Learning to reflect: A unifying approach for data-driven stochastic control strategies