beam_cvar/README.md

beam_cvar

Summary

• Best FA: +0.330
• Oracle calls: 4,358
• Wall time: 4 hours
• Epochs: 191

Algorithm: Beam Search Active Loop

1. Beam search of width=8, guided by ensemble mean μ (deterministic scoring)
2. Beam produces K=25 candidates per epoch directly (no separate Thompson selection)
3. Evaluate K with FA oracle, CVaR fine-tune DDPP, retrain ensemble

Key difference from thompson_cvar: search is directed by μ deterministically. No per-candidate exploration — always prunes toward ensemble mean.

Config

• selection: beam, K=25, beam_width=8
• tau_mode=cvar, tau_quantile=0.80, ddpp_steps_per_epoch=50
• wall_budget_sec=14400 (4hr)

Files

• active_loop_log.jsonl, oracle_progress.csv, results.json