Complex multi-party cooperation service value and quality conflict resolution method

Abstract

The invention discloses a value and quality conflict resolution method for complex multi-party cooperation service, and belongs to the technical field of computer service. The method comprises the following steps: step S1, value conflict modeling in a multi-party cooperation service; s2, negotiation proxy based on large language model fine tuning; s3, negotiating a willingness perception and calculation method; and S4, establishing a complex multi-party cooperation service value conflict negotiation framework. Aiming at the characteristics of the complex multi-party cooperation service, the complexity of the complex multi-party cooperation service is analyzed, the value conflict is converted into a solvable maximum consensus level problem through value modeling, the value and quality conflict of the multi-party cooperation service is converted into a service negotiation problem, and multiple participants are assisted to complete the service quality parameter design of the complex multi-party cooperation service. According to the method, an automatic negotiation framework based on a large model agent is constructed, the negotiation agent is finely adjusted by using a large language model, multiple participants are helped to complete the processes of negotiation conversation, strategy learning, opponent modeling and the like, and negotiation is assisted to be completed.

Description

Technical Field

[0001] The present invention belongs to the field of computer service technology, and relates to a method for resolving conflicts between value and quality of complex multi-party collaborative services, and specifically to complex multi-party collaborative service value conflict modeling, service conflict negotiation agent, and service conflict negotiation optimization framework. Background Art

[0002] Complex multi-party collaborative services usually refer to services that can support complex collaboration and communication needs between multiple teams, departments or organizations. With the acceleration of urbanization, the construction of complex multi-party collaborative services has become one of the key directions for the development of social services. At the same time, the rapid integration of modern services and the rapid emergence of new services have brought challenges to service design and deployment, requiring collaboration and interaction between multiple entities and systems, that is, making full use of the collaborative relationship between various service organization entities to achieve greater service value and innovative value-added. Through multi-party collaborative services, cross-domain and cross-organizational service providers share resources according to a specific collaboration model, and the service value is smoothly circulated and transformed in the large service ecosystem, creating new value-added.

[0003] In complex multi-party collaborative service systems, different service providers, service groups, and systems often have different goals, strategies, and needs, which may lead to conflicts. In complex multi-party collaborative service scenarios, there are complex dependencies between participants and between participants and global services in terms of value, quality, and capabilities, making it difficult for participants to achieve their maximum service value expectations. Effective strategies need to be formulated to resolve this and develop collaborative decision-making mechanisms and negotiation strategies suitable for complex multi-party collaborative services.

[0004] When designing a multi-party collaborative service, it is necessary to comprehensively consider the value demands / value expectations of multiple parties, break down each atomic service in the multi-party collaborative service into specific QoS (service quality parameters), and quantify the value demands of each party into the QoS indicator value of each service. Value conflicts are reflected in the fact that there are significant differences in the QoS indicator values ​​of various services for multiple parties. Therefore, it is necessary to adjust the QoS of multiple parties. The adjustment process is the process of conflict resolution. The adjustment goals can be diverse, such as the minimum cost principle, the fairness principle, the shortest negotiation round, etc. The QoS configuration obtained after optimization and resolution is the QoS configuration for reference when designing multi-party collaborative services. Summary of the invention

[0005] The purpose of the present invention is to solve the above-mentioned problems existing in the prior art and to provide a method for resolving conflicts between value and quality of complex multi-party collaborative services.

[0006] The present invention specifically relates to complex multi-party collaborative service value conflict modeling, service conflict negotiation agent, service conflict negotiation optimization framework, etc. The present invention can evaluate the degree of conflict by modeling the value in multi-party collaborative services and quantifying the consensus level of multiple parties on QoS, thereby determining the specific QoS value that needs to be adjusted. In addition, there may be a nonlinear relationship between the QoS of collaborative services. When a certain QoS is adjusted, it will affect other parameters that have a relationship with it. Therefore, this problem can be modeled as an optimization problem with coupling constraints, and a corresponding negotiation optimization method is proposed. A multi-agent negotiation agent framework is proposed to handle various negotiation problems in multi-party negotiations of collaborative services and to construct an automated and intelligent multi-agent negotiation system. Through technologies such as fine-tuning of large-scale language models, a self-interested agent negotiation agent is trained to complete multi-party negotiation dialogues according to the negotiation intention, and to learn negotiation strategies and opponent modeling tasks.

