LightsOut

[zdu863]

Implementation of the Lights Out game as an env.

• Each action selects one cell, toggling that cell plus its orthogonal neighbors. The goal is to turn all lights off within a step limit.
• Observations are binary grid states, actions are cell indices, and rewards encourage solving efficiently.
• It supports vectorized training and raylib rendering.
• The environment includes a curriculum mechanism driven by an EMA of recent outcomes.
• This EMA adjusts scramble probability over time, so training can start easier and become harder as policy performance improves.