JOURNAL OF LATEX CLASS FILES, VOL. 14, NO.

8, AUGUST 2021 1

Generative AI for Game Theory-based Mobile Networking

Long He, Geng Sun, Dusit Niyato, Fellow, IEEE, Hongyang Du, Fang Mei, Jiawen Kang,
Mérouane Debbah, Fellow, IEEE, and Zhu Han, Fellow, IEEE

Abstract—With the continuous advancement of network tech- intricate strategic interactions typically entails high computa-
nology, various emerging complex networking optimization prob- tional complexity and communication costs among players.
lems opened up a wide range of applications utilizating of game Additionally, game theory models often oversimplify real-
theory. However, since game theory is a mathematical frame-
work, game theory-based solutions often require the experience world complexities for analytical convenience. However, this
and knowledge of human experts. Recently, the remarkable simplification may result in a model that fails to adequately
arXiv:2404.09699v1 [cs.GT] 15 Apr 2024

advantages exhibited by generative artificial intelligence (GAI) capture all relevant factors in practical scenarios, thus affecting
have gained widespread attention. In this article, we propose the accuracy and applicability of the model.
a novel GAI-enabled game theory solution that combines the Solving the aforementioned issues can be challenging since
powerful reasoning and generation capabilities of GAI to the
design and optimization of mobile networking. Specifically, we it traditionally requires a deep understanding of game theory
first outline the game theory and key technologies of GAI, and and solution algorithms, particularly for the newcomers or
then explore the advantages of combining GAI with game theory. those with interdisciplinary backgrounds. Therefore, explor-
Then, we briefly review the advantages and limitations of existing ing the formulation of game-theoretic models from human
research and demonstrate the potential application values of GAI language and obtaining a Nash equilibrium holds tremendous
applied to game theory in mobile networking. Subsequently, we
develop a game theory framework enabled by large language research potential, despite facing the following challenges.
models (LLMs) to realize this combination, and demonstrate the 1) Human language is not explicitly built into the founda-
effectiveness of the proposed framework through a case study in tions of game theory.
secured UAV networks. Finally, we provide several directions for 2) Human language often contains incomplete or vague
future extensions.
information, making it difficult to precisely define all
Index Terms—GAI agent, pluggable LLM module, RAG, game the necessary parameters of a game-theoretic model.
theory, Nash equilibrium.
In particular, with the development of mobile networking,
game theory plays an increasingly important role in network
I. I NTRODUCTION optimization, network configuration and management, as well
as network security and privacy. For example, a game theory-
Game theory studies mathematical models of strategic inter- based method [1] was proposed to optimize user allocation
actions among interdependent rational participants (referred to and transmission power to improve the performance of mobile
as players), who make decisions based on their self-interests edge computing networks. Therefore, exploring solutions to
while their actions influence one another. Given its highly formulate game-theoretic models from human language and
abstract representation of real-life situations, game theory has obtain their Nash equilibrium is also crucial to improve
emerged as a standard analytical tool in various fields to network performance and facilitate network management.
model and comprehend complex systems involving multiple Fortunately, generative artificial intelligence (GAI) tech-
interacting agents. Employing game theory in the real world niques, such as large language models (LLMs) combined
requires both constructing a suitable abstract game model with retrieval-augmented generation (RAG), offer a promising
(i.e., a binding from the cases to game-theoretic language) solution to address the aforementioned challenges. LLMs,
and solving the game (i.e., running an algorithm on the such as GPT-41 , have demonstrated remarkable capabilities in
constructed model to extract the Nash equilibrium). Moreover, understanding and generating human-like text. For example,
analyzing complex game scenarios with numerous players and the authors in [2] proposed a systematic simulation framework
utilizing LLM agents for game theory research, and the simu-
L. He and F. Mei is with the College of Computer Science and lation results demonstrated that the LLMs have the ability to
Technology, Jilin University, Changchun 130012, China (e-mail: he-
long0517@foxmail.com, meifang@jlu.edu.cn). imitate complex human strategic behaviors in socio-economic
G. Sun is with the College of Computer Science and Technology, Jilin Uni- contexts. Furthermore, RAG models are able to enhance the
versity, Changchun 130012, China, and also with the College of Computing reasoning capabilities of LLMs and provide more context-
and Data Science, Nanyang Technological University, Singapore 639798 (e-
mail: sungeng@jlu.edu.cn). aware solutions by leveraging external knowledge sources,
D. Niyato and H. Du are with the College of Computing and Data such as online encyclopedias or databases. Specifically, the
Science, Nanyang Technological University, Singapore 639798 (e-mail: advantages of using GAI in solving the challenges of formu-
hongyang001@e.ntu.edu.sg, dniyato@ntu.edu.sg).
J. Kang is with the School of Automation, Guangdong University of lating game models from human language and to obtain its
Technology, Guangzhou, China (e-mail: kavinkang@gdut.edu.cn). Nash equilibrium in networking are as follows:
M. Debbah is with Khalifa University of Science and Technology, P O Box
• Learning from Data: Generative AI techniques allow
127788, Abu Dhabi, UAE (e-mail: merouane.debbah@ku.ac.ae).
Z. Han is with the University of Houston, Houston TX 77004, USA, and for a data-driven approach to game model formulation.
also with the Department of Computer Science and Engineering, Kyung Hee
University, Seoul 446701, South Korea (e-mail: zhan2@uh.edu). 1 https://openai.com/blog/chatgpt/
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 2