[0007] To achieve the above object, the technical solution adopted by the present invention is:

[0008] A method for resolving conflicts between value and quality of complex multi-party collaborative services, the method comprising the following steps:

[0009] Step S1: Modeling value conflicts in multi-party collaborative services;

[0010] Step S2: Negotiation agent fine-tuned based on a large language model;

[0011] Step S3: Negotiate willingness perception and calculation methods;

[0012] Step S4: Establish a complex multi-party collaborative service value conflict negotiation framework.

[0013] Furthermore, in step S1, the specific process of value conflict modeling in the multi-party collaborative service is:

[0014] Suppose there are M service quality parameters and N service participants, then the VQC matrix is ​​expressed as:

[0015] P=(p i,j ) M×N ,p i,j ∈[0,1]

[0016] P l,t =(p i,j ) M×N , represents the QoS matrix of the i-th participant in the t-th negotiation round; p i,j represents the jth service quality parameter value of the i-th participant. Assume that each participant has its own evaluation function v(p) for the service quality parameter, V:P→R, p∈P, P represents the service quality parameter set, and V represents the value evaluation function;

[0017] but represents the evaluation function of the ith participant on QoS, assuming that V is second-order differentiable with respect to P;

[0018] Negotiate and adjust the matrix X l,t =(x i,j ) M×N , represents the adjustment made by the ith participant after the tth round of negotiation, and its QoS matrix in the t+1th round is represented by P l,t+1 =P l,t +X l,t ;P l,t and X l,t denote the service quality parameters and adjustment matrix in the lth iteration in the tth round respectively;

[0019] Obviously, 0≤p i,j +x i,j ≤1, then the value of x is [-1,1];

[0020] Preference Matrix It represents the preference relationship between the kth participant and the QoS. The preference relationship is approximately calculated using the hierarchical analysis method, and then the square root method is used to approximate the preference relationship:

[0021]

[0022] Then normalize it:

[0023]

[0024] That is, the preference coefficient of the kth participant for each QoS can be calculated;

[0025] The consensus level, that is, the similarity between the QoS matrices of each participant, is calculated as follows:

[0026]

[0027] Where: P l represents the QoS matrix of the i-th participant, X l represents the adjustment matrix of the i-th participant after negotiation;

[0028] The participant's payoff w is Represents the participant's benefit from the current QoS; the selfish loss Δw k That is, the maximum benefit of the user minus the user benefit adjusted after t rounds of negotiations;

[0029]

[0030] Use the adjacency matrix to represent the QoS constraints:

[0031] E=(e i,j ) M×M ,e i,j ∈{0,1}

[0032] If and only if there is a constraint relationship between the i-th service quality parameter and the j-th service quality parameter, then e i,j =1;

[0033] For each participant, the objective function of each round of negotiation is:

[0034]

[0035] Furthermore, in step S2, the negotiation agent fine-tuned based on the large language model is specifically:

[0036] Based on the historical negotiation data of users, a prompt is constructed to fine-tune the large language model. The specific fine-tuning method is as follows:

[0037] Step S21: prompt project;

[0038] Construct four prompt words: <intention>, <item>, <agreement>, <partner_intention>, which are used to integrate text information with participant preference ρ, conflict resolution candidate set CP, negotiation end mark, and opponent negotiation willingness prediction value into prompt;

[0039] Step S22: construct input text <input_text>;

[0040] inputs=<|im_start|>user+<input_text>+<|im_end|><|im_start|>assistant;

[0041] Construct the continuation text <output_text>;

[0042] outputs=<output_text>+<|im_end|>;

[0043] Encode inputs and outputs into inputsId and labels through tokenizer, and build them into input and output of large language models;

[0044] Step S23: load the pre-trained model and tokenizer from the large language model library, set the LoRAConfig parameters, and continue training using the LoRA weights pre-trained by the large language model;

[0045] Step S24: token mapping, mapping the dialogue text roles and text content so that the dialogue data can be adapted to the pre-trained large language model;

[0046] Step S25: Load the SFT trainer and fine-tune by using the labeled data. Load the inputsId and labels in step S24 into the trainer for training.

[0047] Furthermore, in step S3, the negotiation willingness perception and calculation method is specifically:

[0048] Model the preferences and value evaluation functions of multiple parties, i.e., ρ, V; where: ρ represents the preference coefficient, and V represents the value evaluation function;

[0049] The proxy negotiation target is: U = μO + σ (Δw k +Δw -k );

[0050] Where: μ represents the initial consensus coefficient, σ represents the self-interest coefficient, both of which are adjusted hyperparameters; O represents the consensus degree of multiple parties on the collaborative service quality parameters, and W represents the self-interest loss of the participants;

[0051] Define negotiation willingness Q k = {q i |i∈CP k},q∈[0,1], represents the negotiation willingness of participant k for each QoS in the conflict resolution candidate, and the negotiation willingness calculation formula is:

[0052]

[0053] Where: Φ i It indicates the negotiation willingness of the participants on the i-th service quality parameter.

