The Great Vote Count Analysis (Pre-Discussion)

[Titus]

To echo and supplement Blair's point, VCA isn't a perfect algorithm but an imperfect heuristic. I am a big proponent of VCA over reading when I sub into games if the assumptions are met. That doesn't mean VCA is always flawless. If the wrong assumptions get entered, a wrong result pops up frequently.

[Isis]

sometimes after titus makes a really sharp point I imagine her moonwalking away immediately afterwards

[Zantetsu]

I don't doubt that it's possible to build an algorithm to detect posting patterns that indicate scum, I'm surprised someone would do it though because it would take all of the fun out of a game that exists solely for the purpose of having fun.

What is the functional difference between "using an algorithm to detect posting patterns that indicate scum" and "using [your brain] to detect posting patterns that indicate scum?"

Which part of the process would an algorithm "take all the fun out of?" Would it still be fun if you used the algorithm to learn what patterns to look for, but applied those learned principles in live games without the algorithm? If not, why not?

Lots of players apply "tells" that they learned from other players, rather than discovering entirely on their own. Is this functionally different from learning a "tell" (which is what it would functionally be) from an algorithm? Is the difference the fear that it might be more accurate?

Yes it is different, very different. Kind of like playing a sport against someone who uses nonstandard equipment to give themselves a competitive advantage.

Anyway, like I said already, if you're just developing algorithms, you don't have to play in ongoing games. You can just turn them loose on completed games and you won't disturb any games in progress.

[Zantetsu]

Although I suppose that if part of your algorithm uses results of its own votes to further refine reads, then you'd have to do it in live games.

[Zantetsu]

My friend's wife used to play an online version of the game Boggle. One day my friend and I wrote a program to solve the Boggle board for words and we fed her answers while she was playing a game. It was fun for about 30 seconds to watch people being astounded at how quickly she could come up with words. Then it got boring because we weren't really accomplishing anything. We weren't even playing the game, we were just watching our program spit out words. I imagine using a program to find scum for you would feel like that.

[Blair]

You're comparing it to your Boggle program (which you say sucked all the fun out of the game for you) but you also keep suggesting it should only be applied to completed games (which suggests you think a tool like this would suck the fun out of the game

for everyone playing).

I can definitely see how the Boggle program could ruin that game for everyone, not just the player using it, because you would pretty much always win. A similar program for Mafia would not be an auto-win for many reasons, however.

The main one being: It's all subjective. There are no 100% accurate tells. Even if an algorithm could somehow track fake claims, lies, counterwagons, and

scum claims,

it would still yield false positives all the time.

The benefit of a tool like this isn't some sort of magical, 100% accurate scum hunting voodoo - the benefit is in automating some of the more tedious parts of the game (vote count analysis, detecting time/day posting patterns, etc.)

[Blair]

A side-benefit would be finally being able to rigorously test some of our long-held assumptions ("I used to think scum did X more often than town, but after running 2,000 games through this program, I discovered there was no correlation").

[Blair]

I wouldn't compare it to sports teams using unfair equipment, I'd compare it to statistical analysis being introduced to baseball (it was). Did it ruin baseball? (It didn't)

[Zantetsu]

I wouldn't compare it to sports teams using unfair equipment, I'd compare it to statistical analysis being introduced to baseball (it was). Did it ruin baseball? (It didn't)

Hm this is getting pretty far out there but I wonder if players being told in-game where to position themselves for the best chance of making a play against each hitter would take some of the fun out of it for both teams. I'm talking about amateur sports being played for fun here, not pro ball where who cares how much fun they have (not even the players, I'll bet, except when the fun doesn't impact their paycheck).

If the players all had smart phones in their pocket and on every play they were consulting the phone to be told what to expect and where to position themselves for that play ... which I think is a better analogy here ... I can imagine not wanting to bother playing if I felt like the other team was more likely to win because their smart phones were better.

And speaking for myself, I don't like the idea of playing against someone using anything beyond their own note-taking to facilitate their mafia play. I cannot know who if anyone is doing it, nor can I know who if anyone is communicating off-site for personal advantage, but in any case, to the extent that I would be aware of it, I would not like it. But that's just one insignificant person's opinion.

I do like the idea of learning what statistical analysis shows about mafia games. Studying up ahead of time and being well informed seems fine, and interesting, to me.

[yessiree]

I finally got around to writing a program that makes graphical representations of votes. I'm hoping this will help get things moving forward for the great vote count analysis.

