Train an agent with Q-learning or SARSA. Adjust rewards, explore policies, test generalisation, and compare the learned route with breadth-first search.
Environment setup
Maze editor
Enable the editor, then click or drag across maze cells.
Learning parameters
Rewards and penalties
Ready
Wall
Start
Exit
Trap
Bonus
Agent
Best route
Display and interaction
Manual keys: W A S D or arrow keys.
Episode0 / 150
Total reward0.0
Success rate0.0%
Current steps0
Average reward0.0
Best route—
Unique explored1
Convergence0%
Wall collisions0
Trap encounters0
Explore / exploit0 / 0
Training time0.0s
Testing and challenge modes
Standard training
Test result: No test completed.
Generalisation: Not measured.
Shortest path: Not calculated.
Score: 0
Decision inspector
Click an open maze cell to inspect its learned action values and policy.
State: none selected
| Action | Q-value | Policy |
|---|---|---|
| Select a state. | ||
Q(s,a) ← Q(s,a) + α[r + γ max Q(s′,a′) − Q(s,a)]
Reward and steps history
Success and exploration
State visit heatmap
Algorithm comparison
Experiment history and achievements
| Time | Algorithm | Maze | Success | Best | Score |
|---|---|---|---|---|---|
| No saved experiments. | |||||
Live training log
Save, import, and export
How reinforcement learning works here
- The agent observes its current maze cell as a state.
- It chooses an action through exploration or exploitation.
- The environment returns a reward and next state.
- The selected algorithm updates an action value.
- Repeated episodes gradually produce a stronger policy.
Exploration versus exploitation
Exploration tries uncertain actions. Exploitation chooses the strongest learned action. Epsilon controls their balance.
Q-learning and SARSA
Q-learning learns toward the best possible next action. SARSA learns from the next action actually selected.
Reward design
Large exit rewards encourage completion. Step penalties encourage shorter routes. Excessive trap penalties can reduce useful exploration.
Knowledge check