Inconsistent sensor mask when nebula moves

[vkrajnak]

Sensor mask is inconsistent when nebula moves making some nebula tiles visible and some empty tiles invisible. See screenshot below, game seed 1029303108. If sensor mask and tile type were calculated in reverse order, sensor mask would make sense.

Lux-Design-S3/src/luxai_s3/env.py

Lines 644 to 662 in 8188ad5

	state = self.compute_sensor_masks(state, params)
	# Shift objects around in space
	# Move the nebula tiles in state.map_features.tile_types up by 1 and to the right by 1
	# this is also symmetric nebula tile movement
	new_tile_types_map = jnp.roll(
	state.map_features.tile_type,
	shift=(
	1 * jnp.sign(params.nebula_tile_drift_speed),
	-1 * jnp.sign(params.nebula_tile_drift_speed),
	),
	axis=(0, 1),
	)
	new_tile_types_map = jnp.where(
	(state.steps - 1) * abs(params.nebula_tile_drift_speed) % 1 > state.steps * abs(params.nebula_tile_drift_speed) % 1,
	new_tile_types_map,
	state.map_features.tile_type,
	)

[CJBoey]

Addressed in issue #45

[vkrajnak]

While I do not dispute the fact that the order is consistent with the match resolution order, it is inconsistent with the way nebula reduces vision and empty tiles do not reduce vision. Nebula tiles may currently be temporarily visible and empty tiles temporarily invisible. The game rules mention no exception for vision reduction, hence the rules or the code should change.

[StoneT2000]

Thanks for pointing it out @vkrajnak. I think changing the functionality now will cause issues for existing bots and we already made one big balance patch. I think for now it is better to update the rules on this caveat.

[vkrajnak]

The decision is up to you, just bear in mind that either way some bots will issues - those relying on the bug may break if the code changes, those relying on what's written in the rules if the rules change.