It's important to understand how my approach to my research is different than those of other researchers, i.e. what are my distinguishing characteristics as a researcher. If I'm going to succeed where others have failed, then maybe it's because of these differences in my approach. I want to have it written down exactly why I've chosen to do things differently, and what the disadvantages are. Maybe at some point I'd like to change my mind on any of these differences and work more like other researchers, and then it'll be useful to have them written down nicely so I could evaluate them better. Finally, when I talk to other people about my research, they'll be interested to know these differences, because it'll help them place me on the map of the MARL research world. Let's break down these differences to categories: ## Meta differences Let's start with differences that are _around_ the research rather than in the subject matter itself: - [[Communicate results to the public in a fun, down-to-earth way]]. - [[My research results should be accessible to the public on a nice web interface]] ## Differences regarding AI and humanity - [[I don't try to make the AI understand the human world, I try to get it to create its own world that's similar to the human world]]. - [[I shy from suggesting what kind of behavior would be good for human society]]. ## Subject matter differences - [[Be rigorous about the way that the environment changes the agents]], not only about how the agents change the environment. - [[Evaluate agents only in the context of their culture]]. - [[Intrinsic motivation makes the results less interesting]]. - [[Don't try to maximize cooperation before you understand it]]. - [[Win-win and communicate, then introduce a social dilemma]]. - [[Instead of treating the AI as a genius, we treat it as a rat in a maze]]. ## Minor differences - [[SSDs are great, but the concept can be misleading]]. - [[Autocurricula isn't a tool to solve a problem in MARL, it's a natural phenomenon we're reproducing]].