Optimization of Resource Scheduling in Digital Avatar Education Systems Using Distributed Deep Reinforcement Learning

Le résumé fourni par la source

With the acceleration of digital transformation in education, digital human education systems face resource scheduling challenges under high-concurrency scenarios while providing immersive learning experiences [1]. This paper proposes a dynamic resource scheduling framework based on Distributed Deep Reinforcement Learning (D-DRL) [2]. By constructing a multi-dimensional state space (node load vector + task feature matrix) and a composite action space (resource allocation/task migration/bandwidth control), combined with a Pareto-optimal reward function, intelligent scheduling of GPU core resources is achieved [3]. The system adopts a four-layer microservices architecture, deploys a D-DRL decision engine at the resource scheduling layer, integrates an LSTM prediction model at the load balancing layer, implements digital human interaction agents in the microservices cluster, and uses a MySQL+Redis hybrid architecture at the data storage layer. Experimental results show that under 5000 concurrent users, the system achieves 81.2% GPU utilization and 178ms P95 latency, with SLA compliance reaching 92.7%, which is 26.5% higher than the Kubernetes HPA solution. This research provides an efficient and reliable resource scheduling solution for digital transformation in education while offering technical references for other high-concurrency intelligent service systems.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.