Curiosity-Driven Collaborative Request Scheduling in Mobile Edge-Cloud Systems

Abstract

The rapid proliferation of data-intensive and delay-sensitive applications has accelerated the evolution from centralized cloud computing to mobile edge-cloud systems. However, the decentralized and heterogeneous characteristics introduce instability and complexity, making efficient scheduling increasingly challenging. A core problem lies in the joint optimization of request dispatching (RD) and service orchestration (SO), whose decoupled decisions often cause inconsistent objectives and inefficient resource utilization. Furthermore, reinforcement learning-based scheduling methods usually exhibit slow convergence due to sparse and delayed rewards. This paper presents Cur-CoEdge, a curiosity-driven collaborative scheduling framework that integrates hierarchical coordination with intrinsic motivation for adaptive edge-cloud optimization. Cur-CoEdge adopts a multi-timescale learning structure, where we use a multi-agent advantage actor-critic (MAA2C) for per-slot RD. We also apply a graph convolutional network-based A2C (GCN-A2C) for per-frame SO. A bidirectional decision interaction mechanism composed of upper-to-lower guidance, lower-to-upper refinement, and lower-for-upper catering enables consistent coordination across layers. To enhance convergence efficiency under sparse rewards, we develop a curiosity-driven collaborative exploration method, which fuses fixed distance metrics with an attention-based adaptive similarity weighting to model inter-dispatcher curiosity relationships. We theoretically analyze the convergence property of this exploration mechanism and formally prove it. Experiments on a real-world testbed using Alibaba Cluster and PPIO traces demonstrate that Cur-CoEdge achieves up to 40% higher throughput, 26% better time efficiency, and 71% faster convergence compared with existing baselines, showing its potential for large-scale distributed environments.