| Average reward | 0.00 |
|---|---|
| Best reward | 0.00 |
| Average steps | 0.00 |
| Policy efficiency | 0% |
| Exploration | 35% |
| Q-value change | 0.000 |
| Best route length | — |
| Shortest-path gap | — |
| Rank | Reward | Steps | Mode | Date |
|---|---|---|---|---|
| Complete a successful mission to record a score. | ||||
States, actions, and rewards
Each grid position is a state. The agent chooses a movement action and receives a reward. It learns which choices produce valuable long-term outcomes.
Q-learning update
Q(s,a) ← Q(s,a) + α[r + γ max Q(s′,a′) − Q(s,a)]. Learning rate α controls update size. Discount factor γ controls future reward importance.
SARSA update
SARSA updates using the next action actually selected by the policy. It often learns more cautious behaviour when exploration can trigger traps.
Exploration versus exploitation
Exploration tries uncertain actions. Exploitation chooses the best-known action. Epsilon-greedy balances both and gradually reduces exploration through decay.
Reward shaping guidance
Large treasure rewards motivate collection. Move costs encourage shorter routes. Strong trap penalties improve safety but can make the policy overly cautious.