For details about the evaluation criteria for the project, as well as the logistics, please see the course outline in D2L. Here we provide some supplementary materials, some sample topics for projects, and an FAQ which will be expanded throughout the semester as we receive student questions.
Project Proposal and Final Report Template
We will be using the IEEE Conference template in a double column format for the project proposal and final report. The font size should be 10 or 11. The page limit for the proposal is 2 pages including all pictures, tables, and references. The page limit for the final report is 6 pages, including all figures, tables, and references. The use of Latex for writing up these documents is recommended but optional.
Project Ideas
If you are not selecting the individual default project posted to D2L, you are expected to find a project that interests you or your group with the help of the TAs and your instructor. Since the project is the majority of your grade, you are expected to find a project that is considered "significant". Therefore, you must start early. Most groups will choose to replicating some of the prior work on reinforcement learning. For this you can look at the proceedings in NeurIPS, ICML, ICLR, and other conferences. You can also look at kaggle competitions/datasets, or perhaps other open projects on reinforcement learning online.
If you are unsure about what to choose for your project, one good idea is to look at OpenAI's Gym. There you have multiple environments that you can use for your projects. Make sure you also look at the third party environments there. Beside this, you can train reinforcement learning agents to learn to play a game (OpenAI Gym also has multiple games). Here are some examples (you can find more on your own):
If you like to read and reimplement an approach that has been published, make sure you have access to all the necessary components to replicate the results. Some potential papers are (you can find more on your own):
- Curiosity-driven Exploration by Self-supervised Prediction
- Fast reinforcement learning with generalized policy updates
- Meta-Learning through Hebbian Plasticity in Random Networks
- LeDeepChef Deep Reinforcement Learning Agent for Families of Text-Based Games
- Dream to Control: Learning Behaviors by Latent Imagination
- CURL: Contrastive Unsupervised Representations for Reinforcement Learning
- Combining Deep Reinforcement Learning and Search for Imperfect-Information Games for a group of size 1 to read, summarize and experiment with the code (to replicate the results huge compute is needed so just experimenting is enough)
FAQ
Will be populated soon.