Multi-agent game driven advertisement material collaboration method and system

Abstract

The application discloses a kind of multi-agent game driven advertising material cooperation method and system, belong to artificial intelligence technical field.The method includes: S1.advertising material scene hypergraph representation and initial state coding;S2.run multi-agent network in parallel, output local proposal and adaptive fusion, generate global state update input;S3.based on agent utility function, agent conflict potential function and constraint violation function, form unified constraint coupling potential game target;S4.based on constraint coupling potential game target, using original-dual iteration strategy to solve equilibrium material and carry out material update;S5.scene level approximate Shapley credit distribution is carried out, the contribution of each agent and narrative unit is quantified;S6.real delivery feedback and closed loop update agent parameter.The application generates advertising material by multi-agent game cooperation, can consider narrative, brand, conversion, compliance and novelty and realize closed loop optimization.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a multi-agent game-driven method and system for collaborative advertising materials. Background Technology

[0002] In short video ads, feed ads, and short drama ads, the quality of creative materials directly determines click-through rate, conversion rate, sales revenue, and overall campaign efficiency. Existing ad creative generation solutions typically use a single large model to directly generate scripts, or break down the "script generation—brand review—conversion rewriting—compliance verification" process into a sequential workflow. While these methods can shorten the manual production cycle to some extent, their underlying structure is still a sequential module assembly: the previous module provides the result, and the next module then makes adjustments. Therefore, this approach is prone to problems such as local optima, repeated rewriting, and the accumulation of conflicting objectives.

[0003] Specifically, ad creative generation faces at least five mutually constraining optimization objectives simultaneously: First, the script needs narrative coherence and emotional progression; second, the creative must consistently present the brand's tone, core selling points, and persona; third, the creative should maximize click-through rate, conversion rate, and sales performance; fourth, the creative must not violate platform rules, industry regulations, or brand compliance red lines; and fifth, the creative needs to maintain novelty to avoid template-based fatigue and attenuation. Current technologies often only address these contradictions by manually setting fixed weights or prompt word order, failing to establish a calculable, collaborative equilibrium mechanism.

[0004] Furthermore, the single-model direct generation method suffers from two key underlying flaws. First, multiple optimization objectives exhibit a strongly coupled, non-convex relationship within the same solution space. For example, enhancing novelty often disrupts brand consistency, strengthening conversion selling points may compromise narrative naturalness, and compliance constraints create hard boundaries that truncate the plot and expression. Single models cannot explicitly separate the marginal effects of each objective during training or inference, leading to the implicit amplification of one objective while systematically suppressing others. Second, real-world delivery feedback only reflects the overall performance of the material and cannot directly explain "which proxy objective, which narrative unit, or which modification action" caused the change in clicks or conversions. Therefore, existing methods struggle to achieve accurate attribution of responsibility and stable updates.

[0005] Therefore, a new advertising creative generation technology is needed that no longer relies on simple process connections, but instead models multiple objectives as a multi-agent game problem that can be coordinated, adversarial, and achieve equilibrium. At the same time, it should combine real-world delivery feedback to establish a scenario-level credit allocation mechanism, so that creative generation can shift from "post-processing manual repair" to "pre-processing algorithmic equilibrium solution". Summary of the Invention

[0006] The present invention aims to at least partially solve one of the technical problems in the aforementioned related technologies.

[0007] Therefore, the purpose of this invention is to provide a multi-agent game-driven advertising material collaboration method and system, which uses multi-agent game collaboration to generate advertising materials, and can take into account narrative, brand, conversion, compliance and novelty, and achieve closed-loop optimization.

[0008] To solve the above-mentioned technical problems, the present invention is implemented as follows: This invention provides a multi-agent game-driven collaborative advertising creative method, the method comprising: S101. Scene hypergraph representation and initial state encoding of advertising creatives; S102. Run the multi-agent network in parallel, output local proposals and adaptively fuse them to generate global state update inputs; S103. Based on the agent utility function, agent conflict potential function, and constraint default function, a unified constraint coupled potential game objective is formed; S104. Based on the objective of the constrained coupled potential game, the primal-dual iterative strategy is used to solve for the equilibrium materials and update the materials. S105. Perform scene-level approximate Shapley credit assignment to quantify the contributions of each agent and narrative unit; S106. Real-time delivery feedback and closed-loop update of agent parameters.

[0009] In addition, the multi-agent game-driven advertising material collaboration method according to the present invention may also have the following additional technical features: In some implementations, step S101 includes: Segmentation: Divide the material into T narrative units, including shots, dialogue, and selling points; Modeling: Construct a scene hypergraph G=(V,E,H), where V is a narrative unit node, E is a temporal edge, and H is a cross-unit semantic hyperedge; each narrative unit node corresponds to a narrative vector, brand vector, selling point vector, and risk vector; Encoding: Each narrative unit generates narrative, brand, selling point, and risk vectors to obtain an initial global state vector, while retaining the initial anchor vector to control material offset.