[0054] Furthermore, in step S4, the specific steps of establishing a complex multi-party collaborative service value conflict negotiation framework are as follows:

[0055] Step S41: Initialize parameters; initialize N services and M service quality parameters, then the QoS matrix is ​​P = (p i,j ) M×N , initialize the value evaluation function V = (v i,j ) M×N ; Initialize the preference matrix Initialize Agent parameters: Agent = <P, V, ρ>, set the maximum negotiation round L and the maximum single-round negotiation round T, set the initial consensus coefficient μ and the selfishness coefficient σ;

[0056] Step S42: Conduct negotiation;

[0057] Step S43: Obtain the adjustment matrix of each participant, select the optimal negotiation result by calculating the global consensus level O and the self-interest loss Δw, and update and retain the optimal negotiation parameters and negotiation results; Δw represents the self-interest loss;

[0058] Step S44: determine whether the current negotiation round satisfies l>l orO<ε*, where ε* is the set negotiation threshold super parameter; if so, output the optimal negotiation result and end the operation process; if not, execute step S45;

[0059] Step S45: Re-adjust the negotiation parameters of the participants, adjust the QoS matrix P=P+X for the conflict resolution candidate set CP and the adjustment matrix X, adjust the consensus coefficient and the self-interest coefficient μ, σ, return to step S41, and start a new round of negotiation.

[0060] Furthermore, in step S42, the specific steps of the negotiation are as follows:

[0061] Step S421: The Agent calculates the consensus level O of QoS according to the expected quality parameters;

[0062]

[0063] Where: At t = 0, X t =0, then the consensus level of each QoS is calculated;

[0064] Step S422: According to the consensus level O i , retrieve the conflicting QoS set C = {P i |O(P i )>ε}, since the QoS constraint relationship is represented as a graph structure, the relevant conflict paths are searched according to the adjacency matrix E to retrieve the conflict resolution candidate set CP;

[0065] Step S423: The multi-agent agent participates in the negotiation, predicts the negotiation willingness of each participant, and calculates the adjustment matrix X = (x i ) M ;

[0066] Step S424: After the proxy calculates the adjustment matrix X, it recalculates the global consensus level O after the adjustment matrix is ​​given by all parties. If O<ε, execute step S43; otherwise, determine whether the current single maximum negotiation round T is exceeded. If not, go to step S421, otherwise execute step S43 and output the adjustment matrix.

[0067] Furthermore, step S423 includes the following steps:

[0068] Step S4231: Initialize Agent and negotiate parameters;

[0069] Step S4232: Start negotiation; construct input, start polling and speaking according to the role, and output the answer format according to the prompt word; when the prompt word appears <partner_intention>, each negotiation agent outputs the predicted negotiation intention, and calculates the corresponding adjustment matrix X according to the objective function. The objective function is as follows:

[0070]

[0071] Among them: O represents the consensus degree of multiple parties on the quality parameters of collaborative services, W represents the selfish loss of participants;

[0072] Step S4233: Return the adjustment matrix X and return to step S424.

[0073] The beneficial effects of the present invention relative to the prior art are:

[0074] 1. The present invention aims at the characteristics of complex multi-party collaborative services, analyzes their complexity, and transforms value conflicts into solvable problems of maximizing the consensus level through value modeling, and transforms the value and quality conflicts of multi-party collaborative services into service negotiation problems, thereby assisting multiple participants in completing the service quality parameter design of complex multi-party collaborative services.

[0075] 2. The present invention constructs an automated negotiation framework based on a large language model agent. By using a large language model to fine-tune the negotiation agent, it helps multiple parties complete negotiation dialogues, strategy learning, opponent modeling and other processes, and assists in completing the negotiation. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 It is the overall flow chart of the algorithm in the present invention;

[0077] Figure 2 It is the algorithm structure diagram of the present invention;

[0078] Figure 3 It is an automated negotiation flow chart based on multi-party negotiation game;

[0079] Figure 4 Fine-tune the flow chart for negotiation proxy. DETAILED DESCRIPTION

[0080] like Figure 1-Figure 4 As shown, this embodiment discloses a method for resolving conflicts between value and quality of complex multi-party collaborative services, the method comprising the following steps:

[0081] Step S1: Modeling value conflicts in multi-party collaborative services;

[0082] Step S2: Negotiation agent fine-tuned based on a large language model;

[0083] Step S3: Negotiate willingness perception and calculation methods;

[0084] Step S4: Establish a complex multi-party collaborative service value conflict negotiation framework.

