Edge computing is a promising paradigm for enabling low-latency and intelligent services closer to end users. To address increasing demands for real-time responsiveness and scalability, microservice architectures are implemented, enabling applications to be divided and deployed flexibly across edge layers. However, optimal task allocation and resource management remain major challenges in this dynamic environment. Microservice Architecture (MSA) offers a modular design framework compatible with the characteristics of edge computing. In addition, it allows scalability, fault tolerance, and continuous delivery by dividing applications into loosely linked and independently deployable services. These services can be effectively deployed across edge nodes depending on latency, computation demands, and data requirements.

The current research proposes a traffic-aware optimal association and task-offloading architecture optimized for microservice-based edge computing. The proposed approach utilizes a high-precision prediction model for estimating future offloading requests rather than depending on average offloading rates, allowing more accurate resource provisioning for distributed microservices. An optimization-based scheduling approach is introduced to ensure that microservice tasks comply with strict deadlines necessary for task-critical services. Furthermore, the framework allocates computational loads across multiple time steps, enabling adaptive task-offloading decisions that correspond with the dynamic environment. While this framework provides a solid foundation, it still exhibits several limitations. The system depends on pre-learned patterns and lacks real-time adaptability to sudden shifts, which happen frequently in real-world implementation. The decision-making process focuses on short-term optimization, overlooking the long-term impact of present actions. The lack of a feedback mechanism reduces the ability of the system to learn from previous outcomes, and manual parameter adjustment could limit its utility across various workloads and deployment scenarios.

To overcome these limitations and enhance adaptability, we propose a deep reinforcement learning (DRL) approach using the Deep Deterministic Policy Gradient (DDPG) algorithm. This method formulates microservice deployment as a multi-x-objective optimization problem (MOMDP), aiming to minimize interaction costs while maximizing edge resource utilization. Compared to static models, DDPG enables the system to learn directly through continuous interaction with the environment, allowing it to respond to real-time changes. To enhance policy learning and ensure efficiency, expert interventions are integrated to guide resource and service allocation, enabling the agent to avoid suboptimal decisions and achieve faster convergence. As a result, the learning process achieves significantly faster convergence. Comprehensive simulation experiments are conducted to validate the effectiveness of the proposed approach.