Multi-Agent-Based Service Composition Using Integrated Particle-Ant Algorithm in the Cloud

Abstract

The increasing complexity and scale of service-oriented architectures in cloud computing have heightened the demand for intelligent, decentralized, and adaptive service composition techniques. This study proposes an advanced framework that integrates a Multi-Agent System (MAS) with a novel hybrid metaheuristic optimization method, the Integrated Particle-Ant Algorithm (IPAA), to achieve efficient, scalable, and Quality of Service (QoS)-aware service composition. The IPAA dynamically combines the global search capabilities of Particle Swarm Optimization (PSO) with the local exploitation strength of Ant Colony Optimization (ACO), thereby enhancing convergence speed and solution quality. The proposed system is structured into three logical layers—agent, optimization, and infrastructure—facilitating autonomous decision-making, distributed coordination, and runtime adaptability. Extensive simulations using a synthetic cloud service dataset demonstrate that the proposed approach significantly outperforms traditional optimization methods, including standalone PSO, ACO, and random composition strategies, across key metrics such as utility score, execution time, and scalability. Moreover, the framework enables real-time monitoring and automatic re-optimization in response to QoS degradation or Service-Level Agreement (SLA) violations. Through decentralized negotiation and minimal communication overhead, agents exhibit high resilience and flexibility under dynamic service availability. These results collectively suggest that the proposed IPAA-based framework provides a robust, intelligent, and scalable solution for service composition in complex cloud computing environments.

1. Introduction

Cloud computing has transformed the landscape of service-oriented architectures (SOAs). It provides scalable, on-demand, and cost-effective access to distributed services. Within this context, the concept of service composition—combining multiple atomic or composite services into value-added, higher-level services—has become increasingly critical [1,2,3]. Traditionally, service compositions have been orchestrated through centralized systems or client-driven coordination mechanisms. However, as the scale and diversity of services expand, these traditional approaches exhibit inherent limitations. They struggle to ensure scalability, flexibility, and real-time responsiveness in dynamic environments [4,5].

A central challenge in modern service composition lies in selecting optimal combinations of services from a vast and heterogeneous pool of providers. Each provider offers distinct functionalities, cost models, and Quality of Service (QoS) attributes. With the exponential growth in services and providers, manual selection by human operators or static algorithms has become impractical, particularly under dynamic and time-sensitive user requirements [6,7]. The challenge is further compounded in multi-tenant cloud environments characterized by fluctuating workloads and variable service availability [8].

To address these issues, Multi-Agent Systems (MASs) have emerged as a promising paradigm. Agents are autonomous, interactive software entities capable of making local decisions and collaborating with others to solve complex problems in distributed settings [9]. The MAS offers key advantages in cloud-based service composition, including distributed computation, improved fault tolerance, scalability, and real-time decision-making. Nevertheless, effective decision-support mechanisms remain essential for navigating vast search spaces and identifying optimal service compositions under QoS constraints [10,11].

In recent years, metaheuristic optimization algorithms have been integrated into MAS frameworks to enhance decision-making capabilities. Among these, Particle Swarm Optimization (PSO) and Ant Colony Optimization (ACO) are widely applied due to their global search capability and adaptability to complex optimization problems [12,13]. Yet, each technique has its limitations; PSO is prone to premature convergence, whereas ACO often suffers from slow convergence in large-scale scenarios [14].

To overcome these drawbacks, this paper introduces the Integrated Particle-Ant Algorithm (IPAA), a hybrid optimization approach that combines the global exploration ability of PSO with the local refinement and positive feedback mechanisms of ACO. IPAA is embedded within a multi-agent cloud platform, resulting in an intelligent, autonomous, and adaptive framework for service composition.

This study makes the following key contributions:

• We design a cloud-native multi-agent architecture that supports decentralized service composition with real-time negotiation and collaboration between agents.
• We develop the IPAA, a hybrid optimization technique that efficiently searches the solution space while minimizing execution time and enhancing service accuracy.
• We evaluate the proposed system under various scenarios and demonstrate that it outperforms standalone metaheuristic algorithms and traditional methods in terms of QoS, computational efficiency, and scalability.

In summary, this work bridges the gap between autonomous agent-based service coordination and high-performance optimization in cloud environments. By integrating intelligent agent behavior with adaptive heuristic techniques, we provide a scalable and intelligent solution to the ever-growing demands of modern service composition.