[0010] In some of these implementations, the multi-agent network described in step S102 includes at least a script agent, a brand agent, a conversion agent, a compliance agent, and a novelty agent.

[0011] In some implementations, in step S102, each agent outputs a local proposal vector and a local score at the t-th narrative unit, and fuses them through adaptive gating coefficients to generate and update the global state; The adaptive gating coefficients are determined based on agent utility, prediction uncertainty, and gradient sensitivity.

[0012] In some implementations, the unified constraint coupling potential game objective in step S103 is to maximize the weighted agent utility and scenario-level conversion benefits, while minimizing the inter-agent conflict potential and default penalty term, and includes a stability anchor term relative to the initial material.

[0013] In some of these implementations, the agent utility function includes a script agent utility function, a brand agent utility function, a conversion agent utility function, and a novelty agent utility function. The proxy conflict potential function is jointly calculated based on the angle difference, semantic edit distance and rhythm offset of the two agents' proposal vectors for the same narrative unit, and is used to suppress the mutual destruction between script, brand, conversion and compliance suggestions; The constraint default function is a hard constraint function based on the overall compliance risk probability, tolerable risk threshold, brand deviation probability, brand deviation threshold, total duration or total number of words of the material, and upper limit constraint, which imposes penalties when a default occurs.

[0014] In some of these implementations, the script agent utility function is calculated jointly based on cosine similarity, which measures the coherence of adjacent narrative units, hook strength scores of narrative units, and hook enhancement coefficients. The brand intelligent agent utility function is calculated jointly based on the brand target vector, the output of the brand consistency prediction head, the ideal brand caliber distribution set by the brand owner, and the brand deviation penalty coefficient. The utility function of the transformation agent is calculated jointly based on the conditional contribution probability of the narrative unit to the transformation, the time decay weight, and the temporal decay factor. The novel agent utility function is calculated jointly based on a kernel matrix, an identity matrix, and a regularization constant, which are composed of narrative unit similarity.

[0015] In some implementations, the process of solving for the equilibrium material using the primal-dual iterative strategy in step S104 includes: Original variable update: Iteratively optimize the narrative unit vector, project it onto the feasible region, and adjust the update scale using the precondition matrix; Dual variable update: Synchronously update constraint multipliers and proxy conflict multipliers, and strengthen penalties when violations occur; Dynamic adjustment of game temperature parameters: The system is adaptively adjusted according to the intensity of conflict and the degree of breach of contract, so that the system is more inclined to sharp choices when the conflict is high and more encouraged to merge multiple agents when the conflict is low. Convergence: When the iteration limit is reached or the convergence condition is met, output the balanced material solution.

[0016] In some implementations, step S105 includes: Marginal contribution: The Shapley value of each agent's contribution to the material is approximated by random permutation.

[0017] Unit-level credit: Refines agent contributions to each narrative unit, linking attention weights and gradient contributions.

[0018] Parameter update: The agent weights and learning rate are adaptively adjusted based on the credit value, so that agents with higher marginal contributions can receive a larger update range.

[0019] This invention also provides a multi-agent game-driven advertising creative collaboration system, capable of implementing the multi-agent game-driven advertising creative collaboration method described in any of the above embodiments; the system includes: Scene hypergraph construction module: It is configured to complete the structured modeling and initial vector encoding of advertising materials, providing the basic state for subsequent game; Multi-agent proposal module: configured to run multiple types of agents in parallel, output local proposals and adaptively fuse them to generate global state update input; The potential game solving module is configured to solve equilibrium materials through primal-dual iteration based on constraint-coupled potential game objectives, satisfying multi-objective constraints. Credit allocation module: It is configured to quantify the contributions of each agent and narrative unit, providing a basis for adaptive parameter updates; Feedback learning module: It is configured to receive real-world delivery data, generate valuable signals, update the agent parameters in a closed loop, and improve the quality of generated materials.

