LIMIT: Learning Interfaces to Maximize Information Transfer

This is a repository for our paper, "LIMIT: Learning Interfaces to Maximize Information Transfer". We include:

• Implementation of LIMIT
• Custom 2D ("Lights") environment to test LIMIT with an Align Human (see our paper for more information)

Requirements

Requirements are listed in requirements.txt:

• python3
• pytorch $\ge$ 1.12.1
• numpy $\ge$ 1.24.2
• matplotlib $\ge$ 3.7.1
• scipy $\ge$ 1.9.3

Requirements can be installed using pip:

pip install -r requirements.txt

Instructions

To run a demonstration of LIMIT, run python main.py. LIMIT will train with the Align human, and the result should be similar to the plot below:

You can also provide arguments to adjust the behavior of LIMIT:

• --episodes: change the number of episodes that AlignHuman uses LIMIT. Default is 200
• --online: passing this flag enables LIMIT to learn online. Since AlignHuman updates after each interaction, do not an expect a significant performance boost from this. When interacting with real humans, LIMIT performs better when it learns online.