2. Related Work

The challenge of automated service composition has been extensively studied in both academic and industrial contexts. Initial approaches relied heavily on workflow-based or rule-based systems, which required static configurations and manual orchestration of services [15,16]. These early methods, while effective in small-scale or well-controlled environments, lack the flexibility and adaptability required in dynamic and large-scale cloud ecosystems [2].

To overcome these limitations, researchers began exploring semantic web technologies, where ontologies and reasoning engines were employed to match and compose services based on their functional and non-functional descriptions [17]. While this approach improved automation and accuracy in service selection, semantic-based models often suffer from computational complexity and are heavily dependent on the completeness of service annotations, which limits their scalability and practical deployment [4,18].

With the advent of cloud-native architectures and microservice-based platforms, attention has shifted toward more adaptive, scalable, and intelligent composition strategies. Among these, the use of Multi-Agent Systems (MASs) has gained considerable traction. In this paradigm, agents autonomously represent service providers or consumers, interact with other agents, and collaborate to achieve service composition goals [19]. Their inherent reactive and proactive behaviors allow them to dynamically discover services, negotiate terms, and adapt to runtime changes in network or system conditions [20]. However, a MAS alone often falls short in producing globally optimal compositions, particularly when constrained by complex QoS requirements such as latency, availability, reliability, and cost [6].

To enhance the decision-making capabilities of the MAS, researchers have increasingly integrated metaheuristic optimization algorithms into agent frameworks. Two of the most commonly used techniques are ACO and PSO, owing to their simplicity, flexibility, and effectiveness in solving NP-hard problems [12,13]. ACO mimics the foraging behavior of ants and is especially effective in pathfinding and discrete optimization scenarios. In contrast, PSO, inspired by the social dynamics of bird flocks or fish schools, excels in continuous domains and converges quickly in early phases [21].

Despite their strengths, both algorithms exhibit significant limitations when applied independently. ACO often requires a large number of iterations to reach convergence and is prone to premature stagnation due to strong pheromone reinforcement in suboptimal regions [22]. PSO, while efficient in early exploration, may suffer from loss of diversity in later stages, increasing the risk of being trapped in local optima [23].

To overcome these shortcomings, researchers have proposed hybridization techniques that combine the strengths of multiple algorithms. For example, hybrid models integrating PSO with ACO, Genetic Algorithms (GAs), or Differential Evolution (DE) have demonstrated improved convergence rates and higher-quality solutions in application domains such as task scheduling, resource allocation, and semantic web service matching [24,25]. However, many of these hybrid approaches are domain-specific and are not explicitly designed for agent-based distributed execution or real-time adaptability in cloud environments.

Furthermore, existing MAS-integrated optimization models often suffer from centralized computation bottlenecks or incur excessive communication overhead among agents during solution convergence [26]. These limitations hinder their practical applicability in heterogeneous, large-scale, or time-critical service composition scenarios.

As shown in Figure 1, existing service composition approaches can be broadly classified along two key dimensions: (i) execution structure—centralized vs. agent-based, and (ii) optimization strategy—heuristic vs. hybrid. Our proposed framework falls into the lower-right quadrant, representing fully distributed, hybrid metaheuristic-based systems.

To address this gap, we propose a fully distributed framework that embeds a hybrid optimization algorithm—IPAA—within a MAS. This approach leverages the global exploration capability of PSO in early stages and transitions into refined local exploitation through ACO, thereby ensuring both fast convergence and high-quality solutions. Unlike previous methods that require central control or tightly coupled agent coordination, our platform enables agents to autonomously execute optimization tasks and make real-time decisions with minimal overhead.

As shown in Figure 2, the system architecture is structured into three logically separated layers. The agent layer manages intelligent coordination and negotiation, the optimization layer performs metaheuristic computations using the IPAA engine, and the cloud infrastructure layer hosts and monitors the actual services.

Experimental validation under simulated service environments with dynamic workloads has shown that our model outperforms standalone PSO, ACO, and random baselines in terms of both utility score and convergence time, while also maintaining robust scalability and decentralized operability.