[0020] Compared with the prior art, the present invention has at least the following beneficial effects: In this embodiment of the invention, the multi-agent game-driven collaborative advertising material method improves the advertising material generation problem from a single-model or serial revision problem to a computable multi-agent constrained coupled potential game problem. It can simultaneously handle conflicting objectives such as narrative, brand, conversion, compliance and novelty within the same optimization framework, avoiding post-rework modifications. In this embodiment of the invention, the multi-agent game-driven advertising material collaboration method proposes a local proposal-global gating fusion mechanism. Different agents do not sequentially overwrite each other's results, but instead form parallel proposals and interaction kernel fusion at the same narrative unit, which significantly reduces information loss and target oscillation. In this embodiment of the invention, the multi-agent game-driven advertising material collaboration method proposes a primitive-dual equilibrium update algorithm, which allows constraints such as brand safety, compliance red lines, and shot duration to directly enter the generation process as dual variables, avoiding the large amount of invalid generation caused by the existing technology that can only filter after the fact. In this embodiment of the invention, the multi-agent game-driven advertising material collaboration method proposes a scene-level approximate Shapley credit allocation algorithm, which can decompose the overall delivery feedback into specific agents and specific narrative units, thereby achieving refined responsibility attribution and improving continuous learning efficiency. In this embodiment of the invention, the multi-agent game-driven collaborative advertising material method can directly target the real business data loop of advertising companies. It introduces multi-objective feedback such as exposure, click-through rate, conversion rate, transaction amount and cost in the material generation stage, which can more stably generate advertising materials that can tell a story, sell goods and meet brand and compliance requirements.

[0021] The multi-agent game-driven advertising material collaboration system of the present invention can realize the multi-agent game-driven advertising material collaboration method described above, and therefore has at least all the features and advantages of the multi-agent game-driven advertising material collaboration method described above, which will not be repeated here. Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0022] Figure 1 This is a diagram illustrating the overall framework of a multi-agent game-driven collaborative advertising creative method disclosed in one embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings and specific examples and application scenarios.

[0025] In some embodiments of the present invention, a multi-agent game-driven collaborative advertising creative method is provided. By constructing a scene hypergraph representation of advertising creatives, defining multi-agent utility and conflict potential functions, establishing a primitive-dual equilibrium solution mechanism, and introducing a credit allocation algorithm based on approximate Shapley values, the advertising creatives can obtain a computable equilibrium solution among narrative quality, brand consistency, conversion orientation, compliance and security, and novelty, thereby significantly improving the quality of advertising creative generation and delivery performance.

[0026] Please see Figure 1 As shown, in some embodiments of the present invention, the steps of the multi-agent game-driven advertising creative collaboration method include: Step 1: Scene Hypergraph Representation and Initial State Encoding of Ad Creatives The ad creative to be generated or optimized is divided into T narrative units, denoted as X={x1,x2,...,x...}. T Each narrative unit can correspond to a shot, a dialogue segment, a transition sequence, or a segment expressing a selling point. To simultaneously represent chronological order, cross-shot semantic connections, and multi-character relationships, this invention represents the material as a scene hypergraph G=(V,E,H). Where V={v1,v2,...,v...} T} represents the set of nodes; E represents the set of time transition edges between adjacent narrative units; H represents the set of semantic superedges across multiple narrative units, used to depict foreshadowing, plot twists, selling point recall, and brand messaging consistency.

[0027] For the t-th narrative unit, its basic representation vector h(t) is constructed. The dimension d is a unified narrative representation dimension, preferably ranging from 64 to 1024, and more preferably 256 or 512. h(t) can be obtained by linear projection after concatenating the text encoder, video encoder, character encoder, and selling point encoder, and its calculation formula is:

[0028] In the formula, h(t) represents the basic representation vector of the t-th narrative unit, which belongs to the d-dimensional real space; e text (t), e vision (t), e role (t), e sell (t) represent the text semantic vector, the camera visual vector, the character identity and persona vector, and the selling point semantic vector, respectively; concat(·) represents vector concatenation; W h Let b represent a linear mapping matrix. h d represents the bias term; d represents the dimension of the unified narrative representation.

[0029] To ensure that the materials do not deviate from the original materials or initial draft, an initial anchor vector h0(t) is also defined. This vector is obtained from the original script, existing character settings, or basic material codes input by the brand, and is used for subsequent stability constraints.

[0030]

[0031] In the formula, z(1) represents the initial global state vector of the advertising material; MeanPool(·) represents the average pooling operator; T represents the total number of narrative units; and h(t) represents the basic representation vector of the t-th narrative unit.

[0032] Step 2: Construction of a multi-agent parallel proposal network Construct M intelligent agents with different functions but sharing the same material state space. Preferably, M = 5 to 8; in this embodiment, it is preferable to construct five types of core intelligent agents: script intelligent agent A N Brand Intelligent Agent A B Transformation agent A C Compliant Intelligent Agent A R And novel intelligent agent A D If necessary, the agent can be expanded to include a rhythm agent, an audience resonance agent, or a platform-adaptive agent. The m-th agent receives the current local state s(t)={h(t),z(t),c(t)} at the t-th narrative unit and outputs a local proposal vector a(m,t) and a local score q(m,t). Here, c(t) represents the context cache, which can contain summaries of the first k narrative units, brand keyword cache, and constraint state cache.