Specifically, LLMs can be trained on large-scale datasets, A. Game Theory

enabling them to learn from a wide range of game 1) What is game theory about? Game theory is a mathe-
descriptions and strategies. This data-driven learning can matical model and theoretical tool that studies the interaction
help in capturing the patterns, common behaviors, and of rational decision-makers (called players) in the presence of
rationality assumptions exhibited by game players, their conflicts of interest. For example, the time required for a car
strategies and payoffs, enhancing the accuracy and real- driver to get home typically depends not only on the route
ism of the formulated game models [3]. they choose but also on the decisions made by other drivers.
• Incorporating External Knowledge: RAG models, For such scenarios, game theory provides a framework for
which combine LLMs with retrieval-based methods, can analyzing the strategic behavior of rational decision-makers
leverage external knowledge sources to augment the gen- and predicting the outcomes of their decisions. In game
eration process. This is particularly useful when dealing theory, non-cooperative game theory (NGT) is one of the most
with incomplete or vague language descriptions. Further- elemental and important branches, which has been widely
more, possessing the ability to learn from and access applied in various fields. Specifically, it typically implies a
external data sources enhances adaptability. For example, competitive nature where players are not allowed to engage in
we can leverage game-theoretic models derived from eco- enforced cooperation. Therefore, in this brief exposition, this
nomics and effectively apply them to resource allocation paper focuses on a review of NGT.
in networking. By accessing relevant information from 2) How to present a game? The representation of game
databases, RAG models can enhance the accuracy, context theory typically involves mapping a real-world game problem
relevance, and credibility of game model formulation and or scenario into a mathematical model. For NGT, as shown
solution. Moreover, they also improve the trustworthiness in Fig. 1, the strategic form is the most basic form, which
and interpretability of the formulated model. typically consists of three fundamental elements that are
• Scalability and Computational Efficiency: LLMs and players, strategies and payoffs.
RAG models are designed with parallel processing capa- • Players: Players are the participants or decision-makers
bilities, making them computationally efficient for han- involved in the game. A player can be a person, a
dling large-scale game models. This is crucial when machine, or more generally, any decision-making entity.
dealing with complex games involving numerous players Each player has a distinct set of preferences and objec-
and intricate interactions. Moreover, the efficiency of tives, and their choices or actions impact the outcomes
these models enables faster analysis and optimization, of the game.
while reducing the computational burden associated with • Strategies: Strategies represent the available choices or
obtaining Nash equilibrium. actions that players can take in the game. A strategy of
Motivated by these, this paper attempts to provide a the player is their plan of action, which determines how
forward-looking research that delves into the integration of they will act or respond to the actions of other players.
GAI and game theory for networking optimization. To the best • Payoffs: Payoffs in game theory represent the objectives
of the authors’ knowledge, the synergy between GAI and game or utilities that players receive based on the strategies
theory for networking optimization has still been an open issue they choose and the actions taken by other players. The
thus far. The contributions of this paper are summarized as payoffs determine the desirability of different outcomes
follows: for each player, influencing their decisions and strategies.
1) We overview the game theory and introduce different 3) How to analyze the outcome of a game? For NGT,
types of GAI models. Building on this, we provide the Nash equilibrium is a fundamental solution concept for
comprehensive insights into the integration of GAI and predicting the outcome of the game. Specifically, it represents
game theory. a strategy profile in which no player can improve their own
2) We explore a few seminal applications of LLMs and payoff by changing their strategy while keeping the strategies
game theory in networking from different perspectives, of the other players unchanged. In other words, Nash equilib-
providing the guideline on how to integrated these three rium is a stable state where the players lack the incentive to
important technologies to solve practical problems. actively change their strategies.
3) We propose an LLM combined with RAG framework for
networking optimization. Simulation results based on a B. Generative AI Models
real-world networking optimization case study validate GAI is a collection of various AI technologies and models
the effectiveness of the proposed framework. that can learn patterns and regularities from data and then
generate new data. The foundation of GAI mainly includes
the following key technologies.
II. OVERVIEW OF G AME T HEORY AND G ENERATIVE AI LLMs: LLMs are AI models based on Transformers, which
can understand and generate human language through training
In this section, we first introduces the basics of game theory of large-scale text data. Due to their versatility and robust
and several typical AI technologies that forming the foundation generalization capabilities, LLMs can undertake diverse tasks,
of GAI. Following that, we generally explain the potential including text classification, question answering systems, and
support of GAI for game theory. text generation, etc.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 3

