Prototype of a game where a reinforcement learning agent is trained through natural language instructions. This is a research project based at the Montreal Institute for Learning Algorithms (MILA).

Requirements:

• Python 3.5+
• OpenAI gym
• NumPy
• PyQT5
• PyTorch

Start by manually installing PyTorch. See the PyTorch website for installation instructions specific to your platform.

Then, install the minigrid Gym environment:

git clone https://github.com/maximecb/gym-minigrid.git
cd gym-minigrid
pip3 install -e .
cd ..

Then, clone this repository and install the other dependencies with pip3:

git clone https://github.com/maximecb/baby-ai-game.git
cd baby-ai-game
pip3 install -e .

If you are using conda, you can create a babyai environment with all the dependencies by running:

conda env create -f envoriment.yaml

Having done that, you can either add baby-ai-game and gym-minigrid in your $PYTHONPATH or install them in the development mode as suggested above.

The levels directory will contain all the code relevant to the generation of levels and missions. Essentially, this implements the test task for the Baby AI Game. This is an importable module which people can use on its own to perform experiments.

The agents directory will contain a default implementation of one or more agents to be evaluated on the baby AI game. This should also be importable as an independent module. Each agent will need to support methods to be provided teaching inputs using pointing and naming, as well as demonstrations.

In pytorch_rl, there is an implementation of the A2C, PPO and ACKTR reinforcement learning algorithms. This is a custom fork of this repository which has been adapted to work with the gym-minigrid environment. This RL implementation has issues and will hopefully be replaced by a better one soon. One important problem, for instance, is that it is not importable as a module.

The main.py script implements a template of a user interface for interactive human teaching. The version found in the master branch allows you to control the agent manually with the arrow keys, but it is not currently connected to any model or teaching code. Currently, most experiments are done offline, without a user interface.

To contribute to this project, you should first create your own fork, and remember to periodically sync changes from this repository. You can then create pull requests for modifications you have made. Your changes will be tested and reviewed before they are merged into this repository. If you are not familiar with forks and pull requests, I recommend doing a Google or YouTube search to find many useful tutorials on the issue. Knowing how to use git and GitHub effectively are valuable skills for any programmer.

If you have found a bug, or would like to request a change or improvement to the grid world environment or user interface, please open an issue on this repository. For bug reports, please paste complete error messages and describe your system configuration (are you running on Mac, Linux?).

To run the interactive UI application:

./main.py

The environment being run can be selected with the --env-name option, eg:

./main.py --env-name MiniGrid-Fetch-8x8-N3-v0

Basic reinforcement learning code is provided in the pytorch_rl subdirectory. You can perform training using the A2C algorithm with:

python3 pytorch_rl/main.py --env-name MiniGrid-Empty-6x6-v0 --no-vis --num-processes 48 --algo a2c

In order to Use the teacher environment with pytorch_rl, use the following command :

python3 pytorch_rl/main.py --env-name MultiRoom-Teacher --no-vis --num-processes 48 --algo a2c

Note: the pytorch_rl code is a custom fork of this repository, which was modified to work with this environment.

To see the available environments and their implementation, please have a look at the gym_minigrid repository.

If you connect to the lab machines by ssh-ing, make sure to use ssh -X in order to see the game window. This will work even for a chain of ssh connections, as long as you use ssh -X at all intermediate steps. If you use screen, set $DISPLAY variable manually inside each of your screen terminals. You can find the right value for $DISPLAY by detaching from you screen first (Ctrl+A+D) and then running echo $DISPLAY.

The code does not work in conda, install everything with pip install --user.

If you run into error messages relating to OpenAI gym or PyQT, it may be that the version of those libraries that you have installed is incompatible. You can try upgrading specific libraries with pip3, eg: pip3 install --upgrade gym. If the problem persists, please open an issue on this repository and paste a complete error message, along with some information about your platform (are you running Windows, Mac, Linux? Are you running this on a MILA machine?).

The Baby AI Game is a game in which an agent existing in a simulated world will be trained to complete task through reinforcement learning as well as interactions from one or more human teachers. These interactions will take the form of natural language, and possibly other feedback, such as human teachers manually giving rewards to the agent, or pointing towards specific objects in the game using the mouse.

Two of the main goals of the project are to explore ways in which deep learning can take inspiration from human learning (ie: how human babies learn), and to research AI learning with humans in the loop. In particular, language learning, as well as teaching agents to complete actions spanning many (eg: hundreds) of time steps, or macro-actions composed of multiple micro-actions, are still open research problems.

Some possible approaches to be explored in this project include meta-learning and curriculum learning, the use of intrinsic motivation (curiosity), and the use of pretraining to give agents a small core of built-in knowledge to allow them to learn from human agents. With respect to built-in knowledge, Yoshua Bengio believes that the ability for agents to understand pointing gestures in combination with language may be key.

You can find here a presentation of the project: Baby AI Summary

A work-in-progress review of related work can be found here