XdG combined with Machine learning?

[ShanglunLi]

Thank you for providing such an interesting paper to read.

In the paper, you mentioned that XdG cannot implement transfer learning yet, because transfer learning is based on previous experience and knowledge. What if you combine XdG method with machine learning? Use machine learning to treat previous tasks as the training tasks, so that when you encounter a new task, it can based on the training set and XdG to make a more accurate decision. Is there a conflict between these two methods?

Thanks.

[nmasse]

Since the agent might not always have access to past tasks, there's been a lot of recent work of using generative models that can replay past tasks (i.e. https://papers.nips.cc/paper/6892-continual-learning-with-deep-generative-replay.pdf). Was this what you had in mind? If so, it was a good idea!

[jheng18]

I am curious if it's possible that we use XdG and ML methods to generalize some knowledge/rules, compare the new task to the knowledge/rules it already had and update the knowledge/rules each time? I think we human do forget minimal things but tend to remember the generalized rules.