Sample-Efficient Exploration in Visually Complex Reinforcement Learning Tasks

Key information:

Abstract

Reinforcement Learning has shown incredible capabilities to produce agents with superhuman performance for a wide range of tasks. One current major limitation however is the number of interactions these agents need with their environment to achieve this performance. This means that currently many tasks where environment interaction is costly such as real-world training remain off-limits. Sample-Efficient Reinforcement Learning however looks to create agents which can learn with minimal environment interactions and learn new tasks which were previously thought to be infeasible. In addition to learning where environment interactions are costly, sample-efficient methods push the field towards solutions which require much less computation, making the training of agents more sustainable and more inclusive to researchers and practitioners which limited resources.