In addition to foundational studies on MAS, PSO, and ACO, recent MAS–cloud integration research has underscored the significance of scalability, autonomous coordination, and adaptability in distributed cloud–edge environments [27]. For example, AgentFlow proposes a resilient and adaptive MAS orchestration framework that employs publish–subscribe messaging to enable scalable and decentralized coordination within cloud–edge systems. Furthermore, applications in smart manufacturing demonstrate that MAS, when integrated with cloud-based machine learning pipelines, can support intelligent and scalable decision-making in real-world production environments [28]. By incorporating these insights, our IPAA-MAS framework is positioned not only as a hybrid optimization method but also as a pragmatic orchestration system congruent with current cloud-native, agent-based developments.

3. Proposed Methodology

To enable a dynamic, scalable, and QoS-aware service composition in the cloud, this research proposes an intelligent framework that integrates a MAS with a hybrid optimization algorithm, namely the IPAA. This section details the architectural layers, algorithmic mechanisms, and agent coordination logic of the proposed system. Special attention is given to formalizing the optimization strategy through mathematical models and equations.

3.1. System Architecture

The proposed platform adopts a modular, three-tier architecture:

• Agent Layer: This layer is composed of intelligent agents that independently represent service requesters and providers. Each agent can interpret user demands, query for available services, evaluate them against QoS criteria, and collaborate with other agents when necessary. These agents act based on a decentralized coordination model, supporting autonomy, reactivity, and inter-agent communication [4,23].
• Optimization Layer: The core of this layer is the IPAA engine, which is responsible for computing optimal service compositions based on candidate services provided by agents. The optimizer evaluates multiple combinations to select the best service sequences that meet user-defined QoS preferences [21,22].
• Cloud Infrastructure Layer: This layer includes the actual service instances deployed in the cloud. It provides containerized execution environments, service discovery APIs, real-time monitoring tools, and SLA enforcement mechanisms [15,17]. Agents interact with this layer to gather runtime performance metrics.

This layered structure enhances fault tolerance, supports scalability, and separates optimization logic from deployment concerns, making the framework suitable for heterogeneous cloud environments.

3.2. Integrated Particle-Ant Algorithm (IPAA)

The proposed IPAA combines two well-established metaheuristics: PSO for global exploration and ACO for local exploitation. The switch from PSO to ACO occurs dynamically based on the stagnation of fitness improvements.

3.2.1. QoS-Aware Utility Function

A multi-objective utility function is employed to evaluate each candidate composition based on four core QoS parameters: availability, response time, reliability, and cost. The total utility $U_{total}$ for a composition of $n$ services is formally defined in Equation (1):

$$U_{total}={\sum }_{i=1}^n(w_1·A_i+w_2·{\displaystyle \frac{1}{{RT}_i}}+w_3·{AR}_i+w_4·{\displaystyle \frac{1}{C_i}})$$

(1)

Here, $A_i$, ${RT}_i$, $R_i$, and $C_i$ represent the availability, response time, reliability, and cost of the $i$-th service, respectively. The weights $w_1$ through $w_4$ correspond to the relative importance of each QoS attribute and can either be predefined by the user or automatically inferred using preference-learning models or historical user profiles [8,13]. A higher overall utility score, as computed by the objective function, indicates a superior composite service in terms of aggregated QoS performance.

The utility function explicitly optimizes four fundamental QoS attributes: response time, availability, reliability, and cost. These attributes were selected due to their central role in preventing Service-Level Agreement (SLA) violations. The weight assignment mechanism enables users to prioritize specific attributes based on application requirements or historical usage patterns, thereby ensuring a flexible and multi-objective optimization process.

3.2.2. Global Exploration Using PSO

In the first phase of IPAA, Particle Swarm Optimization is used to broadly search the service composition space. Each particle represents a possible composition path. At each iteration $t$, the velocity $v_i$ and position $x_i$ of particle $i$ are updated using the following equations:

$$v_i^{(t+1)}=w·v_i^{\left(t\right)}+c_1·r_1·\left({pbest}_i-x_i^{\left(t\right)}\right)+c_2·r_2·\left({gbest}_i-x_i^{\left(t\right)}\right)$$

(2)

$$x_i^{\left(t+1\right)}=x_i^{\left(t\right)}+v_i^{\left(t+1\right)}$$

(3)

Here, $w$ denotes the inertia weight that governs the balance between exploration and exploitation during the search process. The coefficients $c_1$ and $c_2$ represent the cognitive and social acceleration factors, respectively, which influence a particle’s tendency to move toward its own best-known position and the globally best-known position. The terms $r_1$ and $r_2$ are random values uniformly sampled from the interval [0, 1], introducing stochasticity into the update mechanism.