Fig. 1. The non-cooperative game framework contains multiple players and an environment in which players interact. Each player has a strategy space that
represents the available actions and a utility function that evaluates the player’s payoff from taking a certain action. Additionally, the objective of each player
is to take actions that maximize their own payoff. The environment represents the medium through which players interact, that matches an action profile to a
payoff profile. The flow of player interaction is as follows: in the first interaction (Step 1), each player takes an action from their strategy space and receives
the corresponding payoff. In subsequent interactions (Step i), each player updates their action based on the observed strategies of other players to improve
their own payoff. When the actions of all players no longer change, a Nash equilibrium is obtained.

RAG: RAG is an advanced natural language processing Due to the ability to generate high-quality data and the
technique that integrates information retrieval and generative advantage of modeling complex data distributions, GDMs have
models to enhance the quality and relevance of generated text. been employed in material design, time series forecasting, and
Furthermore, when combined with RAG, LLMs can better text-to-image generation, etc.
understand user queries and generate text results that more
accurately fulfill user needs by accessing external databases. C. GAI for Game Theory
Generative Adversarial Networks (GANs): GANs are a Traditional AI technologies such as deep learning and re-
type of neural network architecture with custom adversarial inforcement learning, have been applied in various successful
learning objectives, and they consists of two main parts that cases in the context of game theory. For example, in [9], the
are the generator and discriminator. The goal of GANs is authors formulated the task offloading problem in vehicular
to train a generator network to generate samples that are edge computing networks as an exact potential game and
similar to real data, and to train a discriminator network utilized a multi-agent distributed distributional deep determin-
to distinguish between samples generated by the generator istic policy gradient approach to attain the Nash equilibrium.
and real samples. Therefore, GANs have high performance in In [10], the authors investigated the dynamic resource trading
generating high-quality multi-modal content, which has led to problem of multi-UAV-assisted industrial IoT networks, which
a large number of applications such as 3D object generation, is modeled as an extended stochastic game. Then, the authors
and image processing. proposed a multi-agent deep reinforcement learning algorithm
Generative Diffusion Models (GDMs): Inspired by non- to solve the formulated stochastic game. In [11], a Stackelberg
equilibrium thermodynamics, GDMs comprise of two in- game was employed to model real-time UAV twin migration
terconnected processes, i.e., a predefined forward process for the emerging UAV metaverse system. The authors designed
that transforms the data distribution into a more simplistic a tiny multi-agent deep reinforcement learning algorithm to
prior distribution (typically a Gaussian distribution), and a approximate the game equilibrium. However, traditional AI
corresponding reverse process that employs a trained neural techniques applied in the context of game theory still en-
network to gradually reverse the effects of the forward process. counter the following challenges:
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 4