[0085] Furthermore, in step S1, the specific process of value conflict modeling in the multi-party collaborative service is:

[0086] In collaborative service design, service providers jointly propose collaborative services, which include several atomic services. Each atomic service has several QoS (quality of service parameters) and determines its value. For each participant in the collaborative service (users and service providers), they have their own value demands, which are mapped to QoS requirements, that is, value demands are quantified as QoS. Conflicts between participants are reflected in the different QoS values. The lower the similarity, the greater the conflict. The ultimate goal is to improve the consensus of multiple participants.

[0087] Suppose there are M service quality parameters and N service participants, then the VQC matrix is ​​expressed as:

[0088] P=(p i,j ) M×N ,p i,j ∈[0,1]

[0089] P l,t =(p i,j ) M×N , represents the QoS matrix of the i-th participant in the t-th negotiation round; p i,j represents the jth service quality parameter value of the i-th participant. Assume that each participant has its own evaluation function v(p) for the service quality parameter, V:P→R, p∈P, P represents the service quality parameter set, and V represents the value evaluation function;

[0090] but represents the evaluation function of the ith participant for QoS (quality of service parameter), assuming that V is second-order differentiable with respect to P;

[0091] Negotiate and adjust the matrix X l,t =(x i,j ) M×N , represents the adjustment made by the ith participant after the tth round of negotiation, and its QoS matrix in the t+1th round is represented by P l,t+1 =P l,t +X l,t ;P l,t and X l,tdenote the service quality parameters and adjustment matrix in the lth iteration in the tth round respectively;

[0092] Obviously, 0≤p i,j +x i,j ≤1, then the value of x is [-1,1];

[0093] Preference Matrix It represents the preference relationship between the kth participant and the QoS (quality of service parameters). The preference relationship is simply calculated using the hierarchical analysis method, and then the square root method is used to approximate the preference relationship:

[0094]

[0095] Then normalize it:

[0096]

[0097] That is, the preference coefficient of the kth participant for each QoS (quality of service parameter) can be calculated;

[0098] The consensus level, that is, the similarity between the QoS matrices of each participant, is calculated as follows:

[0099]

[0100] Where: P l represents the QoS matrix of the i-th participant, X l represents the adjustment matrix of the i-th participant after negotiation (the basic goal is to maximize the consensus level O after adjustment, and the maximum value is obviously 1; and for the convenience of solving, the absolute value term can be replaced by the square term, that is, the consensus function can be differentiable in the value range without changing its convexity);

[0101] The participant's payoff w is Represents the participant's benefit from the current QoS (quality of service parameter); the selfish loss Δw k That is, the maximum benefit of the user minus the user benefit adjusted after t rounds of negotiations;

[0102]

[0103] There are various associations between QoS (quality of service parameters), which in turn constrain the value space between the various quality of service parameters. This constraint relationship is expressed as a constraint function G(X), and the adjacency matrix is ​​used to represent the QoS constraint relationship:

[0104] E=(e i,j ) M×M ,e i,j ∈{0,1}

[0105] If and only if there is a constraint relationship between the i-th service quality parameter and the j-th service quality parameter, then e i,j =1;

[0106] For each participant, the objective function of each round of negotiation is:

[0107]

[0108] This step mainly models the value conflict of multi-party collaborative services, transforms the value conflict resolution problem into maximizing the level of multi-party consensus, and resolves the value conflict in collaborative services by minimizing the QoS (service quality parameter) differences of multiple participants. By adding self-interest losses, the self-interest level of the participants is guaranteed to a certain extent, so that the negotiation results can achieve a global optimal solution at the cost of a certain degree of loss of the interests of the participants.

