Research update: AAMAS and vacation

[Ram Rachum]

Hi everyone!

Retrospective on goals for last month

In last month's update I outlined a list of goals. Here's my update on these goals:

1. Present my research at AAMAS: ☑ Done

   I presented my research in the following formats:

   • 5 minute talk at the ALA workshop

   • Poster session at the ALA workshop

   • 15 minute talk at the COINE workshop

   • Poster session at the main conference

   I've gotten generally positive responses. One topic that came up multiple times is people asking what the applications are, and I said there are no immediate applications and explained how it fits into my research mission.

   There were around 10 people who approached me with questions at the ALA poster session. Some of them were very enthusiastic. Ann Nowé, who is very well-known and active at AAMAS, asked me many questions about the setup of the experiments.

   At the main conference poster session I had around 30 people approach me and ask questions.

   Me with the poster:

   

   
   

   I've also learned about lots of research work at the conference. Jérôme Botoko Ekila is doing work on emergent communication in RL agents and he seems to have a similar direction to mine.

2. Have fun in New Zealand and Japan: ☑ Done... almost

   I didn't do a lot in New Zealand. I barely got out of Auckland. I went hiking at Rangitoto with a group of people from AAMAS. I'm writing this now in Osaka, and Japan is so different that when I think back about my time in Auckland, it just feels like Israel.

   I spent 5 days in Tokyo which was amazing. I gave a talk about my research at the Tokyo AI Safety Benkyoukai, managed by Blaine William Rogers. I did a couple of standup shows and met a few friends I have there. I then went hiking for two days between Magome, Tsumago, Nagiso and Nojiri. A few days ago I reached Osaka for the final part of my vacation. I gave a talk at Kobe university about my research, which was a lot of fun. Now I have around a week more to hang out in Osaka before I come back. I'll probably go on day trips to Kyoto and Nara.

More stuff

Even though I'm on vacation, I've had some time (on flights and bullet trains) to continue working on refactoring the POLA implementation. There were a few Jax-related hurdles that Georg has helped me get over. I now have a better understanding of Jax.

As an example, here's a block of code from the original POLA implementation. It's 100 lines. (Probably would have been 90 lines without the unnecessary blank lines.) I've refactored it down to 33 lines of code in Polina.

One of the attractive properties of Jax is that the runs are deterministic. This is great for me because I made a golden run of POLA before I started refactoring, and on every step I make, I check that the results of the golden run remain identical. When they don't, I know I made a mistake somewhere.

My goals for this month

1. Finish refactoring POLA and get a better understanding of the algorithm.

That's it for now. See you next month!

Ram.