This page hosts the agenda for each TA meeting on mondays.
- Introduction with TA
- Propose CURL paper
- CURL is an improvement on Pixel SAC. Would it be good to also implement the baseline? It is mostly swapping the first couple of CNN layers.
- Currently we have copied the utility classes and training structure from the original repo. Is this okay to do?
- Discuss further research options if implementation turns out to be too trivial.
- Discuss what repo to use as a basis for the reproduction.
- See week 3
- We are unsure how to exactly update the encoder from the SAC side, should we use actor, critic or both?
- Training for 100k timesteps will take +- 15 hours on cheetah run, others might be slower/faster. What is your opinion on this duration?