TABLE I
OVERVIEW OF SURVEY PAPERS ON THE INTEGRATION OF LLM S AND GAME THEORY

Survey Contributions Emphasis

Evaluate the degree of rationality of LLMs in establishing clear desires, refining beliefs
[3]
about uncertainty, and taking optimal actions in the context of game theory. Focus on LLM’s reasoning capabilities.
Analyze the reasoning capabilities of LLMs in competitive environments through game-
[4]
theoretic tasks that require pure logic and strategic reasoning to compete with opponents.
Explore the use of LLM in strategic game experiments, specifically the ultimatum game Focus on the application of LLM in strategy
[5]
and the prisoner’s dilemma. game experiments.
Investigate multi-agent LLMs from a game theoretic perspective, where agents collab-
[6]
oratively solve tasks in competitive environments. Focus on LLM’s collaborative capabilities.
Explore the potential for LLM agents to spontaneously establish collaborative relation-
[7]
ships in the absence of explicit instructions.
Present one possible binding from dialogue to game theory as well as generalizations
[8] Focus on LLM-driven game generation.
of existing equilibrium finding algorithms to this setting.

• Need for Formal Formulation: Traditional AI tech- theory and mathematical method. LLMs combined with
niques typically require a formal formulation of a game- RAG has powerful learning and generation capabilities,
theoretic scenario, i.e., mathematical representations, in- which can formulate accurate mathematical models that
cluding the definition of players, actions, and payoffs. are suitable for gaming scenarios through extensive learn-
However, this formalization process can be challenging, ing of knowledge in related fields.
especially for complex and real-world games where the • Nash Equilibrium Search: Obtaining the outcomes of
rules may not be explicitly defined or known. game theoretic models, i.e., the Nash equilibrium, is often
• Limited Information and Communication: In certain computationally complex. LLMs combined with RAG
games, players may lack complete information about the can extensively explore the strategy space in game theory
game state or the strategies of other players. Traditional and infer possible Nash equilibrium points. Moreover,
AI techniques may struggle to handle such incomplete LLM can enhance reasoning capabilities by learning ex-
or asymmetric information settings, as they rely on the isting methods, which makes obtaining Nash equilibrium
assumption of complete information in their learning and more effective.
decision-making processes. Lesson Learned: The exploration of LLMs combined with
• Computational Complexity: Obtaining Nash equilib- RAG in game theory underscores the critical role of rich
rium for large games can incur significant computational knowledge base and powerful reasoning abilities in simpli-
costs. The search space for Nash equilibrium exponen- fying the utilization of game theory. Specifically, traditional
tially increases exponentially with the number of players game theory typically demands researchers to possess ex-
and strategies [12]. Traditional AI techniques may strug- tensive experience and a deep understanding of game the-
gle to efficiently explore the strategy space and converge ory to provide effective insights into various complex game
to a Nash equilibrium in such cases. scenarios. By engaging in self-learning from large-scale data,
LLMs combined with RAG provide a promising solution for LLM-based game theory successfully encapsulates the afore-
addressing the abovementioned challenges faced by traditional mentioned complex process, thereby effectively reducing the
AI. In the following, we elaborate on the advantages of challenges faced in the applications of traditional game theory.
LLMs combined with RAG in game scenarios from three For example, in the social sciences domain, the integration
perspectives, focusing on game scenario recognition, game of LLMs with game theory provides a valuable tool for
theoretic model formulation, and Nash equilibrium search. experimental research and social simulations [5]. In the field of
economics, LLMs have the potential to revolutionize microe-
• Game Scenario Recognition: Game theory has been conomic analysis by simulating strategic interactions among
widely used in various fields to analyze complex real-life economic agents, which can lead to a better understanding of
situations. Applying game theory to case studies involves market behaviors and policy impacts [4].
initially identifying and extracting key game theory-
related concepts from the context of these cases, includ- III. LLM S - ENABLED G AME T HEORY IN N ETWORKING
ing decision-makers, their strategies, and their payoffs. In this section, we first provide a comprehensive review of
LLMs, trained on large amounts of textual data, possess the existing literature on the integration of LLMs and game
the capability to comprehend natural language, thereby theory, and analyze the contributions and limitations of the
accurately identifying game theory-related concepts from current works. Subsequently, we present some applications of
natural language descriptions of real-life cases. LLMs-enabled game theory in networking and analyze the
• Game Theoretic Model Formulation: Game theoretic combination of LLM and game theory for multi-agent network
model formulation involves mapping a game scenario to a design .
mathematical model to facilitate theoretical analysis. Due
to the diversity and complexity of game scenarios, it is A. Overview of Survey Papers
challenging to formulate suitable mathematical models, Several studies have explored the integration of LLMs and
requiring a broad and in-depth understanding of game game theory, as shown by Table I. However, these studies
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 5

