Marketing scenario-oriented ai agent task scheduling execution method

Abstract

The application discloses a marketing-scene-oriented AI intelligent agent task scheduling execution method, belongs to the technical field of artificial intelligence and digital marketing, and comprises the following steps: S1, collecting multi-dimensional index data and processing; S2, establishing a task priority evaluation model, calculating the comprehensive priority score of a task according to the processed multi-dimensional index data and sorting; S3, constructing a simulated marketing environment, training a reinforcement learning model to obtain an optimized scheduling strategy; S4, modeling a marketing task, an AI intelligent agent and audience elements as nodes and establishing associated edges, training a graph neural network model to predict task execution time and effectiveness; and S5, performing task allocation according to the above content and performing real-time monitoring and adjustment. The marketing-scene-oriented AI intelligent agent task scheduling execution method can make multiple AI intelligent agents work cooperatively in a marketing scene through an intelligent scheduling algorithm, and can quickly and efficiently complete various marketing tasks.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and digital marketing technology, and in particular to an AI intelligent agent task scheduling and execution method for marketing scenarios. Background Technology

[0002] In modern marketing campaigns, the collaborative participation of multiple AI agents (such as chatbots, recommendation engines, and content generation models) has become commonplace. However, existing systems suffer from numerous problems in task allocation, exhibiting low efficiency and slow response times. Often, some agents are overloaded while others remain idle, leading to unreasonable resource allocation. This imbalance not only reduces the timeliness of marketing responses and impacts user experience but also results in a significant waste of computing resources, hindering the efficient execution of enterprise digital marketing campaigns and failing to meet companies' pursuit of high ROI and operational efficiency. Summary of the Invention

[0003] The purpose of this invention is to provide an AI agent task scheduling and execution method for marketing scenarios, in order to solve the problems of low task allocation efficiency, slow response and unreasonable resource allocation in existing marketing AI systems, to achieve efficient collaboration of multiple AI agents in marketing scenarios, to ensure that important marketing tasks are processed in a timely manner, and to avoid idle or overloaded resources.

[0004] To achieve the above objectives, this invention provides an AI agent task scheduling and execution method for marketing scenarios, comprising the following steps:

[0005] S1. Collect and process multi-dimensional indicator data, including user profile data, content priority data, and resource status data;

[0006] S2. Establish a task priority evaluation model, calculate the comprehensive priority score of tasks based on the processed multi-dimensional indicator data, and rank them.

[0007] S3. Construct a simulated marketing environment, determine the state space, action space, and reward signal, and train the reinforcement learning model to obtain an optimized scheduling strategy.

[0008] S4. Model marketing tasks, AI agents, and audience elements as nodes and establish related edges to train a graph neural network model to predict task execution time and effectiveness.

[0009] S5. Assign tasks based on task priority, reinforcement learning scheduling strategy, and graph neural network prediction results, and monitor and adjust them in real time.

[0010] Preferably, the user profile data in step S1 includes collecting users' browsing history, purchasing behavior, search keywords, and dwell time.

[0011] Preferably, step S2 specifically includes:

[0012] S21. The task priority evaluation module uses the user profile data, content priority data and resource status data from step S1 as the core input dimensions.

[0013] S22. Quantify the input indicators of each dimension;

[0014] S23. Assign weights to each dimension indicator based on the characteristics of the marketing scenario;

[0015] S24. Calculate the overall priority score of the task using a weighted summation formula:

[0016] ;

[0017] in, Standardized values ​​for user profile data; A quantified value for content priority data; Standardized values ​​for resource status data; , , All are weighting coefficients, and ;

[0018] S25. Sort all tasks according to their comprehensive priority scores, and the task with the highest score will be given priority to enter the scheduling queue corresponding to the reinforcement learning scheduling strategy training stage in S3.

[0019] Preferably, step S3 specifically includes:

[0020] S31. Construct a simulated marketing environment. Using the task priority ranking result output in step S2 as input, construct a simulated environment that includes task characteristics, AI agent status, and resource constraints. The environment dynamically generates task flow. Task attributes include user value coefficient, content urgency, and resource requirements. Simulate user interaction feedback.

[0021] S32. Determine the core elements of a reinforcement learning model: state space, action space, and reward function;

[0022] S33. A deep Q-network architecture and related training mechanisms are used to train the reinforcement learning model so that it can select the optimal task allocation action strategy under different states in marketing scenario task scheduling.

[0023] S34. After training convergence, the task queue sorted in step S2 is processed to output the optimal action based on the state, forming a dynamic scheduling strategy to ensure that high-priority tasks are assigned to the adapted agent first.

