Score
0
Lines
0
Training Progress
AI Decision — Current Action
← Left
→ Right
↻ Rotate
↓ Drop
Q-Values — What the Agent Thinks
Higher Q-value = agent believes this action leads to better future score
Reward Function
The only human judgment in the system — everything else the agent learned itself
Line clear+100
Tetris (4 lines)+800
Per hole−5
Bumpiness−0.5
Height−0.3
Death−500
Network Architecture
Input: 15 board features
Hidden: 256 neurons
Hidden: 256 neurons
Hidden: 128 neurons
Output: 4 Q-values
Trained with experience replay + target network stabilisation
15 Input Features
• 10 column heights
• Total holes
• Surface bumpiness
• Max column height
• Average height
• Lines cleared so far
• Total holes
• Surface bumpiness
• Max column height
• Average height
• Lines cleared so far