[0033] The local proposal network of the m-th agent can be represented as:

[0034] In the formula, a(m,t) represents the local proposal vector output by the m-th agent for the t-th narrative unit, with the same dimension as h(t); q(m,t) represents the corresponding local score; h(t) represents the local narrative unit vector; z(t) represents the current global state vector; c(t) represents the context cache; θ m f represents the parameter set of the m-th agent; m g represents the proposal generation function; m This represents a local scoring function.

[0035] To avoid mechanical averaging among agents, this invention defines an adaptive gating coefficient α(m,t). This coefficient is related not only to the local score q(m,t), but also to the agent's current uncertainty σ(m,t) and its gradient sensitivity to the overall objective. (m,t) are related, and the calculation formula is:

[0036] In the formula, α(m,t) represents the gating coefficient of the m-th agent at the t-th narrative unit; A(m,t) represents the exponential term obtained by combining the local score, uncertainty, and gradient sensitivity; σ(m,t) represents the current prediction uncertainty of the agent; ε_σ represents the numerical stability term, preferably ranging from 10^-6 to 10^-3; γ_g represents the gradient sensitivity scaling factor, preferably ranging from 0.1 to 5. (m,t) represents the gradient sensitivity of the local utility U_m(X) of the m-th agent to the vector h(t) of the t-th narrative unit, which can be taken as || U_m(X) / h(t)||_2 is the normalized value; τ(t) represents the game temperature parameter, preferably 0.2 to 3; M represents the total number of agents.

[0037] Therefore, the global state is updated as follows:

[0038] In the formula, z(t+1) represents the global state vector at step t+1; LN(·) represents the layer normalization operator; W z and W m Ψ represents the learnable weight matrix; β(m,n,t) represents the interaction strength coefficient between the m-th agent and the n-th agent at the t-th narrative unit, preferably in the range [0,1]; mn (·) represents the kernel function for interaction between intelligent agents.

[0039] Step 3: Constructing the agent utility function, conflict potential function, and constraint default function The key to this invention lies in transforming the multi-agent cooperative generation problem into a computable constrained coupled potential game problem. Let the final material solution be represented as X = {h(1), h(2), ..., h(T)}, and define the utility function U of the m-th class of agents. m (X); Define the conflict potential function C(m,n;X) between agents m and n; Define the default function G for the k-th business constraint. k (X). The unified objective function is written as:

[0040] In the formula, Φ(X,λ,ρ) represents the objective function of the constrained coupled potential game; X represents the material solution composed of T narrative unit vectors; λ represents the set of constrained dual variables; ρ represents the set of agent conflict multipliers; ω m Let represent the utility weight of the m-th agent, with an initial value preferably between [0.05, 0.5] and satisfying Σ. m ω m =1; U m(X) represents the utility function of the m-th agent; C(m,n;X) represents the conflict potential function between the m-th agent and the n-th agent; G k (X) represents the default function for the k-th business constraint; η s represents the stability anchoring coefficient; h0(t) represents the initial anchor vector of the t-th narrative unit.

[0041] In this implementation, at least four types of core utility functions are defined.

[0042]

[0043] In the formula, U N (X) represents the utility of the script agent; cos(·,·) represents the cosine similarity, used to measure the coherence of adjacent narrative units; κ(t) represents the hook strength score of the t-th narrative unit, which can be output by the sentiment peak model, suspense recognition model, or reversal recognition model; ξ h The hook enhancement factor is represented, preferably between 0.05 and 2; T represents the total number of narrative units.

[0044]

[0045] In the formula, U B (X) represents the brand agent utility; b* represents the brand objective vector; π brand (X) represents the brand consistency prediction header output; y brand * Indicates the ideal brand messaging distribution set by the brand owner; ξ b This represents the brand deviation penalty coefficient, preferably ranging from 0.1 to 10.

[0046]

[0047] In the formula, U C (X) represents the utility of the transformative agent; p conv (t) represents the conditional contribution probability of the t-th narrative unit to the transformation; (t) represents the time decay weight of the t-th narrative unit; ρ c represents the time decay factor, preferably 0 to 2; j represents the normalized summation index.

[0048]

[0049] In the formula, U D (X) represents the utility of the novel agent; K(X) represents the kernel matrix composed of narrative unit similarity; I represents the identity matrix; ε d This represents the regularization constant, preferably 10. -4 Up to 10 -1The log det(·) function is used to measure the internal differences and information volume of the material to avoid excessive repetition of shots and expressions.

