Discover Computing
Feasibility-aware warm-initialized QAOA for constrained cloud–fog task scheduling under NISQ-oriented noise
Suresh Kumar Jha, Niroj Kumar Pani, Siddhanta Kumar Singh, Ketan Anand
Abstract
Constrained task scheduling in cloud–fog environments is challenging because heterogeneous resources, latency and energy requirements, and capacity constraints create a combinatorial optimization problem in which QAOA may generate infeasible schedules. This paper proposes a feasibility-aware hybrid quantum–classical framework combining Warm-Initialized QAOA with constraint-aware decoding and joint CP-SAT repair under noisy intermediate-scale quantum (NISQ)-oriented conditions. The scheduling problem is formulated as a quadratic unconstrained binary optimization (QUBO) model that jointly captures latency–energy optimization, deadlines, and resource capacities. The framework is evaluated on 12 structured benchmark instances spanning balanced, capacity-tight, deadline-tight, and heterogeneous regimes, with executable QUBOs containing 20–22 logical qubits. Across 96 production configurations under ideal, low, medium, and high noise conditions, 91 configurations ( $$94.79\%$$ ) produced at least one schedule-feasible sampled candidate before repair; this configuration-level success rate is distinct from the shot-level schedule-feasible probability. The joint CP-SAT repair stage successfully restored feasibility in all five configurations for which no feasible sampled candidate was available, yielding $$100\%$$ post-repair feasibility and reducing the mean constraint-violation count from $$0.09375$$ to $$0$$ . For the 91 configurations with a feasible pre-repair selected candidate, the paired mean optimality gap remained unchanged at $$10.60\%$$ after repair, whereas the post-repair mean over all 96 configurations was $$10.79\%$$ because the five repaired cases entered the post-repair denominator. Matched attribution controls further show that the schedule-feasible probability increases from $$0.44\%$$ for uniform initialization at $$p=0$$ to $$29.86\%$$ for warm initialization at $$p=0$$ and $$30.52\%$$ for Warm-Initialized QAOA at $$p=1$$ , indicating that the dominant feasible-region concentration arises from warm initialization, with a smaller incremental contribution from the shallow QAOA evolution. Additional sensitivity, classical-reference, and GPU scalability experiments characterize parameter dependence, classical competitiveness, runtime, and sampling behavior within the investigated NISQ-oriented simulation regime.