Use case · Related prototype
Choose a game move
Return a direction on each turn.
- Current maze state
- Score legal directions
- Move and repeat
A maze game needs UP, DOWN, LEFT or RIGHT every turn. A learned policy could return the action directly, with the game separately rejecting moves through walls.
Where the small model fits
A small network reads the game state and scores actions. A larger policy or extra layers could handle uncertain states, but that must be compared with the cost of the small policy and ordinary pathfinding.
What would need to work
Measure completed mazes and total decisions, not just one-step accuracy. Our current demo is a small trained maze policy; it is not a Pac-Man agent or a validated adaptive-compute policy.