[0050] Regarding compliance, it is preferable to treat compliance as a hard constraint rather than a general reward. Definition:

[0051] In the formula, G1(X), G2(X), and G3(X) represent the default functions corresponding to compliance risk, brand deviation, and material length, respectively; p risk (X) represents the overall compliance risk probability; δ r This represents the tolerable risk threshold, preferably between 0.01 and 0.2; p brandoff (X) represents the probability of brand deviation; δ b Indicates brand deviation from threshold; L(X) represents total duration or total word count of the material; L max This indicates an upper limit constraint on the length of the material; max(0,·) represents a function that outputs a positive penalty value only when a violation occurs.

[0052] To address conflicts between agents, a conflict potential function is defined as follows:

[0053] In the formula, C(m,n;X) represents the conflict potential function between the m-th agent and the n-th agent; a(m,t) and a(n,t) represent the local proposal vectors of the two agents at the t-th narrative unit, respectively; Δ(m,t) and Δ(n,t) represent the editing action vectors, respectively; r(m,t) and r(n,t) represent the rhythm or expression intensity scalars, respectively; μ1, μ2, and μ3 represent the conflict component weights, which are preferably non-negative and normalized according to the validation set.

[0054] Step 4: Solving the Primitive-Dual Equilibrium and Updating Materials After the unified potential function is determined, this invention does not employ a simple weighted summation followed by a single decoding step, but instead uses a primal-dual iterative strategy to solve for the equilibrium material. The primal variables are the narrative unit vectors in X; the dual variables are the constraint multipliers λ={λ_1,...,λ_2}. K} and the conflicting multiplier ρ={ρ(m,n)}. In the r-th iteration, the original variables are updated to:

[0055] In the formula, X(r) represents the material solution in the r-th iteration; Π Ω [·] denotes the projection operator, used to restrict the updated material representation to the feasible set Ω; η x This represents the update step size of the original variable, preferably 10. -4 Up to 10 -1D(r) represents the diagonal precondition matrix of the r-th iteration; X This indicates that the gradient of the solution X is calculated over the data.

[0056] The dual variable is updated as follows:

[0057]

[0058] In the formula, λ k (r) represents the dual variable of the k-th constraint in the r-th iteration; ρ(m,n,r) represents the conflict multiplier between the m-th and n-th agents in the r-th iteration; [u] + Represents max(0,u); η λ This indicates the step size for the constraint multiplier update, preferably 10. -4 Up to 1; η ρ Indicates the step size of the conflict multiplier update; (m,n) represents the conflict tolerance threshold of the agent, preferably between 0 and 1.

[0059] The game temperature parameter τ(t) can also be adaptively adjusted according to the conflict intensity and the degree of default, making the system more inclined to make sharp choices when the conflict is high and more encouraging multi-agent fusion when the conflict is low.

[0060] In the formula, τ min and τ max Let τ represent the lower and upper limits of the game temperature, respectively, preferably satisfying 0 < τ. min <τ max ≤5; σ(·) represents the sigmoid function; Var(·) represents variance calculation; This indicates that the default function for each business constraint is maximized; mean m<n This represents the average of the conflict potentials across all agent pairs; a τ b τ c τ This represents the scaling factor.

[0061] When the iteration limit R is reached max Alternatively, an equilibrium solution can be output when the convergence condition is met. Preferably, the convergence condition can be set as follows:

[0062] In the formula, ||·|| F Denotes the Frobenius norm; x Represents the numerically stable term; stop This represents the stopping threshold, preferably 10.-4 Up to 10 -2 .

[0063] Step 5: Scenario-level approximate Shapley credit allocation To map real-world delivery feedback to each agent and narrative unit, this invention proposes a scene-level approximate Shapley credit allocation mechanism. Let the overall value function of the content be V(X), representing the unified value obtained by combining impressions, click-through rate, conversion rate, transaction amount, and cost. The marginal contribution to the m-th agent is defined as the Shapley value:

[0064] In the formula, φ(m) represents the Shapley marginal contribution value of the m-th agent; N represents the set of agents, N={1,2,...,M}; S represents any subset of agents that does not contain the m-th agent; S∪{m} represents the set formed after adding the m-th agent to the subset S; |S| represents the size of the subset; V(·) represents the overall value function of the material. Since full enumeration has high complexity, this invention uses a random permutation method for approximation.

[0065]

[0066] In the formula, φ approx (m) represents the approximate Shapley value of the m-th agent; P s (m) represents the set of agents preceding the m-th agent in the s-th random permutation; S sample The number of random samplings is indicated, preferably between 8 and 256, to distinguish it from the aforementioned subset S of intelligent agents.

