Reinforcement Learning: Paradigm, Foundations, and Applications

Abstract

Reinforcement learning (RL) occupies a distinct position within machine learning, differing fundamentally from both supervised and unsupervised paradigms. Unlike supervised learning, RL agents receive evaluative feedback – judgment on the consequences of actions – rather than explicit correct answers, and they learn continuously without a static training-testing distinction. Unlike unsupervised learning, RL operates within an interactive environment with a clear objective: maximizing cumulative reward. Built upon a robust mathematical foundation, RL unifies diverse fields including economics, optimization, game theory, robotics, and cognitive science. This paper provides a self-contained tutorial on RL foundations, covering Markov decision processes, value functions, Bellman equations, temporal-difference learning, Q-learning, policy gradients, and actor-critic methods. The primary research contribution is a reproducible empirical benchmarking study that systematically compares three exploration strategies (ϵ-greedy, Boltzmann, and UCB) in a controlled grid-world environment, providing full experimental specifications, statistical confidence intervals, value iteration baselines, and practical guidelines for exploration strategy selection under varying environmental conditions.