exhibit certain limitations. First, existing research mainly trols and protocols to ensure the confidentiality of network
focuses on exploring the reasoning capabilities of LLMs in resources. Game theory can be used to analyze the interaction
the context of game theory. However, the potential advantages between cyber attacks and defense strategies. By modeling
of applying LLMs to game theory scenarios have not been well the game between attackers and defenders, optimal defense
studied. Second, existing research is mainly based on typical strategies can be derived, thereby improving the network
game theory cases with pre-defined game theory models. security performance. Traditional game theory approaches
However, in various application fields of game theory such as typically formulate optimal defense strategies based on the
the networking, it is challenging to map actual game scenarios static network environments. However, in dynamic network
into appropriate game theory models. Given the potent natural environments, the strategies and behaviors of attackers may
language understanding and generation capabilities of LLMs, evolve over time, which is challenging for dealing with. In
leveraging them to map game scenarios to game models is this case, LLM-based game theory can monitor the changing
a great value topic, which is still lacking research. Finally, network security threats in order to adjust decision-making
existing studies do not consider the combination of LLMs with strategies to deal with the new attacks.
RAG. However, RAG is crucial to improving the capabilities
of LLMs. Therefore, the combination of LLMs and RAG needs C. Combining LLM and Game Theory for Multi-Agent Net-
further exploration. To this end, this paper proposes a frame- work Design
work that applies LLMs combined with RAG to game theory
scenarios in networking to make up for the shortcomings of The convergence of LLM, mobile networks, and multi-agent
existing research. systems represents groundbreaking synergies that are leading
to the emergence of multi-agent LLM network architectures,
B. Potential Applications of LLMs-enabled Game Theory in bringing huge potential for the design of future mobile net-
Networking works [6]. This integration harnesses the power of collective
Spectrum and Interference Management: Spectrum and intelligence and paves the way for self-managing networks.
interference management are important issues in the field However, in a multi-agent LLM network, each LLM agent
of wireless communication networks, involving the effec- not only needs to pursue its individual objectives but also
tive use of limited spectrum resources and power control collaborate with other agents to achieve the collective goals
to reduce wireless interference and improve network per- of the entire network. This balance is crucial in designing
formance. Unlike traditional centralized management solu- efficient multi-agent LLM networks. In this context, game
tions with high computing load, game theory can be used theory can be used to model and analyze the behavior of
to formulate distributed management solutions to effectively multi-agent LLM systems. This includes finding equilibria
alleviate the computing load. However, traditional game theory among agents, enabling them to serve individual objectives
approaches require frequent information interactions between while effectively collaborating to achieve network-level goals.
the service users and central controller to obtain real-time Furthermore, multi-agent reinforcement learning (RL) can
complete information, which results in not only substantial further model the interactions among agents to learn optimal
time taken by users but also heavy computation and com- collaborative strategies and communication protocols among
munication overhead. In this case, the LLMs-enabled game distributed LLM agents. In doing this, the communication
theory approach can use pre-trained models to approximate costs among agents can be reduced.
and learn strategic interaction behaviors in games to predict
user decisions, which supplements the integrity of information IV. LLM S - ENABLED G AME T HEORY F RAMEWORK
to reduce the communication overhead. AUGMENTED BY RAG FOR M OBILE N ETWORK
Resource Allocation: Resource allocation involves the ef- O PTIMIZATION
fective allocation and management of limited network re- In the section, we first propose an LLM-enabled game the-
sources, such as communication resources, computing re- ory framework supported by RAG for network optimization.
sources, and storage resources, to meet the needs of network Then, we conduct a case study on UAV secure communication
users. The limited nature of network resources usually leads optimization to demonstrate the effectiveness of the proposed
to competition among network users, while game theory framework.
provides an analytical framework to understand the decision-
making and strategy selection process among users to obtain
effective resource allocation solutions. However, traditional A. Motivations
game theory approaches to evaluate user satisfaction usually Network optimization is crucial for enhancing the perfor-
rely on pre-defined models, which are inflexible and inaccurate mance, efficiency, and user experience of mobile networks.
in practical scenarios. LLMs-enabled game theory approaches Given the complexity of network optimization problems, game
can develop a more accurate model for evaluating the user theory has been adopted and has played a significant role in
satisfaction by combining historical information and user analyzing various network optimization issues. Nevertheless,
feedback, which can improve the effectiveness of decision- for network designers, employing game theory to analyze
making. networking optimization problems pose challenges that typ-
Network Security: Security constitutes a critical aspect of ically necessitate a solid grasp of mathematical theory and a
network management, which involves various security con- profound comprehension of game theory principles.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 6

