Combining Deep Reinforcement Learning and Search for Imperfect-Information Games