Research

Background, algorithm details, and benchmarks.

Sections

• Background - Evolution Strategies history and theory
• EGGROLL Algorithm - Deep dive into the algorithm
• Benchmarks - Performance comparisons and speedups

Overview

Eggroll Trainer implements the EGGROLL algorithm, a novel Evolution Strategy that achieves 100x speedup over naïve ES methods through low-rank perturbations.

Key References

• EGGROLL Paper/Blog
• GitHub Implementation

Key Innovation

Low-rank perturbations reduce memory and computation:

• Memory: O(mn) → O(r(m+n)) for matrices W ∈ R^(m×n)
• Computation: O(mn) → O(r(m+n))
• Speedup: ~100x for typical models

Yet still achieves high-rank updates through population averaging!

Next Steps

• Read Background for ES theory
• See EGGROLL Algorithm for details
• Check Benchmarks for performance data