Dyca

DyCA (Dynamic Clustering Adaptive) Skill

Reference: LLM4AD_Next instance-aware multi-pool evolution model (DyCA, Dynamic Clustering Algorithm).

1. Method Essence

DyCA is an instance-aware search method. Instead of treating all problem instances uniformly, it dynamically partitions the training instances into clusters and maintains heterogeneous algorithm pools specialized to those clusters. The search then allocates LLM effort across pools based on observed difficulty and cluster stability.

Three heterogeneous pools:

Core mechanisms:

• Per-instance evaluation and cluster-specific fitness
• Dynamic reclustering based on Adjusted Rand Index (ARI) stability
• SOS (Save Our Specialist) inter-cluster collaboration to escape local optima
• Macro (ARI-based) and micro (gap-based) resource allocation

2. Recommended Parameters

See params.yaml in this directory for the recommended parameter configuration.

What Happens During Evolution

1. Build instance feature vectors using n_anchors anchor algorithms
2. Cluster instances via clustering_method into n_clusters clusters
3. Every recluster_interval generations, check ARI stability; recluster if unstable
4. Each generation produces offspring_per_generation new individuals
5. Allocate effort across generalist / specialist / complementary pools
6. Trigger SOS when a cluster stagnates for sos_stagnation_threshold generations
7. using_mode=true freezes clustering and runs only specialist evolution (mature clusters)

Common Pitfalls

• Too few/too many clusters → poor specialization; tune n_clusters
• Reclustering too often → instability; raise recluster_interval / lower ari_threshold
• Anchors too few → weak feature vectors; increase n_anchors
• Complementary pool starved → raise base_complementary_ratio
• Clusters already stable → enable using_mode to stop reclustering

4. Acceptance Criteria

• Instances cleanly clustered; each cluster has a specialist pool
• Reclustering only on ARI instability (not every generation)
• Resource allocation responds to per-cluster difficulty
• SOS triggers on stagnation and recovers a better solution
• Using mode works for already-stable datasets