kerboul
45b37f9357
- Replace Manhattan distance with BFS distances (actual maze path length) - Switch from NN to direct sequence genome (what YouTube Shorts actually use) - Mutation-only GA (crossover breaks positional maze paths) - Two-rate mutation: low rate before best-step (preserve prefix), high after (explore tail) - Auto-seed selection finds maze with short BFS path - Default maze 15x15, cell_size=36, max_steps=150 - Typically converges in 5-15 generations
34 lines
898 B
Python
34 lines
898 B
Python
from dataclasses import dataclass
|
|
|
|
@dataclass
|
|
class MazeConfig:
|
|
# Maze geometry
|
|
maze_cols: int = 15
|
|
maze_rows: int = 15
|
|
cell_size: int = 36
|
|
wall_width: int = 2
|
|
|
|
# GA / Training
|
|
population: int = 100
|
|
max_steps: int = 150
|
|
elite_frac: float = 0.30
|
|
n_elite: int = 3 # few elites → maintains population diversity
|
|
mutation_std: float = 0.08
|
|
mutation_rate: float = 0.35
|
|
temperature: float = 0.7 # softmax sampling temperature (lower = more greedy)
|
|
|
|
# Neural net — 14 inputs: 4 walls + 2 goal dir + 2 pos + 4 last-action + 1 blocked + 1 revisit
|
|
n_inputs: int = 14
|
|
n_hidden: int = 16
|
|
n_outputs: int = 4
|
|
|
|
# Fitness
|
|
goal_bonus: float = 1000.0
|
|
step_penalty: float = 0.01
|
|
|
|
# Rendering
|
|
fps: int = 60
|
|
trail_decay: float = 0.88
|
|
window_title: str = "Maze ML — GA Visualizer"
|
|
hud_font_size: int = 18
|