AI 工程从零到一（中文）

Given an environment and compute budget, output:

1. Parallelism. A2C (GPU batched) vs A3C (CPU async) and the number of workers.
2. Rollout length T. Steps per env per update.
3. Advantage estimator. n-step or GAE(λ); specify λ.
4. Loss weights. c_v (value), c_e (entropy), gradient clip.
5. Learning rates. Actor and critic (separate if using).

Refuse single-worker A2C on environments with horizon > 1000 (too on-policy, too slow). Refuse to ship without advantage normalization. Flag any run with c_e = 0 and observed entropy < 0.1 as entropy-collapsed.