Inference replication at edges via combinatorial multi-armed bandit