[0109] Furthermore, in step S2, the negotiation agent fine-tuned based on the large language model is specifically:

[0110] Through the user's historical negotiation data, a prompt (representation word) is constructed to fine-tune the large language model (fine-tune the large language model to enable the large language model to have negotiation dialogue capabilities and willingness modeling capabilities); the specific fine-tuning method is as follows:

[0111] Step S21: prompt project;

[0112] Construct four prompt words: <intention>, <item>, <agreement>, and <partner_intention> to integrate text information with participant preference ρ, conflict resolution candidate set CP, negotiation end flag, and opponent negotiation willingness prediction value into the prompt (so that the large language model can generate specific task answers based on different prompt word information);

[0113] Step S22: construct input text <input_text>;

[0114] inputs=<|im_start|>user+<input_text>+<|im_end|><|im_start|>assistant;

[0115] Construct the continuation text <output_text>;

[0116] outputs=<output_text>+<|im_end|>;

[0117] Encode inputs and outputs into inputsId and labels through tokenizer, and build them into input and output of large language models;

[0118] Step S23: Load the pre-trained model and tokenizer (model encoder) from the large language model library, set the LoRAConfig parameters, and continue training using the LoRA weights pre-trained by the large language model;

[0119] Step S24: token mapping, mapping the dialogue text roles and text content so that the dialogue data can be adapted to the pre-trained large language model;

[0120] Step S25: Load the SFT trainer (SFT stands for supervised fine-tuning, which is an optimization process for large language models), and load the inputsId and labels in step S24 into the trainer for training by using labeled data for fine-tuning.

[0121] The purpose of step S2 is to fine-tune the large language model so that the large language model has the ability to negotiate dialogue and model intentions.

[0122] (1) Negotiation dialogue generation: Based on the current dialogue context, construct the intention, attitude, and target state representation of the two parties in the dialogue; use the pre-trained large language model as a basis for further fine-tuning and modeling; convert the optimized dialogue strategy into fluent and natural dialogue text, using the language generation capabilities of the large language model and combining domain knowledge for refined rewriting;

[0123] (2) Negotiation strategy learning: Based on the dialogue state, learn the optimal dialogue strategy, generate reasonable responses that can promote negotiation, and define the objective function of negotiation, such as reaching an agreement and maximizing the benefits of both parties. Design reasonable evaluation indicators, such as agreement quality, satisfaction of both parties, negotiation time, etc., and continuously improve the negotiation effect through continuous model optimization and strategy iteration;

[0124] (3) Negotiation / bidding: The large language model negotiation agent generates dialogues and learns negotiation strategies. Through multiple rounds of negotiation interactions or bidding behaviors, it finally reaches the negotiation requirements or completes the transaction. The large language model completes the negotiation transaction / bid by learning historical negotiation interactions / bidding behaviors.

[0125] (4) It is possible to extract key features from rich conversation / behavior data and construct an accurate opponent target preference model to predict the opponent’s emotions, decisions, etc. This can be supported by technologies such as large-scale data accumulation, multimodal information fusion, and reinforcement learning.

[0126] Furthermore, in step S3, the negotiation willingness perception and calculation method is specifically:

[0127] For each agent, the preference matrix and evaluation function largely determine the interest orientation and adjustment direction of the participants, while the conflict originates from the value expectations of each participant.

[0128] Then the conflict can also be regarded as the conflict of multiple parties' preferences and value evaluation functions for QoS (quality of service parameters). Therefore, it is necessary to model the preferences and value evaluation functions of multiple parties in the conflict resolution process, that is, ρ, V; where: ρ represents the preference coefficient, V represents the value evaluation function;

[0129] The proxy negotiation target is: U = μO + σ (Δw k +Δw -k );

[0130] Among them: μ represents the initial consensus coefficient, σ represents the selfishness coefficient, both of which are adjusted hyperparameters; O represents the consensus degree of multiple parties on the collaborative service quality parameters, and W represents the selfishness loss of the participants.

[0131] Calculate Δw -k It is necessary to accurately predict the preference coefficient ρ and value evaluation function V of other participants. In the negotiation dialogue, it is difficult to accurately model the global preference coefficient and value evaluation function. Therefore, the fine-tuned large language model is used to predict the relative preference coefficient ρ of the conflict resolution candidate set CP. The preference coefficient ρ is used as a negotiation willingness parameter to calculate the negotiation willingness of each participant for QoS (quality of service parameter) in the negotiation round.

[0132] Define negotiation willingness Q k = {q i |i∈CP k},q∈[0,1], represents the negotiation willingness of participant k for each QoS (quality of service parameter) in the conflict resolution candidate, and the negotiation willingness calculation formula is:

[0133]

[0134] Where: Φ i It indicates the negotiation willingness of the participants on the i-th service quality parameter.

[0135] In this step, the negotiation dialogue text is analyzed and the parties involved in the negotiation can express their wishes and needs through language or facial expressions, as well as their agreement or disagreement with the suggestions and requests made by other parties. The emotional changes of the participants in the dialogue are analyzed to predict the possibility of reaching a negotiation. This type of method mainly uses deep learning to learn and analyze text information and then predict the negotiation willingness of the participants.