Fig. 2. The LLM-enabled game theory framework. The framework is based on a layered architecture consisting of an input layer, an augmented layer, a
decision layer, and an output layer. The input layer captures user input queries. The augmented layer employs RAG technology to enhance user queries. The
decision layer utilizes a pluggable LLM to generate responses. The output layer returns the generated results to the user.

Inspired by LLM’s outstanding capabilities, we propose an through embedding, the system can learn semantic simi-
LLM-enabled game theory framework to address the above- larities and correlations between inputs.
mentioned challenges. The proposed framework automates • Augmented Layer: The augmented layer is primarily
the mathematical formulation and game theory analysis of implemented through the RAG technique, which is shown
networking optimization problems through simple interactions in detail in Part A of Fig. 3. The RAG can improve
with network designers, which effectively addresses the chal- the quality, accuracy, and relevance of the generated
lenges of applying game theory and enhances the process answers of the system by providing relevant information
of networking optimization. In addition, the framework can related to the user requests [13]. The augmented layer
improve the accuracy of model formulation by incorporating contains a large-scale knowledge base to store rich ma-
relevant domain knowledge learned from a comprehensive terial information, such as academic papers from IEEE
knowledge base. Xplore. Specifically, RAG first acquires rich knowledge
by loading the knowledge base, and then divides the
knowledge into knowledge chunks that are suitable for
B. Proposed Framework embedded search. These knowledge chunks are then
transformed into dense vector representations by using
As shown in Fig. 2, our proposed framework follows a embedding techniques and stored in a vector database.
layered architecture, which consists of four layers that are the Upon receiving the vector representation of the user
input layer, augmented layer, decision layer, and output layer. request, RAG retrieves relevant information from the
• Input Layer: The input layer receives requests from vector database and combines it with the user requests to
users, where multi-modal input allows the users to use form prompts, which are then sent to the decision-making
different types of data as input, such as text, image, layer for inference and decision-making execution.
voice, and video. After receiving a request of the user, • Decision Layer: The decision layer adopts a plugin archi-
prompt engineering guides the system to generate the tecture to select the appropriate LLMs such as New bing,
desired output by designing and constructing appropriate Bard, and GPT4, to formulate decisions. As shown in Part
input prompts. Then, the input request of the user is B of Fig. 3, LLM formulates a response based on the
transformed into a dense vector representation that the query and the selected chunks from the augmented layer.
system can understand and process through embedding The decision layer mainly relies on the reasoning and
to facilitate downstream model processing. Furthermore,
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 7

Fig. 3. The operation flow of RAG and LLM. In Part A, documents are loaded from the knowledge base, segmented into chunks, encoded into vectors, and
stored in the vector database. Then, RAG retrieves the K most relevant chunks to the query of user based on semantic relevance from the vector database.
In Part B, the original query and the retrieved chunks are inputted together into LLM to generate the final response.

