I observed several models that are trained on teaching robots to do grasping tasks, and I noticed that the success rate varies within 7 % in each training session. Also, it can be noted that the validation of each model's training session may have an impact on the performance. It implies that it is neither steady nor efficient. That's why the Deep-RL model's training might take several sessions to find the optimal learning. Also, if we take those several training sessions and put them to the test in various scenarios, we can see that the success rate varies.
- Is the RL-framework impacted by image preprocessing, or is it because the RL is fighting to find the best policy, which may not guarantee proficiency and stability in each learning session?
Well, data and its preprocessing have an impact on the performance. However, as you mentioned, it depends on the network to find the best policy. Note that you should make sure you have observed the requirements, including
1. Enough epochs of training; otherwise the results may even vary widely, depending on the severity.
2. Dependence on the initial seeds and randomness; this certainly has an impact, such that not only in RL but also in other kind of ML it affects results.
3. Using final weights, not intermediate weights
Dear Marwan Qaid Mohammed
I recommend the following articles, these might be an valuable asset:
1. https://machinelearningmastery.com/different-results-each-time-in-machine-learning/
2. https://towardsdatascience.com/best-practices-for-reinforcement-learning-1cf8c2d77b66
3. https://wiki.pathmind.com/deep-reinforcement-learning
Kind Regards
Qamar Ul Islam
Dear Mostafa Kiani Shahvandi
I would really appreciate your valuable comments.
Dear Qamar Ul Islam
Reading your recommended articles seems great and useful; Thank you very much for sharing this information.
- Badges
- Science topic
- Similar topics
- Organizational Psychology
- Training