[0136] The negotiation willingness of the participants is predicted by combining the decision-making behavior of the participants with time information. This type of method predicts the negotiation willingness of the participants through mechanisms or specific formula calculations, such as psychological sentiment calculation, fuzzy algorithms or short-term prediction algorithms such as grayscale prediction.

[0137] Furthermore, in step S4, the specific steps of establishing the complex multi-party collaborative service value conflict negotiation framework are as follows (this step is the implementation of the automated negotiation framework based on the large language model agent, and the overall negotiation algorithm is implemented as shown in steps S41-S45, see Table 1):

[0138] Step S41: Initialize parameters; initialize N services and M service quality parameters, then the QoS matrix is ​​P = (p i,j ) M×N , initialize the value evaluation function V = (v i,j ) M×N ; Initialize the preference matrix Initialize Agent parameters: Agent = <P, V, ρ>, set the maximum negotiation round L and the maximum single-round negotiation round T, set the initial consensus coefficient μ and the selfishness coefficient σ;

[0139] Step S42: Conduct negotiation;

[0140] Step S43: Obtain the adjustment matrix of each participant, select the optimal negotiation result by calculating the global consensus level O and the self-interest loss Δw, and update and retain the optimal negotiation parameters and negotiation results; Δw represents the self-interest loss;

[0141] Step S44: determine whether the current negotiation round satisfies l>l orO<ε*, where ε* is the set negotiation threshold super parameter; if so, output the optimal negotiation result and end the operation process; if not, execute step S45;

[0142] Step S45: Re-adjust the negotiation parameters of the participants, adjust the QoS matrix P=P+X for the conflict resolution candidate set CP and the adjustment matrix X, adjust the consensus coefficient and the self-interest coefficient μ, σ, return to step S41, and start a new round of negotiation.

[0143] Furthermore, in step S42, the specific steps of the negotiation are as follows:

[0144] Step S421: The Agent calculates the consensus level O of QoS according to the expected quality parameters;

[0145]

[0146] Where: At t = 0, X t=0, then the consensus level of each QoS (quality of service parameter) is calculated;

[0147] Step S422: According to the consensus level O i , retrieve the conflicting QoS set C = {P i |O(P i )>ε}, since the QoS constraint relationship is represented as a graph structure, the relevant conflict paths are searched according to the adjacency matrix E to retrieve the conflict resolution candidate set CP;

[0148] Step S423: The multi-agent agent participates in the negotiation, predicts the negotiation willingness of each participant, and calculates the adjustment matrix X = (x i ) M (This problem is a combinatorial optimization problem and can be solved using first-order algorithms such as gradient descent and Newton's method);

[0149] Step S424: After the proxy calculates the adjustment matrix X, it recalculates the global consensus level O after the adjustment matrix is ​​given by all parties. If O<ε, execute step S43; otherwise, determine whether the current single maximum negotiation round T is exceeded. If not, go to step S421, otherwise execute step S43 and output the adjustment matrix.

[0150] Furthermore, the step S423 (in the specific negotiation process, a large language model agent is used to conduct dialogue negotiation. Before the negotiation begins, the user will register the agent with the agent center, set the algorithm input conflict resolution candidate set CP and preference matrix ρ, initialize the role object, and then start the negotiation; the agent will start a negotiation dialogue with other agents according to the negotiation parameters set by the user and the pre-trained negotiation strategy, exchange negotiation opinions, and predict the negotiation willingness Q of each participant. The agent calculates the adjustment matrix X according to the negotiation goal, and then returns) includes the following steps:

[0151] Step S4231: Initialize Agent and negotiation parameters (create a thread for each participant, initialize the negotiation agent; register its own negotiation parameters, conflict resolution candidate set and preference matrix to the agent; initialize the agent dialogue role, need to set the agent role, set the large language model negotiation dialogue recognition role, and map it to the tokenizer);

[0152] Step S4232: Start negotiation; construct input, start polling and speaking according to the role, and output the answer format according to the prompt word; when the prompt word appears <partner_intention>, each negotiation agent outputs the predicted negotiation intention, and calculates the corresponding adjustment matrix X according to the objective function. The objective function is as follows:

[0153]

[0154] Among them: O represents the consensus degree of multiple parties on the quality parameters of collaborative services, W represents the selfish loss of participants;

[0155] Step S4233: Return the adjustment matrix X and return to step S424.

[0156] During the negotiation process, all agents will participate in the negotiation. The negotiation agent will adjust its service parameter expectations according to the proposals put forward by the other party, and maximize its own interests while considering the synergy. The negotiation results will save the best results. When the consensus level threshold or the negotiation round limit is reached, the negotiation stops and the best negotiation result is output.

