Object-sensitive Deep Reinforcement Learning

Deep neural networks have been widely applied in reinforcement learning (RL) algorithms to achieve human-level control in various challenging domains. More specifically, recent work has found outstanding performances of deep reinforcement learning (DRL) models on Atari 2600 games, by using only raw pixels to make decisions [21]. The literature on reinforcement learning is vast. Multiple deep RL algorithms have been developed to incorporate both on-policy RL such as Sarsa [30], actor-critic methods [1], etc. and off-policy RL such as Q-learning using experience replay memory [21] [25]. A parallel RL paradigm [20] has also been proposed to reduce the heavy reliance of deep RL algorithms on specialized hardware or distributed architectures. However, while a high proportion of RL applications such as Atari 2600 games contain objects with different gain or penalty (for example, enemy ships and fuel vessel are two different objects in the game "Riverraid"), most of previous algorithms are designed under the assumption that various game objects are treated equally.