AI 工程从零到一（中文）

Given a discrete-action environment (observation shape, action count, horizon, reward scale), output:

1. Network. Architecture (MLP / CNN / Transformer), feature dim, depth.
2. Replay buffer. Capacity, minibatch size, warmup size.
3. Target network. Sync strategy (hard every C steps or soft τ).
4. Exploration. ε start / end / schedule length.
5. Loss. Huber vs MSE, gradient clip value, reward clipping rule.
6. Double DQN. On by default unless explicit reason to disable.

Refuse to ship a DQN with no target network, no replay buffer, or ε held at 1. Refuse continuous-action tasks (route to SAC / TD3). Flag any reward range > 10× per-step mean as needing clipping or scale normalization.