[0157] Table 1 (Algorithm 1):

[0158] Table 1:

[0159]

[0160] Figure 1 illustrate: Figure 1 The overall flow chart of the algorithm in the present invention is as follows: first, service value modeling is performed to find conflicts, and then multiple rounds of negotiations are conducted to resolve conflicts. When conflicts are resolved, service quality parameters are output;

[0161] Figure 2 illustrate: Figure 2 The algorithm structure diagram of the present invention is mainly composed of three parts: service participant negotiation parameter initialization, service quality parameter initialization and service negotiation. The service quality parameter initialization is first performed to initialize the service quality parameter matrix of the participants, the negotiation adjacency matrix and the negotiation willingness matrix respectively; then the service negotiation is performed to resolve the conflicting service quality parameters;

[0162] Figure 3 illustrate: Figure 3 This is a flowchart of automated negotiation based on multi-party negotiation game, describing the specific process steps of automated negotiation;

[0163] Figure 4 illustrate: Figure 4 Flowchart for fine-tuning a negotiation agent, describing the process of fine-tuning a large language model to become a negotiation agent.

[0164] The above are only preferred specific implementation modes of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical solutions and inventive concepts of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for resolving conflicts between value and quality of complex multi-party collaborative services, characterized by: The method comprises the following steps: Step S1: Modeling value conflicts in multi-party collaborative services; Step S2: Negotiation agent fine-tuned based on a large language model; Step S3: Negotiate willingness perception and calculation methods; Step S4: Establish a complex multi-party collaborative service value conflict negotiation framework.

2. The method for resolving conflicts between value and quality of complex multi-party collaborative services according to claim 1, characterized in that: In step S1, the specific process of value conflict modeling in the multi-party collaborative service is: Suppose there are M service quality parameters and N service participants, then the VQC matrix is ​​expressed as: P=(p i,j ) M×N ,p i,j ∈[0,1] P l,t =(p i,j ) M×N , represents the QoS matrix of the i-th participant in the t-th negotiation round; p i,j represents the jth service quality parameter value of the i-th participant. Assume that each participant has its own evaluation function v(p) for the service quality parameter, V:P→R, p∈P, P represents the service quality parameter set, and V represents the value evaluation function; but represents the evaluation function of the ith participant on QoS, assuming that V is second-order differentiable with respect to P; Negotiate and adjust the matrix X l,t =(x i,j ) M×N , represents the adjustment made by the ith participant after the tth round of negotiation, and its QoS matrix in the t+1th round is represented by P l,t+1 =P l,t +X l,t ;P l,t and X l,t denote the service quality parameters and adjustment matrix in the lth iteration in the tth round respectively; Obviously, 0≤p i,j +x i,j ≤1, then the value of x is [-1,1]; Preference Matrix It represents the preference relationship between the kth participant and the QoS. The preference relationship is approximately calculated using the hierarchical analysis method, and then the square root method is used to approximate the preference relationship: Then normalize it: That is, the preference coefficient of the kth participant for each QoS can be calculated; The consensus level, that is, the similarity between the QoS matrices of each participant, is calculated as follows: Where: P l represents the QoS matrix of the i-th participant, X l represents the adjustment matrix of the i-th participant after negotiation; The participant's payoff w is Represents the participant's benefit from the current QoS; the selfish loss Δw k That is, the maximum benefit of the user minus the user benefit adjusted after t rounds of negotiations; Use the adjacency matrix to represent the QoS constraints: And=(and i,j ) M×M ,And i,j ∈{0,1} If and only if there is a constraint relationship between the i-th service quality parameter and the j-th service quality parameter, then e i,j =1; For each participant, the objective function of each round of negotiation is:

3. The method for resolving conflicts between value and quality of complex multi-party collaborative services according to claim 1, characterized in that: In step S2, the negotiation agent fine-tuned based on the large language model is specifically: Based on the historical negotiation data of users, a prompt is constructed to fine-tune the large language model. The specific fine-tuning method is as follows: Step S21: prompt project; Construct four prompt words: <intention>, <item>, <agreement>, <partner_intention>, which are used to integrate text information with participant preference ρ, conflict resolution candidate set CP, negotiation end mark, and opponent negotiation willingness prediction value into prompt; Step S22: construct input text <input_text>; inputs=<|im_start|>user+<input_text>+<|im_end|><|im_start|>assistant; Construct the continuation text <output_text>; outputs=<output_text>+<|im_end|>; Encode inputs and outputs into inputsId and labels through tokenizer, and build them into input and output of large language models; Step S23: load the pre-trained model and tokenizer from the large language model library, set the LoRAConfig parameters, and continue training using the LoRA weights pre-trained by the large language model; Step S24: token mapping, mapping the dialogue text roles and text content so that the dialogue data can be adapted to the pre-trained large language model; Step S25: Load the SFT trainer and fine-tune by using the labeled data. Load the inputsId and labels in step S24 into the trainer for training.