The term ${pbest}_i$ corresponds to the personal best position discovered by the $i$-th particle, while ${gbest}_i$ denotes the best global position identified by any particle in the swarm. The fitness of each particle is evaluated using the utility function defined in Equation (1), and the particle’s trajectory is adaptively updated in accordance with the resulting fitness values to guide the search toward optimal service compositions.

3.2.3. Local Exploitation Using ACO

After convergence stagnation is detected—based on metrics like reduced variance in fitness or entropy—ACO is activated to perform an intensive local search. Each ant constructs a composition path based on pheromone trails and heuristic desirability. The probability that ant $k$ moves from node iii to $j$ at time $t$ is given by Equation (4):

$$P_{i,j}^k\left(t\right)={\displaystyle \frac{{\left[{\tau }_{i,l}(t)\right]}^{\alpha }·{\left[{\eta }_{i,l}(t)\right]}^{\beta }}{{\sum }_{{l\in Allowed}_k}{\left[{\tau }_{i,l}(t)\right]}^{\alpha }·{\left[{\eta }_{i,l}(t)\right]}^{\beta }}}$$

(4)

${\tau }_{i,l}(t)$ represents the pheromone intensity associated with the edge between node $i$ and node $j$, indicating the historical desirability of selecting that path. The term ${\eta }_{i,l}(t)$ denotes the heuristic value, which typically reflects domain-specific desirability such as the inverse of response latency or cost. The parameters $\alpha$ and $\beta$ are tunable constants that control the relative influence of pheromone concentration and heuristic information during the probabilistic path construction. The set ${Allowed}_k$ defines the feasible candidate services that ant $k$ can visit next, based on the current composition state and task constraints.

Following the construction phase, pheromone levels are updated at the end of each iteration according to Equation (5), allowing the algorithm to reinforce high-utility paths and gradually diminish the influence of suboptimal ones through pheromone evaporation and deposition.

$${\tau }_{i,l}\left(t+1\right)=\left(1-\rho \right)·{\tau }_{i,l}\left(t\right)+{∆}_{{\tau }_{i,l}}$$

(5)

$${∆}_{{\tau }_{i,l}}={\sum }_{k=1}^m{\displaystyle \frac{Q}{U_k}}$$

(6)

$\rho$ denotes the pheromone evaporation coefficient, which controls the rate at which previous pheromone information diminishes over time. The parameter $Q$ is a positive constant that scales the amount of pheromone deposited based on solution quality. $U_k$ represents the utility value of the $k$th ant’s constructed service composition, as defined by the QoS-aware utility function. The variable $m$ indicates the total number of ants participating in the current iteration.

These pheromone update equations serve to amplify the desirability of high-utility service composition paths while simultaneously attenuating the influence of less effective paths, thereby guiding the search toward optimal or near-optimal solutions [13,21].

The transition from PSO to ACO is triggered when convergence stagnation is detected. Specifically, two threshold conditions are applied: (i) the improvement in global best utility remains below 1% for five consecutive iterations, and (ii) the normalized variance of particle positions falls below 0.05, indicating loss of swarm diversity. Once these thresholds are met, the algorithm activates the ACO-based exploitation phase. This well-defined switching criterion ensures reproducibility and prevents premature or delayed transitions.

3.3. Agent Coordination and Workflow

Agents in the proposed framework operate within a decentralized negotiation paradigm, enabling autonomous and collaborative behavior without reliance on a central coordinator. Their activities can be systematically described through a six-step workflow, which facilitates a dynamic service composition and adaptive decision-making.

First, during the request parsing phase, user requests are decomposed into a series of abstract functional tasks, each associated with specific QoS preferences. Each agent interprets and internalizes the overall composition objective, thereby enabling contextual awareness and task responsibility assignment.

Second, in the service discovery phase, agents independently query distributed service registries to retrieve candidate implementations corresponding to their assigned tasks. The retrieved services are subsequently filtered based on baseline compatibility and compliance with QoS constraints such as availability, latency, and cost [16].

Third, the optimization execution phase involves submitting the candidate service sets to the optimization layer. The IPAA processes these inputs and evaluates multiple composition sequences, ultimately returning ranked solutions based on utility maximization criteria.

Fourth, in the negotiation and resolution phase, potential conflicts may arise when multiple agents select overlapping or competing services. To address these conflicts, agents engage in utility-driven negotiation strategies to determine the optimal allocation and resolve contention [4,23].

