Perception-Driven Curiosity with Bayesian Surprise

Extrinsic reward over tme steps in training for the following Atari games: Beamrider, Breakout, Gravitar, River Raid, Private Eye, Space Invaders. The extrinsic reward achieved by the policies trained with ICM, our Bayesian surprise, and extrinsic reward are all shown. Pictures of the game scenes are next to each plot. ● Our method performs comparably or outperforms ICM on each Atari game in our experiments except for Breakout which is the only entrely deterministc game.

以下雅达利游戏在训练过程中的时间步长上的外在奖励：《光束骑手》、《打砖块》、《重力球》、《河流大冒险》、《私家侦探》、《太空侵略者》。展示了通过内在好奇心模块（ICM）训练的策略所获得的外在奖励、我们的贝叶斯惊奇以及外在奖励。每个图表旁边都有游戏场景的图片。●在我们的实验中，除了《打砖块》（这是唯一完全确定性的游戏）之外，我们的方法在每个雅达利游戏上的表现与ICM相当或优于ICM。