Reinforcement learning: consequences arrive later

Picture it

An agent chooses actions that change what happens next. A reward may arrive much later, so learning must connect consequences to earlier decisions. A value function estimates future return; a policy chooses actions.

What to understand

• States, actions, rewards, returns, discounting, and MDPs
• Exploration, credit assignment, and on/off-policy learning

Follow the references

1. 1 · Course

   Hugging Face · Deep reinforcement learning course ↗

   Use Unit 1’s agent-environment loop and the first simple environment.