Some question about the MIP_DQN.py

[Mrdawnzero]

Hello, I'd like to understand how to use the "Actor_MIP" class in the provided code. This part is mentioned as a highlight in your paper, but it seems that the class is not called or utilized in the code. I'm interested in learning how this class should be employed.

[Kacjer]

Same question from me. Thanks.

[EnergyQuantResearch]

This class is called after training is finished when doing real-time implementation. After training, you get a well trained Q network (critic network) and actor network. We don't use this actor network but import the trained Q network into the class. Based on this Q network, we get the optimal action by solving the method provided in the actor class "predict_best_action" by inputting the state.

[lmd123123]

Hello, how to test after training? There is no such link in the code. Could you please provide the process of running the test online described in your article in the code? How to use the trained Q-network in conjunction with MIP. Looking forward to your reply!

[Mrdawnzero]

This class is called after training is finished when doing real-time implementation. After training, you get a well trained Q network (critic network) and actor network. We don't use this actor network but import the trained Q network into the class. Based on this Q network, we get the optimal action by solving the method provided in the actor class "predict_best_action" by inputting the state.

I want to know how you determine that the Q network is well-trained. Additionally, I use the OMLT package to model the trained network with MIP, but I find that the results are not as good as the Q network. I would like to know how you adjust the parameters

[WangDing1030]

Hello, Mr. Hou Shengren. After loading the trained Critic network into the Actor_MIP, it shows that the block is a 'NoneType' object and is not callable, and the formulation is also invalid. How should this be corrected?

[QIUQIUSCI]

Hello, how to test after training? There is no such link in the code. Can you provide in your code the procedure for running tests online as described in your article? How to use a trained Q-network with MIP. Looking forward to your reply! Why don't you open source important content

[Mrdawnzero]

This class is called after training is finished when doing real-time implementation. After training, you get a well trained Q network (critic network) and actor network. We don't use this actor network but import the trained Q network into the class. Based on this Q network, we get the optimal action by solving the method provided in the actor class "predict_best_action" by inputting the state.在执行实时实现时，在训练完成后调用该类。经过训练，您将获得训练有素的 Q 网络 （critic 网络） 和 actor 网络。我们不使用这个 actor 网络，而是将训练好的 Q 网络导入到类中。基于这个 Q 网络，我们通过输入状态来求解 actor 类 “predict_best_action” 中提供的方法，从而获得最佳动作。

I want to know how you determine that the Q network is well-trained. Additionally, I use the OMLT package to model the trained network with MIP, but I find that the results are not as good as the Q network. I would like to know how you adjust the parameters我想知道您如何确定 Q 网络训练有素。此外，我使用 OMLT 包通过 MIP 对经过训练的网络进行建模，但我发现结果不如 Q 网络。我想知道您如何调整参数

Can I talk to you? I also found that it is not possible to achieve the limits of the author's paper我可以和你谈谈吗？我还发现，不可能达到作者论文的极限

how can I talk to you，can you give me your email？

[m18326612413]

This class is called after training is finished when doing real-time implementation. After training, you get a well trained Q network (critic network) and actor network. We don't use this actor network but import the trained Q network into the class. Based on this Q network, we get the optimal action by solving the method provided in the actor class "predict_best_action" by inputting the state.

我想知道你是如何确定 Q 网络训练有素的。此外，我使用 OMLT 包对训练有素的网络进行 MIP 建模，但我发现结果不如 Q 网络。我想知道你是如何调整参数的

Hello! I'd like to ask how you managed to import a trained Q-network into your class, and how you used the OMLT package to build a MIP model of the trained network? 您好 我想请教一下 您是如何解决将训练好的 Q 网络导入到类中的，如何OMLT 包通过 MIP 对经过训练的网络进行建模？ 您有相关的代码？

[ShengrenHOU]

Thanks for pointing this out. You were right: in the older version of the repository, Actor_MIP was defined but not properly wired into the default workflow.

This has now been fixed in the current main branch:

• Actor_MIP is integrated into the default training/evaluation flow
• the training script was refactored to make the call path explicit
• runtime checks and tests were added around this logic

Please note that the full solver-backed Actor_MIP path still requires the local optimization stack (pyomo, omlt, ONNX export support, and gurobi) to be installed.

For the current implementation, the relevant entry points are in MIP_DQN.py, especially Actor_MIP, build_actor_mip, and train_one_seed. The README was also updated accordingly.

I’m closing this issue because the original Actor_MIP is defined but not actually used problem is now addressed in main. If you still hit a specific runtime error with the solver-backed path, please open a new issue with your environment and traceback.