Title: Comparative Analysis of Intrinsic Reward in Reinforcement Learning

Abstract: Reinforcement learning often struggles in environments where rewards are sparse. In such settings, agents can wander aimlessly without sufficient feedback to guide learning. This thesis explores the use of intrinsic rewards—signals that encourage exploration even when extrinsic rewards are sparse. We focus on three main strategies for generating these intrinsic signals: (1) counting how often states are visited, (2) comparing outputs of a randomly fixed network with a trained predictor, and (3) measuring errors in a learned forward model. Our work provides a unified view of these approaches, discusses their strengths and weaknesses, and offers practical advice for choosing and tuning them in real-world tasks. Through experiments in grid-based and Atari-style environments, we show that adding intrinsic rewards can significantly improve exploration and lead to faster policy learning when extrinsic rewards alone are not enough.