HEMO
A framework that treats the harness in automatic operator design as the object of self-evolution, achieving self-modification of the search process for multi-objective combinatorial optimization through single-module updates, discounted UCB scheduling, and persistent memory.
Timeline (last 90 days)
HEMO proposes making the harness of automatic operator design itself the optimization target, and maintaining a persistent harness state that is updated module by module.
HEMO's harness revision logic covers four levels: which piece of experience to retrieve, which operator to improve, which parents to select, and how to orchestrate the search steps.
HEMO adopts a single-module update design to reduce the coupling of the search space, making the effect of each change easier to attribute.
HEMO uses a discounted UCB scheduler to direct adaptation resources toward more promising modules based on changes in each module's utility.
HEMO uses a persistent memory file system that preserves candidate results and diagnostic feedback across harness states, enabling subsequent states to build on previously accumulated optimization experience.
HEMO was validated on eight MOCO tasks and two structurally different evolutionary backbones.
Experimental results show that HEMO performs strongly compared with traditional evolutionary methods, neural network methods and LLM-based AOD baselines, and can effectively transfer to related tasks.
The cost of HEMO is increased system complexity, and how the additional overhead brought by module updates and scheduling varies with task scale still requires follow-up work to answer.
All Facts (8)
HEMO proposes making the harness of automatic operator design itself the optimization target, and maintaining a persistent harness state that is updated module by module.
50%UnverifiedHEMO's harness revision logic covers four levels: which piece of experience to retrieve, which operator to improve, which parents to select, and how to orchestrate the search steps.
50%UnverifiedHEMO adopts a single-module update design to reduce the coupling of the search space, making the effect of each change easier to attribute.
50%UnverifiedHEMO uses a discounted UCB scheduler to direct adaptation resources toward more promising modules based on changes in each module's utility.
50%UnverifiedHEMO uses a persistent memory file system that preserves candidate results and diagnostic feedback across harness states, enabling subsequent states to build on previously accumulated optimization experience.
50%UnverifiedHEMO was validated on eight MOCO tasks and two structurally different evolutionary backbones.
50%UnverifiedExperimental results show that HEMO performs strongly compared with traditional evolutionary methods, neural network methods and LLM-based AOD baselines, and can effectively transfer to related tasks.
50%UnverifiedThe cost of HEMO is increased system complexity, and how the additional overhead brought by module updates and scheduling varies with task scale still requires follow-up work to answer.
50%