What will I learn here?

A discretized Q-table maps angle and angular velocity to three actions: push left, coast, or push right. Height-based rewards and ε-greedy exploration illustrate the core RL loop.

How is reward defined?

Each step rewards swing height (1−cos θ) with a bonus for high swings. Pushing incurs a small cost; extreme speed or angle ends the episode.

What are α, γ, and ε?

α is the learning rate, γ the discount factor, and ε the exploration rate for random actions. Adjust sliders and watch how returns and the policy change.

NN Classifier RL Agent Conv Vision Attention Playground

Playground

Swing RL

Q-learning discovers when to push and when to coast—just like pumping a swing!

Episode0

Learning rate αDiscount γExploration εSpeed

Show formulas

Training settings

Push on the way down, coast on the way up—the Q-table learns this timing from rewards.

Learning rate α: How much each Q update moves. Large values learn fast but can oscillate.
Discount γ: How much future rewards matter now. Closer to 1 weights distant rewards more.
Exploration ε: Random push/coast chance. High ε tries many rhythms; low ε sticks to what worked.

Swing setup

Rope, friction, and wind change the challenge

Reward is swing height (1−cos θ). The agent learns left/right pushes to build amplitude.

Rope:: 1.20 m
Friction:: 0.035
Push:: 2.2

Push opposite to motion near the bottom to add energy
Coasting near the top usually helps

Swing simulator

Purple robot = agent · bar = height

Steps this episode

0.0000

Episode return

0.0000

Height

0.0036

High swings

0.0000

θ -4.9°ω -0.17

Episode return

Higher swings raise the curve

Start training to see per-episode returns

Related lessons