Text this: Adaptive reinforcement learning with active state-specific exploration for engagement maximization during simulated child-robot interaction