Home / Skills tracks / Reinforcement learning
At schoolPost 16 and adults

Reinforcement learning

Agents that learn from reward rather than examples. Build a bandit, train Q-learning on a grid world, then watch a badly written reward get farmed.

3 units, about 3 hours. Free.

1

Learning from reward: the bandit problem

Explain how an agent learns from reward alone, and show why pure greed locks onto a bad choice.

2

Q-learning: values before actions

Implement Q-learning on a grid world and explain the update rule and the discount factor.

3

Reward is a specification, and it gets gamed

Demonstrate reward hacking in code and explain why agents optimise what you wrote, not what you meant.