Back to GalleryQ-Learning
This module visualizes tabular Reinforcement Learning. An agent learns to navigate a grid-world layout, avoiding obstacles to reach a high-reward goal state. The agent maintains a Q-Table representing the expected utility of taking actions in each state.
Key Parameters & Visual Influence:
• RL Algorithm: Q-Learning (off-policy) updates values assuming the greedy optimal action is taken next. SARSA (on-policy) updates values using the actual action chosen by the epsilon-greedy policy, making it more conservative.
• Learning Rate (α) & Discount Factor (γ): Control the update step size and the importance of future rewards vs immediate rewards.
• Epsilon (ε) & Decay: Control the trade-off between exploring random paths and exploiting the best-known route. Epsilon decay gradually reduces exploration as the agent learns.
Dataset Studio
CSV format: Headers in first row, last column = label/target. All values must be numeric.
Example: x1, x2, label
Hyperparameters
AlgorithmChoose off-policy Q-Learning or conservative, on-policy SARSA updates.
Grid SizeCell dimensions of the grid world environment.6
412
Learning Rate (α)Temporal Difference updates step size (learning rate).0.10
0.011
Discount Factor (γ)Discount factor (gamma) defining importance of future rewards.0.950
0.50.999
Epsilon (Explore)Probability of random exploration vs exploitation.0.30
0.011
Max EpisodesMaximum reinforcement episodes to run training loops.500
505000
6 / 7 parameters active