[0067] To further refine this to the narrative unit level, the local credit value of the m-th agent to the t-th narrative unit is defined as:

[0068] In the formula, χ(m,t) represents the local credit value of the m-th agent at the t-th narrative unit; υ(m,t) represents the local contribution strength of the m-th agent at the t-th narrative unit, which can be determined by at least one of attention weight, gradient contribution, and edit amplitude, or by a weighted fusion of the above multiple factors; j represents the narrative unit index; χ(m,t) summed over t equals φ. approx (m).

[0069] Based on the credit score, update the agent weights and update step size:

[0070] In the formula, ω new (m) and ωold (m) represent the utility weight of the m-th agent after the update and the utility weight before the update, respectively; β ω κ represents the weighted smoothing coefficient, preferably ranging from 0.01 to 0.5; κ represents the temperature amplification coefficient, preferably ranging from 0.1 to 10; φ approx (j) represents the approximate Shapley value of the j-th agent.

[0071]

[0072] In the formula, η new (m) represents the learning rate updated by the m-th agent; η0 represents the base learning rate; ξ η σ represents the scaling factor; φ ε represents the standard deviation of the Shapley values ​​over the most recent rounds; φ This represents the stabilizing term. This formula allows agents with higher marginal contributions to receive larger update increments.

[0073] Step 6: Closed-loop update of real-world delivery feedback This invention introduces real-world delivery feedback to enable the creative collaboration algorithm to continuously learn from ad delivery results. For the i-th creative sample, let its exposure be Imp. i Click-through rate (CTR) i Conversion rate CVR i The transaction amount is GMV i The consumption cost is Cost i The retention or retention quality index is Ret i Then, the unified value signal is defined as:

[0074] In the formula, Y i CTR represents the uniform value signal of the i-th material sample. i CVR i GMV i Ret i and Cost i These represent the click-through rate, conversion rate, transaction amount, retention or dwell quality index, and cost of consumption for the i-th creative sample, respectively; norm(·) represents the normalization operator, which can use minimum-maximum normalization, quantile normalization, or z-score standardization; ν1 to ν5 represent the business weights, all of which are non-negative numbers, and can be determined according to the campaign objectives satisfying ∑ j ν j =1.

[0075] The supervisory objective for transforming an agent can be written as:

[0076] In the formula, L C This represents the supervised training loss of the transforming agent; (i) represents the value prediction of the conversion agent for the i-th material sample; Y(i) represents the unified value signal constituted by the actual delivery feedback; N sample λ represents the sample size; reg θ represents the regularization coefficient. C This indicates the transformation of agent parameters.

[0077] The script agent, brand agent, and novelty agent can share the credit allocation signal and update it respectively through the following objectives:

[0078] In the formula, L m Let represent the training loss of the m-th agent; χ(m,t,i) represent the credit value of the m-th agent at the t-th narrative unit on the i-th material sample; e*(t,i) represent the final adopted editing action; s(t,i) represent the state of the i-th material sample at the t-th narrative unit; π m λ represents the action distribution of the m-th agent; m Ω represents the weight of the regularization term; m (θ m ) represents the regularization term of the m-th agent parameter; m∈{N,B,D,R} corresponds to script, brand, novelty and compliance agents respectively.

[0079] Thus, this invention forms a complete algorithmic closed loop, from material scene modeling, multi-agent game solving, scene-level credit allocation to real-world delivery feedback updates.

[0080] The key parameters and their values ​​in this invention are described in Table 1.

[0081] Table 1

[0082] The present invention will be further illustrated below through several specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. Any parameter substitution, function substitution, or structural equivalent change without departing from the core idea of ​​the present invention should fall within the scope of protection of the present invention.

[0083] Example 1: Collaborative Generation Example of Short Drama Ad Script In this embodiment, the material to be generated is a 45-second short drama advertisement, which is divided into T=9 narrative units, corresponding to: opening conflict, character introduction, pain point introduction, selling point placement, escalation of misunderstanding, product proof, emotional reversal, call to action, and brand exposure at the end. The initial script draft, character settings, product information, brand tone description, and platform compliance rules are used as inputs to form an initial scene hypergraph.

[0084] The five core intelligent agents perform the following tasks: the script agent is responsible for maintaining the causal order, emotional progression, and hook design; the brand agent constrains the brand tone, category description, and advertising message; the conversion agent is responsible for arranging the order of selling points, the location of calls to action, and the strength of benefits; the compliance agent is responsible for identifying absolute terms, sensitive words, exaggerated promises, and prohibited expressions; and the novelty agent is responsible for reducing template repetition and enhancing the freshness of the scenario.