Fifth, during the execution and monitoring phase, agents coordinate the invocation of the composed services using distributed workflow engines. In parallel, they monitor real-time system performance to ensure compliance with predefined Service-Level Agreements (SLAs). In the event of QoS degradation or SLA violation, agents are programmed to trigger re-optimization procedures to maintain service quality [12].

Finally, in the learning and adaptation phase, agents continuously collect performance logs and feedback metrics from past executions. These data are exploited to refine future decision-making processes using reinforcement learning techniques or statistical inference models, thereby enabling long-term adaptability and sustained performance improvement.

Agent coordination is implemented using a modified Contract Net Protocol (CNP). During the negotiation phase, each agent submits a utility-based bid representing its candidate service, while conflicts are resolved through a score-based arbitration mechanism. If consensus cannot be achieved, a re-composition process is automatically triggered to preserve workflow continuity. This decentralized negotiation model enhances the resilience of the MAS layer and ensures adaptive behavior under dynamic cloud conditions.

3.4. Advantages and Contributions

The proposed methodology offers several critical benefits:

• The hybrid IPAA balances rapid convergence and local optimization, outperforming traditional metaheuristics in complex environments [21,22].
• The QoS-aware utility model (Equation (1)) provides a formal, flexible basis for evaluating service compositions.
• Runtime adaptability is enabled through decentralized agent negotiation and real-time monitoring.
• The layered architecture and mathematical rigor allow for extensibility, reproducibility, and deployment in heterogeneous cloud environments.

Together, these contributions support an intelligent, efficient, and fault-tolerant service composition tailored to dynamic user needs. The robustness of the framework arises from the agent-based re-optimization mechanism, which adapts to service failures in real time. Its intelligence is derived from the hybrid PSO–ACO optimization, which balances exploration and exploitation. Scalability is ensured through the layered MAS architecture, which enables distributed negotiation and computation without centralized bottlenecks.

4. Experimental Results

To evaluate the effectiveness, efficiency, and scalability of the proposed multi-agent service composition platform using the IPAA, we conducted a series of controlled simulations.

4.1. Experimental Setup

All simulations were conducted in a controlled computing environment to ensure reproducibility. The testbed was implemented using Python 3.11 with the SimPy simulation engine (Python 3.11). The experiments were executed on a system with an Intel Core i7-12700H CPU and 32 GB RAM, ensuring negligible hardware-related noise in performance evaluation.

To emulate a realistic cloud service environment, a synthetic dataset comprising 300 atomic services was generated. These services were categorized into ten functional domains, including payment processing, data storage, and analytics. Each service instance was assigned four key QoS attributes—availability, response time, reliability, and cost. The value ranges of these QoS attributes are summarized in

Table 1. QoS parameter ranges for simulated services.

QoS Parameter	Minimum	Maximum	Distribution
Availability	0.85	0.999	Uniform
Response Time (ms)	100	1500	Normal (μ = 600, σ = 200)
Reliability	0.70	0.99	Uniform
Cost (USD)	0.10	5.00	Normal (μ = 1.5, σ = 1.0)

, which were selected to reflect the characteristics of cloud services reported in prior studies [29,30].

Specifically, response time values followed a normal distribution (μ = 600 ms, σ = 200), calibrated based on latency observations reported in benchmarking studies. In contrast, availability and reliability were generated using uniform distributions to capture service heterogeneity. Prior to optimization, all QoS attributes were normalized into the range [0, 1] to account for differences in scale and units, thereby ensuring fairness in utility computation.

Each composition task required selecting one service per function, resulting in a search space of up to 10³⁰ possible combinations. This combinatorial complexity made it an ideal test scenario for evaluating the efficiency of heuristic optimization techniques.

To evaluate the effectiveness of the proposed IPAA, we conducted a comparative analysis against three baseline methods: PSO [12], ACO [13], and a Random Composition (RC) strategy. The RC method serves as a non-heuristic benchmark, where services are selected randomly without any optimization logic.

Each algorithm was executed 30 independent times on the same experimental dataset. This repeated execution was performed to mitigate the effects of stochastic variability inherent in metaheuristic algorithms, and the average performance metrics were recorded for analysis. Key evaluation criteria included the overall utility score, execution time, and consistency of solution quality.

