DL4J to learn the task execution steps

DL4J
1

Dear members,

I am trying to model a network which can take :-

  1. Input - A Target Goal (e.g. Go To Cinema Hall)
  2. Output - A sequence of steps (e.g. Go To Bus Stop → Onboard A Bus → Drop at Location X → Walk 100 meters.

My training data is a similar kind of Objective - Task Sequence Mappings.

Can anyone please suggest a suitable approach using DL4J to address this problem and model suitable solutions?

P.S. I am new to DL4J and just getting started with the framework.

Thanks in advance,

Cheers,
Amar

2

HI Amar:

Are you referencing any papers when attempting to do this?
This doesn’t really sound like a well defined problem.
Trying to just use supervised learning to tell a machine to pick an instruction to do doesn’t really sound appropriate to even try.

You could look in to RL for this, but generally if it’s not a well defined simple problem like classifying something or predicting a value you’re not really going to make much progress.

3

Hi Adam,
Thanks for replying to my post. My problem is that, I have an input data where I have gathered the steps to address a goal. E.g. Goal - Cook A Turkey , has steps

  1. Marinate 2. Light_Burner 3. Roast 4. Cut in Pieces

Another goal could be Cook Chicken, has steps 1. Marinate 2. Light_Burner 3. Boil 4. Fry and so on. Given such a set of a large amount of input comprising of Goals and steps mapping, can the system lean how to address a goal such as “Cook a duck”.

My input data is in the form of a hierarchical eclipse EMF-based model, and I was wondering if the machine can learn the EMF model which has such goal and steps mapping. The reason I am looking at deeplearning is that I see that there is a possibility to represent a model (EMF model) as a neuron and then let the learning happen acorss multiple neurons.

Could you please suggest if this sounds good ? Or you think there is a problem getting this done via dl4j ??

Thanks and cheers,
Amar

4

It sounds like we’re not really applying the fundamentals of machine learning here.
If you’re not referencing a specific paper or something I would recommend sticking with a rules based approach. Your problem is not well defined enough to really go anywhere.

There are limited forms of NLP you can do, but again you would need to implement a paper to do that.
I would suggest going back to the drawing board and coming back with a more specific approach before I can make a recommendation on how to proceed.

5

Thanks again Adam.
This is the paper that I would like to refer.

6

@amarbanerjee23 This: https://arxiv.org/pdf/1801.02294.pdf

isn’t really good for actions though. It’s clearly targeting items. Those domains aren’t remotely close.
The closest you might get is a session based recommendation: Session-Based Recommendations | Papers With Code

Even then it’s a bit of a stretch. How big is your dataset? Have you built it yet?