Token Engineering Commons Gitcoin Modelling Co-Lab, 19 Feb 2021

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Gitcoin Modelling Co-Lab #5.5 Research Topics

Description

Join @gitcoin, @block_science, @tokengineering & QV/QF academics and enthusiasts to discuss the collusion hunting cadCAD model of the Gitcoin Grants ecosystem.

Research plan: https://hackmd.io/QCCJWZE0Ru27X_GRk6UKjQ
Repo: https://github.com/gitcoinco/gitcoin_cadcad_model
Slides: https://docs.google.com/presentation/d/1DehR5ANrRKMMawXNW-e_SvHEi1b0yyU1w401t93PUdg/edit?usp=drivesdk

More information about this research initiative:
https://www.notion.so/blockscience/Gitcoin-Modelling-Co-Lab-TL-DR-Public-4e115525dbc0430289bb78d60ad79a04

Learn more about the Token Engineering Academy here:
https://tokenengineeringcommunity.github.io/website/docs/academy-welcome/

Read more about Block Science: https://block.science/

Learn more about Token Engineering: https://tokenengineeringcommunity.github.io/website/

More about the Token Engineering Commons: https://medium.com/token-engineering-commons/kicking-off-the-token-engineering-commons-be6a253cba81

cadCAD is an open-source Python package that assists in the processes of designing, testing and validating complex systems through simulation.

Discover more at https://cadcad.org/
Join the cadCAD Community Discord: https://discord.gg/hs4dAN9
Telegram channel: t.me/cadCAD_org
Twitter: https://twitter.com/cadcad_org
Get up to speed on our forum: https://community.cadcad.org/

A

Think um roger like this is super important for all of us to learn to do so. I would probably agree with octopus that it's worth doing it even if we have to move slow.

B

Sure I'm just I was thinking I'll, be the guy going. um What does that mean quite a bit? So that's for sure.

C

Not sure if we have some minutes left to introduce some more questions in case sean and matt you'd like to share some of your thoughts, or maybe I don't know if you have already nailed your questions.

D

Yeah sorry is late, guys, um sports stuff. uh I have no idea what my question is gonna, be I'm still still very much feel like I'm trying to get up to speed, um so I'm definitely interested in any. Like math working group, I feel like uh very similar to roger in the same boat, where, like I come from a formal, math background, but like I haven't done in a long time, um so I'm really feel like I'm just trying to wrap my mind around the problem and understanding the terms that everybody else is talking about.

D

um I threw some questions out there. Some things I was interested in, but I mean this is like getting up to speed here- is like not that trivial, um so yeah, I'm still really I'm. The fact that you guys are talking about this is encouraging.

A

I'll put that as an extra plus one to the focused on onboarding and legibility, and making this work overall more accessible without again sacrificing the technical so like helping people get from join to I understand this seems like the most important thing that we could do.

D

Yeah for sure I feel like that yeah there's just a lot to sort of like I feel like when we listen to danilo like in his sessions, like you know, he's been in this for seems like a few years and he's deep in like graph theory and like a lot of things that are like you can sort of pull up to speed like relatively quickly. They know what's going on, but to be able to contribute, I think is another level um yeah.

D

So that's kind of what I think at least like from, like I'm a total newcomer. No octopus is like a total newcomer. So for guys like us, it's like that's very tough, even though you know we're all relatively technical.

A

Yeah, I guess the point I'm trying to make is that these sessions exist to identify and close that gap right like you know, if we were just trying to go full speed ahead, I mean honestly, we have our own internal sessions for our project. We are doing a project me and danilo and jaja, and some of the members of the block science team and we're going to continue to run sort of full speed ahead on that, based on the needs of the git coin platform and like our relationship with them.

A

But these sessions with the token engineering community are as much about transferring knowledge and like learning to communicate and collaborating as anything else so sort of just sort of understand that it's not a problem that we need to slow down and work through stuff.

A

It just needs to be identified so that we can accomplish this goal of making those fields more accessible, making it easier for people to understand how to practicalize them, because I think one of the reasons people haven't used some of these methods for some time is, they haven't been given a clearer path between their formal mathematical skills or education and the practical problems they face on a day-to-day basis, and if we can help fill that in then I think we can sort of unlock those skills too.

A

A

B

Quick quick question: yj anderson in particular. What uh what kind of models were you looking to use? Because I recall you mentioning reinforcement.

B

D

B

I'm definitely discussing more on that too. uh I just don't see the rl part yet, but I'd love to discuss with you more on that on the side.

