📝 The tricks in Learn to optimize in TNCO sycamore

下面 theta 表示 TNCO任务的“解”，它是一个表示量子电路收缩顺序的tensor， 计算 `theta.argsort()` 就能获得有序的 edge_id，表示依次收缩某一条边。

1. 维持了两个 ReplayBuffer，一个保存了模型实时迭代产生的theta，另一个保存了得分较好的theta。这样让得分好的theta不至于被 ReplayBuffer 的 FIFO 规则删掉。

计算 `keep_score`
https://github.com/AI4Finance-Foundation/ElegantRL_Solver/blob/41a58d0ecb9daeddfa635d19a3741b0a29162342/rlsolver/rlsolver_learn2opt/tensor_train/TNCO_H2O.py#L418-L419

根据 `keep_score` 找出需要保存到 buffer0 的 theta
https://github.com/AI4Finance-Foundation/ElegantRL_Solver/blob/41a58d0ecb9daeddfa635d19a3741b0a29162342/rlsolver/rlsolver_learn2opt/tensor_train/TNCO_H2O.py#L372-L375

2. 从更好的解附近开始搜索

`historical_theta` 是一个得分好的theta，我们对它加上噪声并在它附近开始迭代
https://github.com/AI4Finance-Foundation/ElegantRL_Solver/blob/41a58d0ecb9daeddfa635d19a3741b0a29162342/rlsolver/rlsolver_learn2opt/tensor_train/TNCO_H2O.py#L402-L408

我们也从 保存了较好的 theta 的 buffer0 里随机选出的theta 开始迭代
https://github.com/AI4Finance-Foundation/ElegantRL_Solver/blob/41a58d0ecb9daeddfa635d19a3741b0a29162342/rlsolver/rlsolver_learn2opt/tensor_train/TNCO_H2O.py#L410-L419

3. 在训练迭代器之后，使用迭代器进行推理

可以看到
- 训练阶段 `opt_loop(..., unroll=unroll, if_train=True)`
- 推理阶段 `opt_loop(..., unroll=unroll * 2, if_train=False)`
https://github.com/AI4Finance-Foundation/ElegantRL_Solver/blob/f2d89778f9d9bb20ddfcfa72521f3e0265b8befe/rlsolver/rlsolver_learn2opt/tensor_train/TNCO_H2O.py#L472-L479