Create an online version of the plan recognizer.

Just port the code.

Read and present to us the main idea of SAS+ representation.
Presentation date: April 13.

Train an LSTM to receive the encoded representation of a series of states and predict the goal.

Alternatively, we should create a network with a CNN that processes the input image and creates an embedding for the LSTM layer.

We will use this LSTM to compare to our current results.

Take photos of an and use our approach to make a good demo.
This issue is just to remember that we shall try this one more time before discarding it.

Are inputs from sxnor limited to {-1,0,1}?
If they are, the following code is a simplified sxnor:

def sxnor(se,so):
    if se == so:
        return 1 if se == 1 else -1
    return 0

Research about it and check if it can be a better way to create the image representation.
A good start: GANs

The observations must be states and not actions so we can use the recognizer in the demo.

• Find what is being used to make such auto-train;
• Use evolutionary algorithms to find the best network topology.

We need to refine the user interface and connect with the goal recognizer in the server.
The following image is a suggestion of the demo workflow:

We need to discuss:

• if the candidate goals are generated by:
  • client (randomly or user-defined)
  • server (randomly or file-based).
• How to send the new state (image?).

Complete task of implementing FDR conversion. Still ways to go in both research and implementation of this technique.

Currently, it is a mess. Transform generate_pddl.py in a class and move some of the methods to other files.

Follow the instructions on PAL-18.

Due on May 18.

Submission Procedure

We solicit workshop paper submissions relevant to the above call of the following types:
Long papers -- up to 8 pages + unlimited references / appendices
We will accept papers in any of the IJCAI, ICML, AAMAS, or NIPS formats. Submissions are not anonymous and should include author information.
IJCAI authors-kit

Use yield when generating the actions.

We need to decide which of the repos we will stick with.
The (possibly) ideal solution is to make latplan a submodule of this repo.