Artificial Intelligence: Reinforcement Learning in Python

Artificial Intelligence: Reinforcement Learning in Python by Lazy Programmer (Lazy Programmer Inc.) is the strongest budget-priced introduction to classical reinforcement learning, but it is foundations-only despite the deep-RL marketing wrapper. Across 179 lectures in 22 sections (roughly 18-20 hours), you build every algorithm by hand in plain Python and NumPy: epsilon-greedy / UCB1 / Thompson Sampling bandits, Markov Decision Processes, dynamic programming, Monte Carlo, TD(0), SARSA, Q-learning, and linear function approximation, finishing with a Q-learning stock-trading project. Note the accuracy caveat: this course does NOT teach Deep Q-Networks from scratch despite some catalog blurbs implying it; DQN with experience replay and target networks lives in the instructor's separate 'Advanced AI: Deep Reinforcement Learning' courses. Public sentiment is consistently positive (Udemy lists 4.7/5 across roughly 10,700 ratings; the instructor's own deeplearningcourses.com page also shows 4.7/5), with learners praising the from-scratch, theory-plus-code approach. It is excellent if you want to genuinely understand tabular RL before touching neural nets, and a poor fit if you came expecting hands-on deep RL.

A high-quality, rigorously from-scratch grounding in classical/tabular reinforcement learning at a very low price, but only worth taking if you (a) actually want the fundamentals rather than deep RL, and (b) meet the real math/Python prerequisites (calculus, probability, NumPy, linear regression). The catalog's 'deep Q-networks from scratch' framing overstates the scope, so buy it for what it truly is.

Best for: Python programmers and aspiring ML/data-science practitioners who want to understand reinforcement learning at a technical, build-it-yourself level before jumping to deep RL libraries. It suits people who learn well from paired theory-then-code lectures, who are comfortable with calculus, probability, OOP, NumPy and linear/logistic regression, and who value implementing bandits, MDPs, dynamic programming, Monte Carlo and Q-learning by hand.

Skip if: Complete programming beginners, anyone wanting hands-on Deep Q-Networks / policy gradients / actor-critic (those are in the instructor's separate Advanced/Deep RL courses), people who dislike heavy theory or want a fast plug-and-play library tour, and learners who prefer the canonical Sutton & Barto academic treatment with formal exercises and rigorous proofs.