4. The method for resolving conflicts between value and quality of complex multi-party collaborative services according to claim 3, characterized in that: In step S3, the negotiation willingness perception and calculation method is specifically as follows: Model the preferences and value evaluation functions of multiple parties, i.e., ρ, V; where: ρ represents the preference coefficient, and V represents the value evaluation function; The proxy negotiation target is: U = μO + σ (Δw k +Δw -k ); Where: μ represents the initial consensus coefficient, σ represents the self-interest coefficient, both of which are adjusted hyperparameters; O represents the consensus degree of multiple parties on the collaborative service quality parameters, and W represents the self-interest loss of the participants; Define negotiation willingness Q k = {q i |i∈CP k },q∈[0,1], represents the negotiation willingness of participant k for each QoS in the conflict resolution candidate, and the negotiation willingness calculation formula is: Where: Φ i It indicates the negotiation willingness of the participants on the i-th service quality parameter.

5. The method for resolving conflicts between value and quality of complex multi-party collaborative services according to claim 1 or 4, characterized in that: In step S4, the specific steps of establishing a complex multi-party collaborative service value conflict negotiation framework are as follows: Step S41: Initialize parameters; initialize N services and M service quality parameters, then the QoS matrix is ​​P = (p i,j ) M×N , initialize the value evaluation function V = (v i,j ) M×N ; Initialize the preference matrix Initialize Agent parameters: Agent = <P, V, ρ>, set the maximum negotiation round L and the maximum single-round negotiation round T, set the initial consensus coefficient μ and the selfishness coefficient σ; Step S42: Conduct negotiation; Step S43: Obtain the adjustment matrix of each participant, select the optimal negotiation result by calculating the global consensus level O and the selfishness loss Δw, and update and retain the optimal negotiation parameters and negotiation results; Δw represents the selfish loss; Step S44: determine whether the current negotiation round satisfies l>lorO<ε*, where ε* is the set negotiation threshold super parameter; if so, output the optimal negotiation result and end the operation process; if not, execute step S45; Step S45: Re-adjust the negotiation parameters of the participants, adjust the QoS matrix P=P+X for the conflict resolution candidate set CP and the adjustment matrix X, adjust the consensus coefficient and the self-interest coefficient μ, σ, return to step S41, and start a new round of negotiation.

6. The method for resolving conflicts between value and quality of complex multi-party collaborative services according to claim 5, characterized in that: In step S42, the specific steps of the negotiation are as follows: Step S421: The Agent calculates the consensus level O of QoS according to the expected quality parameters; Where: At t = 0, X t =0, then the consensus level of each QoS is calculated; Step S422: According to the consensus level O i , retrieve the conflicting QoS set C = {P i |O(P i )>ε}, since the QoS constraint relationship is represented as a graph structure, the relevant conflict paths are searched according to the adjacency matrix E to retrieve the conflict resolution candidate set CP; Step S423: The multi-agent agent participates in the negotiation, predicts the negotiation willingness of each participant, and calculates the adjustment matrix X = (x i ) M ; Step S424: After the proxy calculates the adjustment matrix X, it recalculates the global consensus level O after the adjustment matrix is ​​given by all parties. If O<ε, execute step S43; otherwise, determine whether the current single maximum negotiation round T is exceeded. If not, go to step S421, otherwise execute step S43 and output the adjustment matrix.

7. The method for resolving conflicts between value and quality of complex multi-party collaborative services according to claim 6, characterized in that: Step S423 includes the following steps: Step S4231: Initialize Agent and negotiate parameters; Step S4232: Start negotiation; construct input, start polling and speaking according to the role, and output the answer format according to the prompt word; when the prompt word appears <partner_intention>, each negotiation agent outputs the predicted negotiation intention, and calculates the corresponding adjustment matrix X according to the objective function. The objective function is as follows: Among them: O represents the consensus degree of multiple parties on the quality parameters of collaborative services, W represents the selfish loss of participants; Step S4233: Return the adjustment matrix X and return to step S424.

Images