A

The reward function would generally in this setting, be the actual matched funds, or at least the efficiency sort of the match funds per fund. The challenge that you get into, though, is um if you're modeling a reinforcement learner against that you're actually effectively modeling an attacker. So what's interesting is you would be creating taking real historical scenarios or synthesized artificial historical scenarios with counterfactual scenarios using some of the data from the previous ones, the real ones and then you'd be running repeated simulations and you'd.

A

Let your attacker play and they would play with a strategy that was learned through the reinforcement, learning algorithm and what you would really be doing is is training an attacker.

A

It's an interesting idea, though I am not a hundred percent sure how you would use it to detect collusion it's more effective for um training a colluder, so um I mean I guess the open question would be like. If you could train a colluder, you could measure you could use it to sort of find optimal strategies quickly and then thus measure optimality gap more efficiently.

A

So there's some cool stuff that could come from that, but my instinct listening to what you're saying is that by using real historical scenarios and constructed counterfactual scenarios, you could actually develop a reinforcement learning agent. It would just be an attacker.

A

A

Yeah, I agree with that. It makes sense. I think it's pretty cool. I think the trick, though, is that, if you're looking into reinforcement, learning most of the meat of it comes from the the game that you create and the measure of of you know the utility function and then, after that, the reinforcement learning at this point has been reduced to use these packages and tools and proof the answer pops out but like what is the answer, you got the answer to and that's actually baked into how you defined the game.

A

They were playing and the scenarios so to speak and the utility function they had, but thinking about it in terms of the cad cad models and the um uh the actual outcomes in in the quadratic funding, I would argue, it's probably feasible using some of the research infrastructure that we've already built.

B

Yeah I'm down to sit down for a day and try to crack down on the mdp like just model as an mdp and see where it goes from there.

D

Yeah, I definitely am interested in that topic too. I have like background in that uh type of stuff sort of like a attacking models and then defending against them, and I think pretty well described exactly like how the research would go, but I still think there's like some interesting stuff. You could.

E

Learn very cool awesome. Well, uh yeah. I think there's there's a lot a lot to explore here. uh So glad glad we had this uh opportunity to explore all of the exciting initiatives coming out of the token engineering academy, research group and it seems like we will need some uh kind of spin out working groups.

E

If it's you know a bunch of uh academy participants collaborating on specific models- or you know these spin-out math working group um yeah happy to to mime and of course the sessions we have already booked are, of course available, um but there's definite um opportunity for async work on on some of these initiatives as well, so um yeah we can keep the discord channel as a place to touch base I'll drop in the notes and the recording from this call and yeah happy to discuss further async and we will be having these open hacking and open academic.

E

Oh sorry, the open hacking session is next week as well. uh The next uh one of these sessions will be two weeks from now and.

A

E

A

E

A

We should reserve, as a kind of you know, office hours for the the research planning, and you know. Obviously, people can reach out in advance if they have agenda items they want to hit. We talked a bit about the you know: convex, optimization, linear, algebra, algebraic graph theory stuff. We talked about the reinforcement learning stuff.

A

I think both of those are good topics, for you know, um sort of research planning and discussing of methods and sharing knowledge and even like you know in some bits, maybe even pulling in some materials and talking about it in more detail to help people get on closer to the same level in those topics. um But if I, if I had to sort of recommend, I think we should use these sessions sort of like office hours with a an agenda set like loosely in advance, similar to what we've done for like the one hive sessions.

A

Where you know the goal is to have a sort of community driven. But then we have the sort of daniello myself and you know, maybe others, depending on the topic coming in to provide you know like support, answer, questions or recommendations, but still keeping it largely driven by what the community wants to work on.

E

Yeah, absolutely and.

A

E

Actually, a really good point: there are some other kind of open collaboration, modeling sessions going on um one with the one hive community luna swarm. There may be some interest of people we're talking about like upscaling and kind of, like you know, learning the language of all of these multiple intersecting discipline.

E

So definitely these skills can be applied in multiple different areas. There are some other open community explorations um involving cad and kind of this network network science and and um yeah a lot of the same concepts so happy to invite anyone here who who is interested in this stuff to participate in any of these other open collaborations.

E

um They can be found on the token engineering calendar or the cad cad calendar, and, if you're really interested, I can, I can shoot you a link or it's all in the um the notion document. That's of the the link repo for uh for everything going on in this initiative so share some of those in the channel again.

E

uh If anyone needs anything in particular, please feel free to reach out and dm me, and I can hook you up some links to get you more information, um but I think that's all for today, unless there are any other final questions or closing thoughts um good to let everyone else head off to their next call.

B

uh Jeff really quick, so I arrived a bit late in the beginning. uh Where can I find that document? So I can add just some of my notes on my questions.

E

I will drop it in the channel right now.

B

Thank you very much.

C

Thanks guys, bye-bye cool, thank you so much for today.

C