RL-Based Fragment Allocation and Replication for Distributed Heritage Multimedia Databases

Mahmoud, Shymaa Hosny; Badr, Nagwa; Abdelraouf, Ahmed E; Ali, Mohamed Ibrahim

doi:10.21608/ijicis.2025.411803.1419

RL-Based Fragment Allocation and Replication for Distributed Heritage Multimedia Databases

Document Type : Original Article

Authors

¹ faculty of computer and information science, information system, Ain shams university

² Department of Information Systems, Faculty of Computer and Information Sciences, Ain Shams University, Cairo, 11566, Egypt

³ Museums and Archaeology Sites Management Department Faculty of Archeology, Ain Shams University.

10.21608/ijicis.2025.411803.1419

Abstract

Optimizing fragment allocation and replication in distributed heritage multimedia databases is crucial for minimizing query execution costs in dynamic, resource-constrained environments. Existing heuristics, such as VFAR (Vertical Fragment Allocation and Replication) often neglect inter-fragment dependencies, join relationships, and site capacity limitations, which can result in inefficient allocations. This paper introduces RL-FAWM (Reinforcement Learning based Fragment Allocation and Replication for Workload-aware Multimedia Systems), a Q-learning framework that models fragment placement as a Markov Decision Process. RL-FAWM incrementally learns effective allocation and replication strategies over multiple episodes using a Q-table, allowing the system to adapt to dynamic workload changes. The approach incorporates structured workload matrices, including read frequency (FRM), manipulation intensity (FMM), and co-access patterns (FCAM), along with inter-site communication costs and strict storage capacity constraints. A cost estimator provides real-time feedback to guide the learning process toward globally optimal and constraint-compliant configurations.Experimental results on a distributed multimedia case study demonstrate that RL-FAWM consistently reduces execution costs and avoids site overloading, outperforming VFAR while minimizing inter-site join costs. These findings underscore the potential of reinforcement learning for adaptive, scalable, and constraint-aware data management in digital preservation systems.

Keywords