Introduction to Reinforcement Learning

Canonical starting point for modern RL taught by the lead architect of AlphaGo and AlphaZero: agent-environment loops, Markov state, policy, value functions, and exploration vs. exploitation.

Builds the mathematical bedrock of RL step-by-step from Markov chains to MDPs, Bellman expectation equations, and nonlinear Bellman optimality equations.

Proves contraction convergence of iterative policy evaluation, policy iteration, value iteration, and generalized policy iteration.

Contrasts unbiased high-variance Monte Carlo returns against low-variance bootstrapped Temporal-Difference learning and eligibility traces TD(λ).

Derives GLIE convergence for on-policy SARSA, importance sampling, and Watkin's off-policy Q-Learning algorithm.

Diagnoses the 'deadly triad' of function approximation + bootstrapping + off-policy training, and shows how experience replay and fixed target networks stabilized Atari DQN.

#Reinforcement Learning Course by David Silver# Lecture 7: Policy Gradient Methods (updated video thanks to: John Assael).

#Reinforcement Learning Course by David Silver# Lecture 8: Integrating Learning and Planning.

This lecture dissects the fundamental dilemma between exploration and exploitation in reinforcement learning, introducing techniques like epsilon-greedy and upper confidence bound (UCB) algorithms. It examines how these strategies balance discovering new information with leveraging existing knowledge to maximize rewards.

This session explores the application of reinforcement learning principles to classic games, detailing methods like Monte Carlo Tree Search (MCTS) and alpha-beta pruning. It illustrates how RL agents can learn optimal strategies in competitive environments with perfect or imperfect information.