ARCHIVES

Smart Services Such As Health Monitoring, Camera-Based Access Control And Ride-Hailing Must Decide What To Compute, Where To Compute It And When. Sending Every Reading To The Cloud adds Delay And Network Load, While Running Every Model On The Device Drains Its Battery And Memory. Edge Computing Places Processing Between These Two Extremes, And Learning-Based Schedulers And Dispatchers Decide How The Available Processors, Servers And Vehicles Are Shared. The Two Lines Of Work Are Usually Studied Apart: Edge Studies Focus On Architecture And Analytics, Whereas Scheduling Studies Focus On Operating Systems, Clusters Or Mobility Platforms. As A Result, Little Guidance Exists On How A Placement Decision And A Sharing Decision Should Be Made Together, Or Which Learning Method Fits Which Decision. This Review Examines Edge Intelligence And Adaptive Resource Management For Smart Services Under One Device-Edge-Cloud View. It Covers The Characteristics And Analytics Of Edge Computing, On-Device And Edge Inference For Vision And Health Services, Learning-Based Scheduling Inside Operating Systems And Clusters, And Learning-Based Dispatch On Ride-Hailing Platforms. The Review Finds That Edge Analytics Lowers Latency And Bandwidth For Monitoring And Vision Services; That Ensemble And Deep Classifiers Can Run Close To The User For Face, Emotion, Age And Gender Recognition; And That Reinforcement Learning Outperforms Fixed Rules When A Decision Changes Later System States. A Hybrid Context-Aware Scheduler Cut Response Time By 31.7% Against The Linux Completely Fair Scheduler, And An Actor-Critic Dispatcher Reduced Mean Passenger Waiting Time To 5.95 Minutes Against 8.68 Minutes For Proximal Policy Optimization. Open Problems Include The Joint Optimization Of Placement And Scheduling, Energy-Aware Rewards, Privacy Of Personal Data Processed At The Edge, Safe Fallback Behaviour For Learned Policies And The Absence Of Shared Benchmarks Across Service Domains.