Adaptive trajectory-constrained exploration strategy for deep reinforcement learning.