[0024] Preferably, step S32 specifically includes:

[0025] State space definition: The result of merging the task priority ranking in step S2. Real-time load of intelligent agents Resource utilization rate and task context vector , forming the state vector ;

[0026] Action space definition: Actions represent decisions that assign tasks to agents. The discrete space is defined as follows: ,in, This represents the sorted task index in step S2; Indicates the available agent index;

[0027] Set the reward function: ;

[0028] in, For the task In intelligent agents Execution time on , , As weight, Penalty for resource load:

[0029] ;

[0030] in, This is the resource load threshold, used to determine whether the agent is overloaded.

[0031] Preferably, step S33 specifically includes:

[0032] A deep Q-network architecture is adopted, and experience replay mechanism is used for storage. Sample, update Q value using target network:

[0033] ;

[0034] in, For learning rate, As a discount factor, Main network parameters, For the target network parameters, the loss function uses mean squared error: Parameters are optimized using gradient descent, and the target network parameters are synchronized every N steps. - A greedy strategy balances exploration and exploitation, specifically:

[0035] Initialize exploration probability: Set the initial exploration probability. and final exploration probability and exploration probability decay rate ;

[0036] Select action: in each state Next, generate a random number between [0,1]. ,like Then randomly select an action. Uniform sampling from the action space is used to explore new task allocation methods; if Then choose the action that maximizes the Q-value predicted by the current Q-network. Utilize the strategies already learned;

[0037] Update the exploration probability: After each training iteration, update the exploration probability according to the following formula. :

[0038] ;

[0039] As training progresses, This reduces the likelihood of the agent utilizing the optimal strategy it has already learned, thus reducing the exploration of new actions.

[0040] Preferably, after training convergence in step S34, the task queue sorted in step S2 is processed according to its state. Output optimal action This forms a dynamic scheduling strategy to ensure that high-priority tasks are assigned to the appropriate intelligent agents first.

[0041] Preferably, step S4 specifically includes:

[0042] S41. Integrate nodes and edges into a heterogeneous graph. ,in For the set of all nodes, For the set of all edges;

[0043] S42. A graph neural network is used for node feature aggregation. The feature update formula for each layer is:

[0044] ;

[0045] in, For nodes In the Features of the layer For nodes The neighborhood group, For nodes and its neighboring nodes The weight of the edges between them. For node degree, and For learnable parameters, For activation functions;

[0046] S43. Predict task execution time and results:

[0047] For task-agent pairs Execution time is predicted by the output layer of a graph neural network. The loss function is the mean squared error:

[0048] ;

[0049] in, This refers to the actual execution time. The number of samples;

[0050] For the task-agent-audience triplet Predicting task conversion rate The loss function is cross-entropy:

[0051] ;

[0052] in, This represents the actual conversion rate.

[0053] Preferably, step S5 constructs an allocation scoring function based on task priority ranking, reinforcement learning scheduling strategy, and graph neural network prediction results to select the optimal agent, specifically as follows:

[0054] For the task and intelligent agents ,score for:

[0055] ;

[0056] in, Assign a score based on the overall priority of the task; To enhance the value of the actions output by the learning model; , , These are the weighting coefficients;

[0057] For the sorted task queue, select the agent with the highest score: ,in, This is the set of available intelligent agents.

[0058] Therefore, this invention employs the aforementioned AI agent task scheduling and execution method for marketing scenarios. Through multi-dimensional indicator-driven task priority evaluation, reinforcement learning-optimized adaptive scheduling strategies, and graph neural network-supported intelligent decision-making, it effectively solves the problems of low task allocation efficiency, slow response, and unreasonable resource allocation in existing marketing AI systems. It can intelligently coordinate the efficient operation of multiple AI agents based on real-time marketing environments and business needs, ensuring timely processing of important marketing tasks while avoiding resource idleness or overload. In practical applications, it can significantly improve the speed and accuracy of marketing response, enhance user experience, and increase the utilization rate of marketing resources, bringing higher ROI and operational efficiency to enterprises' digital marketing activities.

[0059] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0060] Figure 1 This is a flowchart of the AI ​​intelligent agent task scheduling and execution method for marketing scenarios according to the present invention.

[0061] Figure 2 This is a flowchart illustrating the process of calculating and prioritizing computational tasks according to an embodiment of the present invention.

