Reinforcement Learning

Teach AI agents to make decisions through trial and error — the technology behind game-playing AI, robotics, and autonomous systems.

✅ What You'll Learn

• The exploration vs exploitation dilemma
• Q-Learning with epsilon-greedy strategy
• Multi-armed bandit problem
• Policy gradient methods (REINFORCE)

🎮 What Is Reinforcement Learning?

🎯 Real-World Analogy: Think of training a puppy. You don't show it 10,000 labelled examples of "sit" (supervised learning). Instead, the puppy tries random actions, and you give a treat (reward) when it sits. Over time, the puppy learns: "sitting when I hear 'sit' = treat!" That's reinforcement learning.

Unlike supervised learning (learn from labels) or unsupervised learning (find patterns), RL learns from interaction. An agent takes actions in an environment, receives rewards, and learns a policy — a strategy for choosing actions that maximise long-term reward.

🧩 Core Components

• Agent — The learner/decision-maker
• Environment — The world the agent acts in
• State — Current situation
• Action — What the agent can do
• Reward — Feedback signal (positive or negative)
• Policy — Strategy mapping states → actions

Try It: Multi-Armed Bandit

Explore the exploration vs exploitation trade-off with slot machines

Try it Yourself »

Python

import numpy as np

# Multi-Armed Bandit: The Exploration vs Exploitation Dilemma
# Imagine 4 slot machines — which one pays the most?

np.random.seed(42)

# True reward probabilities (agent doesn't know these!)
true_probs = [0.2, 0.5, 0.75, 0.3]
n_arms = len(true_probs)

def pull_arm(arm):
    """Simulate pulling a slot machine lever"""
    return 1 if np.random.random() < true_probs[arm] else 0

# Strategy 1: Epsilon-Greedy
def epsilon_greedy(n_rounds, epsilon=0.1):
    counts = np.zeros(n_arm
...

Try It: Q-Learning Grid Navigation

Train an agent to find the shortest path through a 4×4 grid

Try it Yourself »

Python

import numpy as np

# Q-Learning: Teaching an agent to navigate a grid world
# The agent learns which actions lead to rewards

# Grid: 4x4, Agent starts at (0,0), Goal at (3,3)
grid_size = 4
n_states = grid_size * grid_size
n_actions = 4  # up, down, left, right
actions = ["UP", "DOWN", "LEFT", "RIGHT"]

# Q-table: stores expected reward for each (state, action) pair
Q = np.zeros((n_states, n_actions))

# Reward: +10 at goal, -1 per step (encourages shortest path)
def get_reward(state):
    if s
...

Try It: Policy Gradient (REINFORCE)

Learn a policy directly — scales to complex environments

Try it Yourself »

Python

import numpy as np

# Policy Gradient: Learning a POLICY directly (not a value table)
# Used in complex environments where Q-tables are too large

def softmax(x):
    exp_x = np.exp(x - np.max(x))
    return exp_x / np.sum(exp_x)

# Simple environment: choose action 0, 1, or 2
# Action 2 gives best reward on average
n_actions = 3
true_rewards = [1.0, 2.0, 5.0]  # expected rewards

# Policy parameters (will be learned)
np.random.seed(42)
theta = np.zeros(n_actions)
learning_rate = 0.1

print("===
...

📋 Quick Reference

Algorithm	Approach	Best For
Q-Learning	Learn value of (state, action)	Small, discrete environments
SARSA	On-policy Q-learning	Safer exploration
DQN	Q-learning + neural network	Atari games, large states
Policy Gradient	Learn policy directly	Continuous actions
PPO	Stable policy optimisation	Robotics, ChatGPT RLHF

💡 Pro Tip: ChatGPT uses Reinforcement Learning from Human Feedback (RLHF). Humans rank model responses, and RL fine-tunes the model to produce responses humans prefer. This is why ChatGPT sounds helpful — it was literally trained to maximise human approval!

🎉 Lesson Complete!

You can now build RL agents that learn from interaction! Next, learn how to deploy ML models to production.

Reinforcement Learning

✅ What You'll Learn

🎮 What Is Reinforcement Learning?

Try It: Multi-Armed Bandit

Try It: Q-Learning Grid Navigation

Try It: Policy Gradient (REINFORCE)

📋 Quick Reference

🎉 Lesson Complete!

Cookie & Privacy Settings