PACE: Proactive Adaptive Constrained Edge-Decision for Safe and Generalizable Online Decision-Making

Edge computing requires safe, efficient, and generalizable online decision-making, yet existing methods suffer from reactive constraint handling, fragmented scheduling frameworks, and poor generalization. We propose PACE, a unified framework shifting from reactive remediation to proactive anticipation. PACE integrates a Proactive Constrained Policy Optimizer with preemptive penalty and constraint-aware intrinsic rewards, a Nested Index Scheduler with closed-form policies for preemptive and non-preemptive AoI minimization, and a Generalizable Multi-Objective Offloading Network with histogram encoding and masking for single-policy generalization. Experiments on safe locomotion, MEC scheduling, and multi-objective offloading show PACE achieves highest returns with strict constraint satisfaction, reduces AoI by up to 61.84%, and attains near-optimal Pareto performance within 0.3% of the upper bound using a single policy.