[0062] Figure 3 A flowchart illustrating the optimized scheduling strategy in an embodiment of the present invention;

[0063] Figure 4 This is a flowchart illustrating the task execution time and performance prediction in an embodiment of the present invention. Detailed Implementation

[0064] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0065] Please see Figures 1-4 A task scheduling and execution method for AI intelligent agents in marketing scenarios includes the following steps:

[0066] S1. Collect and process multi-dimensional indicator data, including user profile data, content priority data, and resource status data. User profile data is collected through the company's existing user behavior tracking system, gathering data such as user browsing history, purchase behavior, search keywords, and dwell time on platforms like websites and apps. This data is cleaned to remove invalid information (such as duplicate records and incorrectly formatted data). Then, feature extraction is performed, transforming user behavior into quantifiable features, such as user spending power (judged based on historical purchase amounts), interests and preferences (determined based on browsing and search content), and activity level (calculated based on login frequency and dwell time), building a complete user profile database. Content priority data is determined by marketers based on the goals and strategies of marketing activities, setting evaluation criteria for content priority. For example, for promotional content, priority is determined based on the urgency of the activity (remaining time before the activity ends) and expected conversion rate; for regular marketing content, priority is determined based on its importance to brand promotion and user retention. These evaluation criteria are transformed into specific quantitative indicators. The system scores marketing content to determine the priority of each content task. Monitoring programs are deployed on each AI agent and the servers running those agents to collect real-time data on the number of tasks being processed, task progress, average processing time, and server resource status data such as CPU utilization, GPU utilization, memory usage, and network bandwidth usage. The collected data is transmitted in real-time to a central data processing center for aggregation and formatting for subsequent analysis and use.

[0067] S2. Establish a task priority evaluation model, calculate and rank the comprehensive priority scores of tasks based on the processed multi-dimensional indicator data, specifically including:

[0068] S21. The task priority evaluation module uses the user profile data, content priority data and resource status data from step S1 as the core input dimensions.

[0069] S22. Quantify the input indicators of each dimension;

[0070] For user profile data, user characteristics are converted into a 0-1 standardized user value coefficient (e.g., the coefficient for high-value users is 0.8-1.0).

[0071] The content priority data directly adopts the priority value determined in step S1 (such as a 1-10 scale).

[0072] Resource status data is quantified by calculating a resource matching index (such as the reciprocal of the agent's current load rate).

[0073] S23. Assign weights to each dimension indicator based on the characteristics of the marketing scenario. The weights can be dynamically adjusted according to business objectives.

[0074] S24. Calculate the overall priority score of the task using a weighted summation formula:

[0075] ;

[0076] in, Standardized values ​​for user profile data; A quantified value for content priority data; Standardized values ​​for resource status data; , , All are weighting coefficients, and ;

[0077] S25. Sort all tasks according to their comprehensive priority scores, and the task with the highest score will be given priority to enter the scheduling queue corresponding to the reinforcement learning scheduling strategy training stage in S3.

[0078] S3. Construct a simulated marketing environment, determine the state space, action space, and reward signal, and train the reinforcement learning model to obtain an optimized scheduling strategy, specifically including:

[0079] S31. Construct a simulated marketing environment. Using the task comprehensive priority ranking result output in step S2 as input, construct a simulated environment that includes task characteristics (such as user profile tags and content priority scores), AI agent status (such as current load and historical processing success rate), and resource constraints (such as server CPU / GPU utilization threshold). The environment dynamically generates task flows. Task attributes include user value coefficient (from user profile data in S2), content urgency (from priority scores in S2), and resource requirements (such as estimated computation time). Simulate user interaction feedback.

[0080] S32. Determine the core elements of the reinforcement learning model: state space, action space, and reward function, specifically:

[0081] State space definition: The result of merging the task priority ranking in step S2. Real-time load of intelligent agents Resource utilization rate and task context vector , forming the state vector ;

[0082] Action space definition: Actions represent decisions that assign tasks to agents. The discrete space is defined as follows: , indicating that the first The task is assigned to the first There are 10 intelligent agents, among which... This represents the sorted task index in step S2; Indicates the available agent index;

[0083] Set the reward function: ;

[0084] in, For the task In intelligent agents Execution time on , , As weight, Penalty for resource load:

[0085] ;

[0086] in, This is a resource load threshold used to determine if the agent is overloaded; when the agent's actual load... Exceeding the threshold will trigger a resource load penalty, prompting the model to reduce the allocation of tasks to the overloaded agent; if If the threshold is ≤threshold, the penalty is 0, allowing normal task allocation to balance the agent's load and optimize resource utilization.

