PACE: Primitive-Aware Code Evolution for Automated Algorithm Design
PACE keeps useful algorithm fragments alive after their original programs fail.
The paper proposes Executable Algorithmic Primitives, persistent code units that can be retained and transferred across evolving programs. Its operators are designed to preserve those primitives structurally during algorithm search. Primitive selection is guided with Thompson sampling based on parent-relative performance gains, without extra evaluation datasets. Experiments across four tasks found competitive algorithms while preserving valuable components. ArXiv · AI/CL/LG's note
The paper proposes Executable Algorithmic Primitives, persistent code units that can be retained and transferred across evolving programs. Its operators are designed to preserve those primitives structurally during algorithm search. Primitive selection is guided with Thompson sampling based on parent-relative performance gains, without extra evaluation datasets. Experiments across four tasks found competitive algorithms while preserving valuable components. ArXiv · AI/CL/LG's note
score 4