eduzhai > Applied Sciences > Engineering >

Noisy Agents Self-supervised Exploration by Predicting Auditory Events

  • king
  • (0) Download
  • 20210506
  • Save

... pages left unread,continue reading

Document pages: 13 pages

Abstract: Humans integrate multiple sensory modalities (e.g. visual and audio) to builda causal understanding of the physical world. In this work, we propose a noveltype of intrinsic motivation for Reinforcement Learning (RL) that encouragesthe agent to understand the causal effect of its actions through auditory eventprediction. First, we allow the agent to collect a small amount of acousticdata and use K-means to discover underlying auditory event clusters. We thentrain a neural network to predict the auditory events and use the predictionerrors as intrinsic rewards to guide RL exploration. Experimental results onAtari games show that our new intrinsic motivation significantly outperformsseveral state-of-the-art baselines. We further visualize our noisy agents behavior in a physics environment and demonstrate that our newly designedintrinsic reward leads to the emergence of physical interaction behaviors (e.g.contact with objects).

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×