Curiosity-Driven Exploration

Project Info:

Team Size: 3

Technologies/Tools used: Pytorch Python

Project Timeline: September 2021 - December 2021

Video Demonstration

Motivation

Several real world environments/games provide rewards infrequently or only at the end of specific tasks (sparse-reward settings). In such situations, where extrinsic rewards are infrequent, we can implement an intrinsic reward mechanism to help the agent to explore the environment. We implemented the system proposed in the paper: Curiosity-Driven Exploration by Pathak et al. The Intrinsic Curiosity Model (ICM) which is the central idea proposed by the paper, is learned by the agent without any rewards from the environment.