4.2. Evaluation Metrics

To comprehensively evaluate the performance of each algorithm, we employed four key evaluation metrics designed to capture both the effectiveness and efficiency of the service composition process.

• Average Utility Score: This metric quantifies the overall quality of the composed service by applying the QoS-aware utility function previously defined in Equation (1). A higher utility value reflects a more desirable trade-off among the four QoS parameters—availability, response time, reliability, and cost-efficiency.
• Execution Time: This refers to the total computational time (measured in seconds) required by the optimization algorithm to return the best composition plan. It provides insight into the algorithm’s real-time applicability in dynamic cloud environments.
• Success Rate: This metric denotes the proportion of trials (out of 30 runs) in which the algorithm achieved a utility score greater than 85% of the maximum theoretical utility. It is used to assess the robustness and consistency of the optimization method under stochastic variations.
• Convergence Iterations: This represents the number of iterations required for the algorithm to converge, where convergence is defined as the point at which the utility improvement between successive iterations falls below 1%. It reflects the stability and convergence speed of the algorithm.

Together, these metrics offer a balanced evaluation framework that captures the quality of solutions (utility and success rate) as well as the computational efficiency (execution time and convergence behavior). This combination ensures a holistic understanding of how each algorithm performs under realistic cloud service composition scenarios. The combination of these four metrics ensures a holistic evaluation: the utility score reflects solution quality, execution time and convergence iterations capture efficiency, and success rate evaluates robustness. Together, they provide a comprehensive measure of overall algorithmic superiority.

4.3. Results and Comparative Analysis

4.3.1. Utility Score Performance

As summarized in

Table 2. Comparison of optimization methods (average over 30 runs).

Algorithm	Avg. Utility ↑	Execution Time (s) ↓	Success Rate (%) ↑	Convergence Iterations ↓
IPAA	0.912	1.42	96.7	23
PSO	0.878	1.35	89.2	30
ACO	0.861	2.72	86.3	47
RC	0.693	N/A	48.1	N/A

, the proposed IPAA outperforms the benchmark methods in terms of average utility, execution efficiency, success rate, and convergence iterations. Notably, the IPAA consistently achieves the highest utility values, indicating its ability to generate high-quality service compositions. Furthermore, its superior success rate and reduced convergence iterations demonstrate both robustness and computational efficiency.

The IPAA consistently outperformed the traditional methods across all metrics. While PSO demonstrated faster convergence in early stages, it tended to stabilize around suboptimal solutions. ACO showed a stronger exploitation capability but at the cost of a longer execution time. The random composition baseline performed poorly, validating the necessity of intelligent optimization in complex service environments.

4.3.2. Convergence Behavior

The convergence behavior of the three algorithms is shown in Figure 3, which illustrates how the utility score evolved over 60 optimization iterations.

In the early iterations (1–20), the IPAA followed a trajectory similar to PSO, demonstrating rapid global exploration. However, unlike PSO, the IPAA transitioned smoothly into local exploitation using ACO, leading to steady refinement and higher final utility. ACO alone required nearly twice as many iterations to converge and often suffered from slower exploration in early stages.

This observation empirically confirms the theoretical design of the IPAA, where PSO’s fast convergence is balanced by ACO’s accuracy through sequential integration.

4.3.3. Utility Distribution Robustness

Figure 4 presents box plots of utility scores for each algorithm over 30 independent runs. The IPAA achieved an average utility score of 0.912 ± 0.021, while PSO and ACO reported 0.878 ± 0.034 and 0.861 ± 0.039, respectively. This additional reporting provides stronger evidence of statistical consistency across independent runs.

This statistical consistency is important in real-world deployments, where service availability and user requests vary unpredictably. A robust optimization performance helps ensure SLA satisfaction across diverse operating conditions.

4.3.4. Scalability Evaluation

To evaluate scalability, the number of candidate services per function was gradually increased from 10 up to 1000, thereby simulating higher workloads and service density. The execution times, summarized in Figure 5, demonstrate that the IPAA maintained a near-linear scaling trend.

Specifically, the runtime increased from 1.4 s (100 candidates) to 5.7 s (1000 candidates). In contrast, ACO exhibited an exponential growth in execution time, while PSO scaled linearly but consistently achieved lower utility scores in larger search spaces. These findings confirm that the proposed IPAA framework preserves practical scalability and robustness, making it suitable for deployment in modern cloud environments where the number of services continues to grow dynamically.

