Experience Replay For Learning Multiple Environments

Stekovic, S. (Speaker)

Activity: Talk or presentationTalk at conference or symposiumScience to public

Description

In this work, we adapt the standard experience replay approach for the task of learning multiple similar environments. Particularly for our task, we consider a robot learning decision making for multiple corridor environments. While using single monocular images for the observation states, the agent learns to predict the reward that is related to the maximum traveled distance that can be reached for the given action. Based on the reward predictions, the agent than decides which action to take.
Period9 May 2018
Held atBMVA Symposium on Reinforcement Learning in Computer Vision
Event typeOther
LocationLondon, United Kingdom

Keywords

  • Reinforcement
  • Learning
  • Robotics
  • Computational Intelligence
  • Vision