ACDL 2019 Projects on BabyAI
Maxime Chevalier-Boisvert & Giuseppe Nicosia
“I share with many others a dream of learning machines which learn like
babies do, with the help of humans, about how their environment works,
acquiring gradually more complex skills, understanding, and language.
Motivated by this vision, we have developed the BabyAI platform. If you
want to do research on instruction following and/or language grounding,
consider using it: 10^19 synthetic instructions, 19 levels of varying
difficulty. Work done by @MILAMontreal with the help of @Element_AI.”
Yoshua Bengio
For students who must pass the ACDL 2019 Course with an exam/project in order to obtain the ECTS points, this year ACDL 2019 has the following project on BabyAI Platform.
Obviously all the participants are invited to work on the project on BabyAI.
BabyAI is a platform for simulating language learning with a human in the loop. This is an ongoing research project based at Mila.
https://github.com/mila-iqia/babyai
BabyAI: A Platform to Study the Sample Efficiency of Grounded Language Learning
https://openreview.net/forum?id=rJeXCo0cYX
BabyAI is an open-ended grounded language acquisition effort at Mila. The current BabyAI platform was designed to study data-efficiency of existing methods under the assumption that a human provides all teaching signals (i.e. demonstrations, rewards, etc.).
If you have questions on the code or otherwise you can write to Maxime Chevalier-Boisvert (chevalma@iro.umontreal.ca) and Dima Bahdanau (dimabgv@gmail.com).
It is strongly recommended to approach this project with a group!
At the end of the course, the team members will present the results to the lecturers and the directors.
It follows the list of projects on BabyAI:
- To perform experiments on curriculum learning and manually designing curriculums (or automatically sampling which levels to train on based on learning progress). This paper may be inspirational (https://arxiv.org/abs/1707.00183).
- To play with unsupervised exploration for representation learning, having agents learn things on their own in the environment without supervision (e.g. how to navigate from room to room). It seems clear that much of what an agent has to learn is not task-specific, but rather understanding the basic mechanics of how it can interact with the world.
- Projects on model-based reinforcement learning for BabyAI: learning a predictive model and using it for planning which action to take, or using optimization algorithms (or stochastic search) to evolve agent models that learn faster.
- Hyper-parameter tuning.
- Automatic agent architecture search. https://arxiv.org/abs/1808.05377
- Inverse reinforcement learning.