4.3.5. Sensitivity Analysis of QoS Weights

To assess the robustness of the proposed QoS-aware utility function, a comprehensive sensitivity analysis was conducted by systematically varying the relative weights assigned to different QoS attributes. When the cost parameter was assigned twice the weight of reliability, the average utility score exhibited a modest decline of 3.2%; however, the IPAA consistently outperformed both PSO and ACO under this configuration. Conversely, when reliability was prioritized, the IPAA achieved a success rate exceeding 95%, markedly higher than the baseline algorithms. These results demonstrate that the proposed framework exhibits a stable and reliable performance across heterogeneous user preference profiles.

4.4. Discussion

The experimental results provide compelling evidence that the proposed IPAA-based framework significantly improves both optimization quality and computational efficiency compared with traditional methods. By integrating global exploration through PSO with local refinement via ACO, the framework consistently achieves high-utility compositions, faster convergence, and greater reliability. Furthermore, the decentralized, agent-based execution model enhances system resilience, enabling agents to re-optimize dynamically in response to runtime service failures or SLA violations. The synergy of algorithmic intelligence and architectural modularity positions the framework as a practical and scalable solution for intelligent cloud service composition.

In addition to synthetic simulations, the framework is designed to handle unpredictable noise in real-world environments. For instance, sudden variations in response latency or random service downtime can be mitigated through the agent-based re-optimization mechanism, which triggers localized re-composition upon SLA violations. Although the current study relies primarily on synthetic datasets to ensure reproducibility and controlled evaluation, future research will validate the framework against publicly available datasets, such as the QWS Web Service dataset and large-scale cloud traces from commercial providers (e.g., Amazon AWS, Microsoft Azure). Such validation will strengthen the external validity of the results and further demonstrate the robustness of the proposed IPAA framework under heterogeneous and dynamic service conditions.

The hybrid nature of the IPAA inevitably introduces additional computational overhead compared with standalone algorithms. Complexity analysis indicates that the average time complexity of the IPAA is O (N·logN), with an approximately 12% higher runtime cost than PSO in equivalent conditions. Despite this overhead, the execution times remained within sub-second levels for most configurations, rendering the framework feasible for real-time and SLA-sensitive cloud applications.

5. Conclusions

In this study, we proposed a novel framework for a cloud-based service composition that integrates a MAS with a hybrid metaheuristic optimization algorithm, the IPAA. The primary goal of this framework is to address the increasing complexity, scale, and dynamism of service-oriented architectures in modern cloud environments. By leveraging the autonomous decision-making capabilities of agents and the complementary strengths of PSO and ACO, the framework facilitates intelligent, efficient, and QoS-aware service compositions.

The proposed IPAA method was designed to balance global exploration and local exploitation effectively. This is achieved through a dynamic transition mechanism. The optimizer begins with the fast convergence of PSO and then refines the solutions using the guided exploitation of ACO. We also developed a QoS-driven utility model (Equation (1)) that enables multi-objective evaluation across availability, response time, reliability, and cost, making the solution highly adaptable to user-defined requirements.

To evaluate the performance and validity of our approach, we conducted extensive simulations using a synthetic dataset of 300 cloud services across 10 functions. The results, as reported in Section 4, clearly demonstrate that our method outperforms conventional optimization approaches in terms of utility score, convergence speed, and robustness. Specifically, the IPAA-based system achieved an average utility of 0.912—significantly higher than that of PSO (0.878), ACO (0.861), and Random Composition (0.693)—while also maintaining a shorter execution time and faster convergence. Moreover, the IPAA demonstrated superior scalability, remaining efficient as the number of candidate services increased up to 500, a critical requirement for real-world applications.

Finally, the distributed architecture further enhanced the system’s fault tolerance and responsiveness. Agents were able to detect SLA violations and trigger real-time re-composition with minimal disruption, an essential capability for resilient service orchestration in volatile environments.

Despite promising results, the present study has two main limitations: (i) reliance on synthetic datasets and (ii) evaluation within a controlled cloud simulation environment. Future research will address these limitations by conducting experiments with large-scale real-world service datasets, examining the energy efficiency of hybrid optimization in cloud infrastructures, and extending the architecture to a federated MAS framework for distributed, cross-cloud orchestration. These directions are expected to enhance the generalizability, sustainability, and scalability of the proposed approach.