decision-making capabilities of LLMs based on large- problem, use an appropriate game theory model to analyze
scale data training. the problem, and design effective algorithms to solve it,
• Output Layer: The output layer presents the decision which relies on rich professional knowledge. With the help
results given by the decision layer to the users. of the proposed framework, the optimization process can be
automatically implemented through three rounds of interaction
C. Case Study: UAV Secure Communication Optimization between the network designer and the GAI agent. Therefore,
In the section, we use a UAV secure communication opti- the proposed framework can effectively solve the challenge
mization case to demonstrate the proposed framework. faced by the traditional optimization process.
Scenario Description: As shown in Fig. 4, consider a UAV Evaluation Results: The evaluation result module (a) of
communication network consisting of N legitimate UAV users, Fig. 4 shows the convergence of the algorithm generated by the
a destination ground base station, a friendly UAV jammer, proposed framework (i.e., GAI algorithm). From this figure,
and a malicious ground eavesdropper. These legitimate UAV we can see that the GAI algorithm reaches a stable state,
users communicate with the destination base station through N i.e., Nash equilibrium, after five iterations. The simulation
orthogonal channels while facing the risk of eavesdropping by results illustrate that the proposed framework can obtain a
the eavesdropper. The friendly UAV jammer has the capability Nash equilibrium solution to the game theory problem. More-
to transmit cooperative jamming signals simultaneously over over, the evaluation result module (b) of Fig. 4 shows the
these channels to enhance the overall secrecy rate against the performance of the GAI algorithm with respect to different
eavesdropper. The optimization objective is to maximize the number of channels, where the expert algorithm indicates that
aggregate security rate of all legitimate users through optimal the algorithm designed by network experts can be regarded
power allocation of the friendly jammer across each channel. as the optimal algorithm, and the EPR algorithm indicates
Framework Configuration: In the framework configura- that the average power allocation strategy is a benchmark
tion, we call the ChatGPT 4 model through the OpenAI API algorithm. Correspondingly, we can observe that the GAI
to implement the pluggable LLM module. Furthermore, the algorithm has the same performance as the expert algorithm.
RAG module is built based on the LangChain. The evaluation results demonstrate the effectiveness of the
Framework Operation Flow: The LLM interactive mod- proposed framework.
ule in Fig. 4 shows the optimization flow of the proposed
framework to solve the UAV secure communication opti- V. F UTURE D IRECTIONS
mization problem. In traditional optimization process, a net- In this section, we present three major future directions for
work designer needs to formulate a mathematical optimization the enhancement and extension of GAI-enabled game theory.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 8

Fig. 4. The experiment results of the wireless secure communication optimization case. The scenario module presents a system model of the considered scenario.
The interaction module showcases the functionality of the proposed framework. The RAG augmentation module emphasizes the operational mechanism of
RAG. The evaluation results module demonstrate the effectiveness of the proposed framework.

A. GDM-enhanced Non-Cooperative Game VI. C ONCLUSION

For non-cooperative game scenarios in networking with
In this article, we explored the integration of LLMs with
numerous players and complex decision spaces, obtaining a
game theory, and concluded the advantages of LLM for game
Nash equilibrium is often challenging computationally due to
theory. Following this, we presented the potential applications
the need to search through a high-dimensional, exponentially
of LLM combined with game theory in networking. Based on
large space. GDM has outstanding capabilities in modeling
these, we proposed a framework that utilized RAG to combine
complex data distribution and data generation [14]. With the
LLM with game theory, aiming to achieve the intelligence
help of GDM, it is possible to enhance the search efficiency
of game theory in network applications. We conducted a
for Nash equilibrium and greatly reduce the computational
case study on UAV secure communication optimization to
complexity of complex non-cooperative games.
validate the effectiveness of the proposed framework. Finally,
we discussed several potential research directions for future
B. Performance Evaluation of GAI-enabled Game Theory
extensions.
It is crucial to design a comprehensive performance evalua-
tion model to evaluate the performance of GAI combined with
game theory. This is because effective evaluation models can R EFERENCES
provide directions for guiding and improving the GAI-enabled
game theory framework. Furthermore, intelligent evaluation [1] P. Lai, Q. He, F. Chen, M. Abdelrazek, J. Hosking, J. Grundy, and
Y. Yang, “Online user and power allocation in dynamic NOMA-based
criteria should be designed to evaluate GAI-enabled game mobile edge computing,” IEEE Trans. Mob. Comput., vol. 22, no. 11,
theory instead of human evaluation in the future. pp. 6676–6689, November 2023.
[2] S. Mao, Y. Cai, Y. Xia, W. Wu, X. Wang, F. Wang, T. Ge, and
F. Wei, “Alympics: LLM agents meet game theory – Exploring strategic
C. GAI-enabled Game Theory for Future Networks decision-making with AI agents,” arXiv preprint arXiv:2311.03220,
As a key technology for future 6G wireless communications, 2023.
[3] C. Fan, J. Chen, Y. Jin, and H. He, “Can large language models serve as
the space-air-ground integrated network (SAGIN) has been rational players in game theory? A systematic analysis,” in Proc. AAAI
proposed to provide seamless communication coverage and Conf. Artif. Intell., vol. 38, no. 16, March 2024, pp. 17 960–17 967.
high-speed connections [15]. Due to the inclusion of hetero- [4] J. Duan, R. Zhang, J. Diffenderfer, B. Kailkhura, L. Sun, E. Stengel-
Eskin, M. Bansal, T. Chen, and K. Xu, “Gtbench: Uncovering the
geneous network technologies, i.e., ground-based network, air- strategic reasoning limitations of LLMs via game-theoretic evaluations,”
based network, and space-based network, the deployment and arXiv preprint arXiv:2402.12348, 2024.
network management of SAGIN becomes more challenging. [5] F. Guo, “GPT agents in game theory experiments,” arXiv preprint
The vast potential of GAI-enabled game theory in networking arXiv:2305.05516, 2023.
[6] H. Zou, Q. Zhao, L. Bariah, M. Bennis, and M. Debbah, “Wireless
is anticipated to synergize with the emerging SAGIN technol- multi-agent generative AI: From connected intelligence to collective
ogy, catalyzing the advancement of SAGIN. intelligence,” arXiv preprint arXiv:2307.02757, 2023.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 9

