Controls
1×
Live
- Episode
- 0
- Step
- 0
- Last reward
- 0
- Episode length
- 0
- Buffer
- 0 / 10000
- Epsilon
- 1.000
- Loss
- —
- Mean Q
- —
Training curves
CartPole live
The pole's tip color shifts from blue (stable) to red (falling). Action is rendered briefly as a small arrow on the cart.
About
DQN (Deep Q-Network) trains a small MLP to estimate Q-values for CartPole's 4D state. The agent's goal is to maximize the cumulative reward by keeping the pole balanced. The target network stabilizes training, and ε-greedy exploration gradually shifts toward greedy exploitation.
See src/algorithms/dqn-agent.ts for the agent,
src/envs/cartpole.ts for the environment, and
src/ui/app.ts for this UI's entry point.