ARCHITECTURAL DECISION-MAKING USING REINFORCEMENT LEARNING IN LARGE-SCALE SOFTWARE SYSTEMS

Abstract

Architectural decision-making in large-scale software systems plays a crucial role in determining performance, scalability, and maintainability. Traditional methods, such as rule-based and heuristic-based systems, often fall short in managing the complexity and dynamism of modern software environments. This study explores the application of reinforcement learning (RL) to address these limitations. Leveraging advanced RL techniques, including Deep Q-Networks (DQN) and Proximal Policy Optimization (PPO), the research proposes a novel approach for optimizing architectural decisions.

The RL-based system was evaluated against traditional methods, demonstrating superior performance in several areas. The RL system achieved a decision accuracy of 90%, closely aligning with expert architects' decisions. It also outperformed traditional systems in decision-making speed, with an average time of 60 seconds, compared to 120 and 180 seconds for rule-based and heuristic systems, respectively. Furthermore, the RL approach exhibited strong adaptability, handling dynamic changes and constraints with a score of 85. Overall, it improved system performance by 80%, enhancing response time, scalability, and maintainability.