[7] Z. Wu, S. Zheng, Q. Liu, X. Han, B. I. Kwon, M. Onizuka, S. Tang,

R. Peng, and C. Xiao, “Shall we talk: Exploring spontaneous collab-
orations of competing LLM agents,” arXiv preprint arXiv:2402.12327,
2024.
[8] I. Gemp, Y. Bachrach, M. Lanctot, R. Patel, V. Dasagi, L. Marris, G. Pil-
iouras, and K. Tuyls, “States as strings as strategies: Steering language
models with game-theoretic solvers,” arXiv preprint arXiv:2402.01704,
2024.
[9] X. Xu, K. Liu, P. Dai, F. Jin, H. Ren, C. Zhan, and S. Guo, “Joint task
offloading and resource optimization in NOMA-based vehicular edge
computing: A game-theoretic DRL approach,” J. Syst. Archit., vol. 134,
p. 102780, January 2023.
[10] M. S. Abegaz, H. N. Abishu, Y. H. Yacob, T. A. Ayall, A. Erbad, and
M. Guizani, “Blockchain-based resource trading in multi-UAV-assisted
industrial IoT networks: A multi-agent DRL approach,” IEEE Trans.
Netw. Serv. Manag., vol. 20, no. 1, pp. 166–181, March 2023.
[11] J. Kang, Y. Zhong, M. Xu, J. Nie, J. Wen, H. Du, D. Ye, X. Huang,
D. Niyato, and S. Xie, “Tiny multi-agent DRL for twins migration
in UAV metaverses: A multi-leader multi-follower Stackelberg game
approach,” IEEE Internet Things J., pp. 1–1, 2024.
[12] Y. Babichenko, S. Barman, and R. Peretz, “Simple approximate equi-
libria in large games,” in Proc. ACM EC, 2014, pp. 753–770.
[13] A. Asai, Z. Wu, Y. Wang, A. Sil, and H. Hajishirzi, “Self-RAG: Learning
to retrieve, generate, and critique through self-reflection,” arXiv preprint
arXiv:2310.11511, 2023.
[14] H. Cao, C. Tan, Z. Gao, Y. Xu, G. Chen, P.-A. Heng, and S. Z. Li, “A
survey on generative diffusion models,” IEEE Trans. on Knowl. Data
Eng., pp. 1–20, 2024.
[15] J. Liu, Y. Shi, Z. M. Fadlullah, and N. Kato, “Space-air-ground integrated
network: A survey,” IEEE Commun. Surv. Tutorials, vol. 20, no. 4, pp.
2714–2741, Fourthquarter 2018.