[0085] In this embodiment, the preferred settings are: d=256, M=5, η x =0.02, η λ =0.05, η ρ =0.02, τ min =0.4, τ max =2.2, δ r =0.06, δ b =0.08, S=32. If the original script is relatively mature, then η s It is advisable to increase η appropriately, such as by 0.5 to 2, to reduce the offset relative to the initial draft; if the material only contains the theme and not an initial draft, then η s It can be reduced to 0.05 to 0.2.

[0086] After the primal-dual iteration, the system outputs a complete advertising material scheme including the script text, character dialogue, camera movement, subtitle information, and the insertion position of selling points. If the compliance agent detects that the risk value of a specific narrative unit exceeds the threshold, the corresponding multiplier λ1 will automatically increase, thereby suppressing the further enhancement of exaggerated expressions and high-risk rhetoric in that narrative unit.

[0087] Example 2: An example of updating agent weights using real delivery feedback In this embodiment, the advertising company places multiple versions of the same product on an in-feed advertising platform. For the i-th product sample, the impression count Imp is collected. i Click count (Clk) i Conversion rate (Ord) i Transaction Amount (GMV) i and actual cost i CTR i =Clk i / Imp i CVRi =Ord i / max(Clk i 1) As a basic indicator, and compared with GMV i and Cost i Together they constitute a unified value signal Y i .

[0088] If, after a round of campaigning, the overall click-through rate of the creative materials is high but the conversion rate is low, the Shapley marginal contribution value of the conversion agent may be negative, while the marginal contribution of the script agent and the novelty agent may be positive. In this case, the system uses ω... m and η m The adaptive update mechanism automatically reduces the conversion agent's preference for strategies that are "too hypey and have weak sales results," while increasing the influence of the brand and script agents in the generation of subsequent materials.

[0089] Furthermore, in the local credit value χ m,t At the level of analysis, if a narrative unit is identified as a key location leading to high user clicks but low conversions, the strength of its selling points, dialogue length, and camera pacing will be adjusted locally in the next generation, rather than completely rewriting the entire script. Therefore, this invention enables fine-grained material optimization, rather than a coarse-grained overall rewrite.

[0090] Example 3: System Structure Example Corresponding to the above method, the present invention also provides a multi-agent game-driven collaborative advertising material system. The system includes at least: a scene hypergraph construction module, a multi-agent proposal module, a potential game solving module, a credit allocation module, and a feedback learning module.

[0091] The scene hypergraph construction module is used to complete narrative unit segmentation, basic vector encoding, and scene hypergraph modeling; the multi-agent proposal module is used to run script, brand, conversion, compliance, and novelty agents in parallel; the potential game solving module is used to calculate the unified potential function Φ(X,λ,ρ) and perform primal-dual equilibrium updates; the credit allocation module is used to estimate the agent marginal contribution based on the approximate Shapley value; and the feedback learning module is used to receive real business feedback such as exposure, clicks, conversions, transactions, and costs, and update the parameters of each agent.

[0092] The system can be deployed on an advertising company's material production platform, creative platform, automatic editing system, or delivery operation platform. It can run independently as an offline script collaborative generation module, or it can be used as an online creative optimization subsystem to form a closed-loop linkage with the material delivery process.

[0093] In summary, this invention addresses the problems of unified modeling of multi-agent constrained coupled potential game problem in existing technologies by refining the collaborative generation of advertising materials as a multi-agent constrained coupled potential game problem, and supplementing it with primal-dual equilibrium updates and an approximate Shapley credit allocation mechanism. This technical solution possesses clear algorithmic innovation, engineering feasibility, and adaptability to advertising business.

[0094] Any part of this invention not described in detail can be referred to in the prior art or in the art known to those skilled in the art. This embodiment does not limit such part and will not describe it in detail here.

[0095] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of the present invention.

Claims

1. A method for multi-agent game driven advertisement material collaboration, characterized in that, The method includes: S101. Scene hypergraph representation and initial state encoding of advertising creatives; S102. Run the multi-agent network in parallel, output local proposals and adaptively fuse them to generate global state update inputs; S103. Based on the agent utility function, agent conflict potential function, and constraint default function, a unified constraint coupled potential game objective is formed; S104. Based on the objective of the constrained coupled potential game, the primal-dual iterative strategy is used to solve for the equilibrium materials and update the materials. S105. Perform scene-level approximate Shapley credit assignment to quantify the contributions of each agent and narrative unit; S106. Real-time delivery feedback and closed-loop update of agent parameters. 2.The multi-agent game driven advertisement material collaboration method of claim 1, wherein, Step S101 includes: Segmentation: Divide the material into T narrative units, including shots, dialogue, and selling points; Modeling: Construct a scene hypergraph G=(V,E,H), where V is a narrative unit node, E is a temporal edge, and H is a cross-unit semantic hyperedge; each narrative unit node corresponds to a narrative vector, brand vector, selling point vector, and risk vector; Encoding: Each narrative unit generates narrative, brand, selling point, and risk vectors to obtain an initial global state vector, while retaining the initial anchor vector to control material offset. 3.The multi-agent game driven advertisement material collaboration method of claim 1, wherein, The multi-agent network described in step S102 includes at least a script agent, a brand agent, a conversion agent, a compliance agent, and a novelty agent.