[0087] S33. A deep Q-network architecture and related training mechanisms are used to train the reinforcement learning model, enabling it to select the optimal task allocation strategy under different states in marketing scenario task scheduling. Specifically:

[0088] A deep Q-network architecture is adopted, and experience replay mechanism is used for storage. Sample, update Q value using target network:

[0089] ;

[0090] in, For learning rate, As a discount factor, Main network parameters, For the target network parameters, the loss function uses mean squared error: Parameters are optimized using gradient descent, and the target network parameters are synchronized every N steps. - A greedy strategy balances exploration and exploitation, specifically:

[0091] Initialize exploration probability: Set the initial exploration probability. and final exploration probability and exploration probability decay rate ;

[0092] Select action: in each state Next, generate a random number between [0,1]. ,like Then randomly select an action. Uniform sampling from the action space is used to explore new task allocation methods; if Then choose the action that maximizes the Q-value predicted by the current Q-network. Utilize the strategies already learned;

[0093] Update the exploration probability: After each training iteration, update the exploration probability according to the following formula. :

[0094] ;

[0095] As training progresses, The gradual reduction in size causes the agent to increasingly utilize the optimal strategies it has learned, while reducing the exploration of new actions.

[0096] S34. After training convergence, sort the task queue from step S2 according to its state. Output optimal action This forms a dynamic scheduling strategy to ensure that high-priority tasks are assigned to the appropriate intelligent agents first.

[0097] S4. Model marketing tasks, AI agents, and audience elements as nodes and establish related edges. Train a graph neural network model to predict task execution time and effectiveness. Specifically, this includes:

[0098] S41. Integrate nodes and edges into a heterogeneous graph. ,in For the set of all nodes, For the set of all edges;

[0099] S42. A graph neural network is used for node feature aggregation. The feature update formula for each layer is:

[0100] ;

[0101] in, For nodes In the Features of the layer For nodes The neighborhood group, For nodes and its neighboring nodes The weight of the edges between them. For node degree, and For learnable parameters, For activation functions;

[0102] S43. Predict task execution time and results:

[0103] For task-agent pairs Execution time is predicted by the output layer of a graph neural network. The loss function is the mean squared error:

[0104] ;

[0105] in, This refers to the actual execution time. The number of samples;

[0106] For the task-agent-audience triplet Predicting task conversion rate The loss function is cross-entropy:

[0107] ;

[0108] in, This represents the actual conversion rate.

[0109] S5. Tasks are allocated based on task priority, reinforcement learning scheduling strategies, and graph neural network prediction results, and are monitored and adjusted in real time, specifically as follows:

[0110] For the task and intelligent agents ,score for:

[0111] ;

[0112] in, Assign a score based on the overall priority of the task; To enhance the value of the actions output by the learning model; , , These are the weighting coefficients;

[0113] For the sorted task queue, select the agent with the highest score: ,in, The set of available intelligent agents (resource constraints must be met: , For the task Increased load rate).

[0114] Real-time monitoring and adjustment:

[0115] Monitoring indicators include:

[0116] Deviation between actual execution time and graph neural network prediction: ;

[0117] Real-time load rate of the agent: (Updated every 5 seconds);

[0118] Task timeout rate: Number of timed-out tasks / Total number of tasks.

[0119] Dynamic adjustment strategy:

[0120] like (If the deviation exceeds 20%), incremental training of the graph neural network model is triggered (updating the parameters of the last layer with new data).

[0121] like (Overload) The agent's unfinished tasks are reallocated according to S3's reinforcement learning strategy: in, For load rate A collection of intelligent agents.

[0122] Therefore, this invention adopts the above-mentioned AI intelligent agent task scheduling and execution method for marketing scenarios. Through multi-dimensional indicator-driven task priority evaluation, it can comprehensively and accurately judge the importance and urgency of tasks, ensuring that important tasks are given priority. The adaptive scheduling strategy optimized by reinforcement learning enables task allocation to be adjusted in real time according to environmental changes, improving the efficiency of resource utilization under complex marketing activities. The intelligent decision-making supported by graph neural networks can capture the potential correlation patterns between tasks and between tasks and intelligent agents, achieving more accurate task allocation.

