Attempt to Optimize Wolf-Sheep Performance

[gadhvirushiraj]

Summary

Attempt at optimizing Sheep Agent movement using Mesa's PropertyLayer and vectorized operations to determine the best moves.

Motive

The current movement logic in the wolf-sheep model performs redundant checks when agents search their neighborhood for valid moves. This optimization builds on changes from #2503.

Implementation

Two PropertyLayer: sheep_pos and grass_pos are added and updated. The agent determines the best move by extracting a subsection (neighbor radius - 1) of the PropertyLayer and applying vectorized boolean operations. Refer Issue #2565.

However, performance drops for Wolf-Sheep (large), likely because the overhead of updating a larger grid offsets the benefits. A larger grid with higher agent density maybe yield better results, I think (not sure).

Usage Examples

--

Additional Notes

I tested multiple variations of this approach. Full-layer vector operations proved expensive compared to the original get_neighborhood method.

An, imporvment can be a better way to update grass_pos (grass tracking layer) and updating wolf_pos directly from Agent's move(). Edit : Implemented

[github-actions]

Performance benchmarks:

Model	Size	Init time [95% CI]	Run time [95% CI]
BoltzmannWealth	small	🔵 -1.0% [-1.5%, -0.5%]	🔵 +0.3% [+0.2%, +0.5%]
BoltzmannWealth	large	🔵 -2.6% [-3.4%, -1.8%]	🔵 -1.5% [-2.2%, -0.7%]
Schelling	small	🔵 -2.1% [-2.3%, -1.9%]	🔵 -0.2% [-0.3%, -0.1%]
Schelling	large	🔵 -0.4% [-9.0%, +8.7%]	🔵 +1.1% [-0.4%, +2.5%]
WolfSheep	small	🔵 -1.4% [-1.9%, -0.9%]	🔵 -2.9% [-9.0%, +4.0%]
WolfSheep	large	🔵 -0.5% [-0.9%, -0.2%]	🔴 +40.5% [+39.1%, +42.2%]
BoidFlockers	small	🔵 +0.8% [+0.1%, +1.4%]	🔵 -0.5% [-0.7%, -0.3%]
BoidFlockers	large	🔵 -0.1% [-0.7%, +0.5%]	🔵 +0.1% [-0.1%, +0.4%]

[github-actions]

Performance benchmarks:

Model	Size	Init time [95% CI]	Run time [95% CI]
BoltzmannWealth	small	🔴 +5.9% [+4.6%, +7.1%]	🔵 +0.2% [-0.1%, +0.4%]
BoltzmannWealth	large	🔵 +17.8% [-0.7%, +53.8%]	🔵 +3.4% [+0.0%, +6.9%]
Schelling	small	🔵 +0.5% [+0.2%, +0.7%]	🔵 +0.4% [+0.3%, +0.6%]
Schelling	large	🔵 +4.8% [+0.9%, +11.3%]	🔵 +4.6% [+2.1%, +7.5%]
WolfSheep	small	🔴 +5.2% [+4.6%, +5.7%]	🔴 +163.2% [+144.1%, +182.2%]
WolfSheep	large	🔴 +4.7% [+3.6%, +5.7%]	🔴 +153.9% [+148.5%, +159.1%]
BoidFlockers	small	🔵 +0.2% [-0.6%, +1.2%]	🔵 -1.0% [-1.1%, -0.8%]
BoidFlockers	large	🔵 -1.4% [-1.9%, -0.9%]	🔵 -0.9% [-1.1%, -0.7%]

[EwoutH]

@gadhvirushiraj Thanks a lot for your PR! Our benchmarks currently show a performance decrease, both the small and large size. What would you like us to do for this PR?

[gadhvirushiraj]

Hello @EwoutH,

I've been digging into the recent performance drop and noticed that it doesn't seem from updating positions in PropertyLayer, but rather from the operations within move(). Am I overlooking something here, is there a better way?

I'm open to exploring alternative approaches if you have any suggestions. This PR has been a good learning opportunity for me in getting to grips with Mesa's modeling :) Thanks

[EwoutH]

I would start with profiling both the old and new implementation, to see where compute time is spend in both.