4. The multi-agent game-driven advertising material collaboration method according to claim 1, characterized in that, In step S102, each agent outputs a local proposal vector and a local score at the t-th narrative unit, and fuses them through adaptive gating coefficients to generate and update the global state; The adaptive gating coefficients are determined based on agent utility, prediction uncertainty, and gradient sensitivity.

5. The multi-agent game-driven advertising material collaboration method according to claim 1, characterized in that, The unified constraint coupling potential game objective in step S103 is to maximize the weighted agent utility and scenario-level conversion benefits, while minimizing the inter-agent conflict potential and default penalty term, and includes a stability anchor term relative to the initial material.

6. The multi-agent game-driven advertising material collaboration method according to claim 5, characterized in that, The agent utility functions include script agent utility functions, brand agent utility functions, conversion agent utility functions, and novelty agent utility functions; The proxy conflict potential function is jointly calculated based on the angle difference, semantic edit distance and rhythm offset of the two agents' proposal vectors for the same narrative unit, and is used to suppress the mutual destruction between script, brand, conversion and compliance suggestions; The constraint default function is a hard constraint function based on the overall compliance risk probability, tolerable risk threshold, brand deviation probability, brand deviation threshold, total duration or total number of words of the material, and upper limit constraint, which imposes penalties when a default occurs.

7. The multi-agent game-driven advertising material collaboration method according to claim 6, characterized in that, The utility function of the script agent is calculated jointly based on the cosine similarity used to measure the coherence of adjacent narrative units, the hook strength score of the narrative unit, and the hook enhancement coefficient. The brand intelligent agent utility function is calculated jointly based on the brand target vector, the output of the brand consistency prediction head, the ideal brand caliber distribution set by the brand owner, and the brand deviation penalty coefficient. The utility function of the transformation agent is calculated jointly based on the conditional contribution probability of the narrative unit to the transformation, the time decay weight, and the temporal decay factor. The novel agent utility function is calculated jointly based on a kernel matrix, an identity matrix, and a regularization constant, which are composed of narrative unit similarity.

8. The multi-agent game-driven advertising material collaboration method according to claim 1, characterized in that, Step S104, which uses the primal-dual iterative strategy to solve for the equilibrium material, includes: Original variable update: Iteratively optimize the narrative unit vector, project it onto the feasible region, and adjust the update scale using the precondition matrix; Dual variable update: Synchronously update constraint multipliers and proxy conflict multipliers, and strengthen penalties when violations occur; Dynamic adjustment of game temperature parameters: The system is adaptively adjusted according to the intensity of conflict and the degree of breach of contract, so that the system is more inclined to sharp choices when the conflict is high and more encouraged to merge multiple agents when the conflict is low. Convergence: When the iteration limit is reached or the convergence condition is met, output the balanced material solution.

9. The multi-agent game-driven advertising material collaboration method according to claim 1, characterized in that, Step S105 includes: Marginal contribution: The Shapley value of each agent's contribution to the material is approximated by random permutation. Unit-level credit: Refines agent contributions to each narrative unit, linking attention weights and gradient contributions. Parameter update: The agent weights and learning rate are adaptively adjusted based on the credit value, so that agents with higher marginal contributions can receive a larger update range.

10. A multi-agent game-driven collaborative advertising material system, characterized in that, The system is capable of implementing the multi-agent game-driven collaborative advertising creative method as described in any one of claims 1 to 9; the system includes: Scene hypergraph construction module: It is configured to complete the structured modeling and initial vector encoding of advertising materials, providing the basic state for subsequent game; Multi-agent proposal module: configured to run multiple types of agents in parallel, output local proposals and adaptively fuse them to generate global state update input; The potential game solving module is configured to solve equilibrium materials through primal-dual iteration based on constraint-coupled potential game objectives, satisfying multi-objective constraints. Credit allocation module: It is configured to quantify the contributions of each agent and narrative unit, providing a basis for adaptive parameter updates; Feedback learning module: It is configured to receive real-world delivery data, generate valuable signals, update the agent parameters in a closed loop, and improve the quality of generated materials.

Images