Reinforcement learning overview

Author: Irakli Koberidze
Keywords: reinforcement learning
Annotation:

This paper provides a comprehensive overview of the evolution of Reinforcement Learning (RL), focusing on the integration of Q-learning with function approximation techniques. Starting with foundational principles such as Markov Decision Processes (MDPs) and Dynamic Programming (DP), we trace the development of key concepts including Temporal Difference (TD) learning and the landmark Q-learning algorithm. Despite significant advancements, traditional RL algorithms encounter challenges in scalability due to high-dimensional state and action spaces, leading to inefficiencies in value function computation and policy learning. To mitigate these issues, function approximation methods, including linear models and deep neural networks, have been employed to generalize value functions from limited data. This paper systematically reviews these developments, emphasizing the pivotal role of Q-learning combined with function approximation in advancing scalable RL solutions.

Lecture files:

გამოწრთობით დასწავლის მიმოხილვა [ka]