Anyway, the games I scrapped are from the archive here
viewtopic.php?f=53&t=29549

The repository
https://github.com/blookvoodow/VCV

more information can be found in the README file

You can easily generate these graphs yourself as well without any programming skills, you just need a little bit of knowledge running NodeJS and Python scripts

Open to suggestions/comments on how to present the data better

Here are 5 examples

Spoiler: graphs

[Psyche]

hmm
the semester is over today
so maybe it's good timing

[gobbledygook]

I finally got around to writing a program that makes graphical representations of votes. I'm hoping this will help get things moving forward for the great vote count analysis.

Anyway, the games I scrapped are from the archive here
viewtopic.php?f=53&t=29549

The repository
https://github.com/blookvoodow/VCV

more information can be found in the README file

You can easily generate these graphs yourself as well without any programming skills, you just need a little bit of knowledge running NodeJS and Python scripts

Open to suggestions/comments on how to present the data better

Here are 5 examples

Spoiler: graphs

What do you think these images tell us about that game in particular?

[Psyche]

yknow i asked that question before and i only pretended to understand the answer

[yessiree]

one graph is one game, you can kinda get a holistic overview of what happened in that game

I guess I should post the README here for ease of access

Node

Each node in the network represents a player slot in the game. Each node includes all the replacements to a particular slot.

Edge

Each edge in the network represents a vote cast from one node to another. For simplicity, edges are undirected, meaning if two players cross-vote one another, only one edge will exist between them. However, edges are more complicated than nodes since weights are attached to them.

Weight

As the network is weighted, the weight of an edge represents the severity of conflicts between two nodes. Since an edge presents a vote, and votes are used for pressuring or lynching a player, it can be interpreted that the higher the weight, the higher the degree of conflict.

Weights are calculated based on duration of the vote. The reasoning is that votes generate pressure and may or may not start wagons. Therefore, the longer the vote is cast, the greater the pressure and the degree of conflict between the two players. Of course, this may not always be the case, but our assumption is the voters will be forced to defend their votes when confronted by other players, and they will either retract their votes or keep their votes as a result.

Weights are measured in milliseconds and normalized by 1 hour.

the labels on the nodes

L = lynched
K = killed

and the they are followed by the time of the incident, so
L d 1 = lynched day 1
K n 1 = killed night 1
endgamed, survived and won should be self-explanatory

a solid edge can be interpreted as the two nodes having a lot of conflict and interactions
a dotted edge means the two slots barely voted each other

[skitter30]

oooooh this is interesting, i'm going to take a closer look over the weekend!

which games were these 5?

also, why did you choose to assign weights in {length of irl times that has passed}? my first instinct for the weights would have been maybe how many posts the vote was on the person

[yessiree]

Spoiler: 1

in this one, scum got steamrolled.
you can see two "clusters" by looking at nodes whose edges are solid lines. one includes the mafia that got lynched d1, two townies were on him consistently, and another townie was pressuring one of the two, but he somehow got nightkilled
the scum that got lynched d2 didn't really focus on anyone, never placed his vote on any slot for a long period of time

you kinda look at it like this

pedit: I believe the games are in the same order as the archive posted by implosion here
viewtopic.php?p=11775190#p11775190

[yessiree]

actually no, scratch that, these got mixed up when I uploaded them to imgur

mini normal 2081
2082
2085
2076
2080

[yessiree]

also, why did you choose to assign weights in {length of irl times that has passed}? my first instinct for the weights would have been maybe how many posts the vote was on the person

it was easier to grab the timestamps of the posts and compute the difference, but post count difference is also doable I suppose

[skitter30]

fair enough! was just wondering

[Psyche]

I can't reboot this weekend but I'll be putting this back on my radar

[gobbledygook]

Spoiler: 1

in this one, scum got steamrolled.
you can see two "clusters" by looking at nodes whose edges are solid lines. one includes the mafia that got lynched d1, two townies were on him consistently, and another townie was pressuring one of the two, but he somehow got nightkilled
the scum that got lynched d2 didn't really focus on anyone, never placed his vote on any slot for a long period of time

you kinda look at it like this

pedit: I believe the games are in the same order as the archive posted by implosion here
viewtopic.php?p=11775190#p11775190

Thank you for this!
I am curious to see how the length of posts would change the representations.

I can’t immediately discern anything “universal” from the graphical representation, but the sample size is probably too small for that.

[yessiree]

graphs can be easily generated. I just didn't wanna do it since I wanted to get some feedback on the samples first

im gonna change time delta to post count delta and see what happens

[yessiree]

Changelog:

• changed weights for a single vote to be total post count instead of total time
  
• Added game # to title
  
• split weights into 5 categories, 800+, 500+, 150+, 30+, and 30-
  
• Misc style changes

Spoiler: actually deliver on promises Pog?

[gobbledygook]

Thank you for doing that yessiree. I don’t think the graphical representations changed too much, though.

[petapan]

psyche did you really spend 8 goddamn years on this project