[0123] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for AI agent task scheduling and execution for marketing scenarios, characterized in that, The method comprises the following steps: S1, collecting multi-dimensional index data and processing, the multi-dimensional index data including user portrait data, content priority data and resource state data; S2, establishing a task priority evaluation model, calculating the comprehensive priority score of the task according to the processed multi-dimensional index data and sorting; S3, constructing a simulated marketing environment, determining the state space, action space and reward signal, and training the reinforcement learning model to obtain an optimized scheduling strategy; S4, modeling the marketing task, AI agent and audience elements as nodes and establishing associated edges, and training a graph neural network model to predict the task execution time and effectiveness; S5, according to the task priority ranking, reinforcement learning scheduling strategy and graph neural network prediction result, constructing an allocation score function, selecting the optimal agent, and specifically: Step S3 specifically comprises: S31, constructing a simulated marketing environment, taking the task comprehensive priority ranking result output in step S2 as input, constructing a simulated environment containing task features, AI agent state and resource constraints, dynamically generating a task flow, and simulating user interaction feedback; S32, determining the core elements of the reinforcement learning model: state space, action space and reward function; S33, using a deep Q network architecture and related training mechanism to train the reinforcement learning model, so that it selects the optimal task allocation action strategy in different states in the marketing scenario task scheduling; S34, after training convergence, output the optimal action according to the state for the task queue sorted in step S2, form a dynamic scheduling strategy, and ensure that high-priority tasks are allocated to adaptive agents first; Step S32 specifically comprises: State space definition: task priority ordering result of fusion step S2 , agent real-time load , resource usage rate and task context vector , constitute state vector ; Action space definition: An action represents a decision of assigning a task to an agent, the discrete space is defined as: where, denotes the sorted task index in step S2; denotes the available agent index; Set the reward function: ; wherein, is a task is an execution time, on an agent , , is a weight, is a resource load penalty: ； wherein, is a resource load threshold value for determining whether the agent is overloaded.

2. The marketing-scene-oriented AI agent task scheduling execution method according to claim 1, characterized in that: The user portrait data in step S1 includes collecting the user's browsing records, purchase behavior, search keywords and dwell time.

3. The marketing scenario-oriented AI agent task scheduling execution method according to claim 1, characterized in that, Step S2 specifically comprises: S21, the task priority evaluation module takes the user portrait data, content priority data and resource state data in step S1 as the core dimensions of the input; S22, quantitatively processing the input index of each dimension; S23, according to the characteristics of the marketing scenario, assigning weights to each dimension index; S24, calculating the task comprehensive priority score using the weighted sum formula: ； wherein, is a standardized value of the user profile data; is a quantized value of the content priority data; is a standardized value of the resource status data; , , are all weight coefficients, and + + ; S25, according to the comprehensive priority score, all tasks are sorted, and the task with the highest score is preferentially entered into the scheduling queue corresponding to the reinforcement learning scheduling strategy training link in S3.

4. The marketing scenario-oriented AI agent task scheduling and execution method according to claim 1, characterized in that, Step S33 specifically comprises: Using a deep Q-network architecture, storing samples through an experience replay mechanism and updating Q-values using a target network ； wherein, is the learning rate, is the discount factor, is the main network parameter, is the target network parameter, and the loss function adopts mean square error: Optimize the parameters by gradient descent, synchronize the target network parameters every N steps, and use - Greedy strategy balances exploration and utilization, specifically: initial exploration probability: set initial exploration probability and final exploration probability , and exploration probability decay rate ; Select action: in each state Next, generate a random number between [0, 1] If , randomly select an action , uniformly sample from the action space to explore new task allocation ways; if , select the action that maximizes the Q value predicted by the current Q network, , use the learned strategy; Update the exploration probability: After each training iteration, update the exploration probability according to the following formula : ； As the training progresses, Decreases, so that the agent exploits the optimal policy it has learned while reducing exploration of new actions.

5. The marketing-scene-oriented AI agent task scheduling and execution method according to claim 4, characterized in that: After the training converges, the task queue sorted in step S2 is sorted according to the state Output the optimal action , forming a dynamic scheduling strategy to ensure that high-priority tasks are assigned to the adaptive agent first.

6. The marketing scenario-oriented AI agent task scheduling execution method according to claim 5, characterized in that, Step S4 specifically comprises: S41, integrate the nodes and edges into a heterogeneous graph wherein for all node sets, for all edge sets; S42, using a graph neural network to aggregate node features, and the feature update formula of each layer is: ； wherein, is a node In a first characteristic of a layer, is a node is a neighbor set of a node, is a weight of an edge between a node and its neighbor nodes is a node degree, is a node degree, and is a learnable parameter, is an activation function; S43, predicting the task execution time and effectiveness: For task-agent pairs , predict execution time through graph neural network output layer , loss function is mean square error: ； wherein, is the actual execution time, is the number of samples; For task-agent-audience triplets Predicting task conversion rates The loss function is cross-entropy: ； wherein, is the actual conversion rate.

7. The marketing-scene-oriented AI agent task scheduling execution method according to claim 6, characterized in that, Step S5 according to the task priority ranking, reinforcement learning scheduling strategy and graph neural network prediction result, constructs an allocation score function, selects the optimal agent, and specifically: To the task And the agent , score Is: ； wherein, is a task synthesis priority score; is an action value output by the reinforcement learning model; , , is a weight coefficient; The sorted task queue is selected in this way to select the highest scoring agent: wherein, is the set of available agents.

Images