Dear Matthias, great work! I have a curiosity, could you kindly tell how the agent training will be carried out? Also, how would you set the rewards for such RF tasks? Thank you!
Dear Alessandro, thank you for your question.
The training can be performed with different commonly used agents. Currently, I investigate which agent, and parameter configuration is best suited. The simulation is controlled via an interface and is part of the environment. The current layout is the input of the simulation, the values of the target variables are the output. These are normalized and weighted and then used as a reward for the training. Furthermore, invalid solutions are penalized / not rewarded.
Hi Matthias
thanks for the insights.
Have you already been able to validated whether this approach actually leads to better results than a purely manual approach?
And how do you exactly deal with the complex requirements regarding MEP (TGA-Anforderungen)? Are you just looking at connection points?
Best
Matthias
I haven`t directly validated the approach compared to pure manual planning. But I compared it with expectations I defined based on the transport intensity between the functional units. The approach was able to fulfill these expectations. A comparison between this approach, a manual, and a heuristic approach is planned as future work.
The MEP requirements are an interesting and challenging point. Since I`m focussing on the early stages of layout planning I dont consider all MEP requirements in detail. At this moment, I only incorporated a few requirements. For that, each position of the grid has defined properties regarding the requirements. Furthermore, the functional units have defined requirements that need to be fulfilled. The fit of the requirements of both parts is compared and transferred to the reward function. In the future, I will investigate how to incorporate the requirements in more detail to enhance the quality of the solution.
Dear Matthias, great work! I have a curiosity, could you kindly tell how the agent training will be carried out? Also, how would you set the rewards for such RF tasks? Thank you!
Dear Alessandro, thank you for your question.
The training can be performed with different commonly used agents. Currently, I investigate which agent, and parameter configuration is best suited. The simulation is controlled via an interface and is part of the environment. The current layout is the input of the simulation, the values of the target variables are the output. These are normalized and weighted and then used as a reward for the training. Furthermore, invalid solutions are penalized / not rewarded.
Hi Matthias
thanks for the insights.
Have you already been able to validated whether this approach actually leads to better results than a purely manual approach?
And how do you exactly deal with the complex requirements regarding MEP (TGA-Anforderungen)? Are you just looking at connection points?
Best
Matthias
Hello Matthias,
I haven`t directly validated the approach compared to pure manual planning. But I compared it with expectations I defined based on the transport intensity between the functional units. The approach was able to fulfill these expectations. A comparison between this approach, a manual, and a heuristic approach is planned as future work.
The MEP requirements are an interesting and challenging point. Since I`m focussing on the early stages of layout planning I dont consider all MEP requirements in detail. At this moment, I only incorporated a few requirements. For that, each position of the grid has defined properties regarding the requirements. Furthermore, the functional units have defined requirements that need to be fulfilled. The fit of the requirements of both parts is compared and transferred to the reward function. In the future, I will investigate how to incorporate the requirements in more detail to enhance the quality of the solution.
Best
Matthias