory management agent; represents the inventory management utility value; Indicates the total profit of the system.

Images