Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[TolyK]

Not looking at your role PM until N1 is like being a sleeper agent. That's not exactly fair...

By the way, on the off chance that I want to do machine learning on bay12 mafia games, have there been any tools to extract text from the forum posts? I know LurkerTracker probably uses something along those lines, I just don't want to write a tool to parse it myself if I don't have to.

[doll]

Putting other players (who aren't on your team, in particular) at a disadvantage is kind of the point of (the application of) theory in games like mafia.
I was, of course, wrong to agree with Leafsnail and Persus13. Lynchable teammates are not in any way desirable in mountainous as only townie-mafia matter and lynch-able mafia are by definition not townie - pressure makes townie players appear more townie and scummy players appear more scummy, therefore there is no greater or lesser reason to exert pressure on teammates. Scum's goal in mountainous can be summed up as 'have the game end with at least one mafia alive' which is not at all a numbers game and entirely 100% down to one's ability to lead, control and avoid the lynch. One's ability to do so is wholly independent of the strength of ones teammates where one's teammates are ignorant, and decreases alone as one's teammates become better informed because one relies then on the competence of one's teammates in not spilling impression towards ones role indelibly, that is, in the heart of hearts of the other and not in the mind thereof.
What's the strongest method available to the town for attempting to select a random lynch, anyway? I haven't thought of a methodology which doesn't open itself up to scum manipulation or require the loss of town members. Is there a gold standard in place?

[Leafsnail]

You can win on your own, but it's a bad bet. You're hoping that you, who are acting like a townie, will get lynched after all the other players who are also acting like townies. Really if you don't trust yourself to use the information of who your partner is effectively you can't have much faith in your ability to play the game.

You could base the lynch on a real-life event such as a lottery draw. You could also have the players encrypt numbers and add them up to decide the lynch, although encryption would be another thing you'd have to ban if players started using it.

[TolyK]

I could write out a game-theoretic tree of this situation, but we first need to agree on a set of rules.

Are all players unaware of their roles, or is only 1 player unaware? And then, does the unaware player play exactly like Town or slightly differently?
Also, how do lynches occur? (example: town has a weight of "t" to be lynched, mafia "m", unaware "u", and the probability of one getting lynched is their weight divided by total weight) (another example: each player has a "scumminess level" that is a random variable, distributed differently whether Town, Mafia, or Unaware, and the one with the highest scumminess value gets lynched).

Simplest would be just assuming Unaware players act like Town, and Mafia-acting players have a slightly larger "weight" of getting lynched.

[doll]

You can win on your own, but it's a bad bet. You're hoping that you, who are acting like a townie, will get lynched after all the other players who are also acting like townies. Really if you don't trust yourself to use the information of who your partner is effectively you can't have much faith in your ability to play the game.

You're hoping that you or anyone else on your team, who are acting like townies, will get lynched after all but (remaining scum) of the other players who are also acting like townies.
As I said, "persuasive position is more important than technical position re: information, and thus that an increased information base is very useful to scum seeking to influence other players" - and so you would look at your role PM. You would, however, use your information advantage as a tool to manipulate town players because you know that they know that they are unclear town and so you can address them appropriately. You should not use that information to avoid leading a lynch on a scummy teammate because only townie mafscum are valuable, and by definition of your ability to lead a lynch on them you are less lynchable than they are, so a compromise in your position is unacceptable.

You could base the lynch on a real-life event such as a lottery draw. You could also have the players encrypt numbers and add them up to decide the lynch, although encryption would be another thing you'd have to ban if players started using it.

Encryption with randomly generated initial numbers should work, so far as I can see.

[doll]

Probability Mapping
I'd like to request for some mechanical help in examining the outcomes predicted by the assumption that skill is purely a metric of ability to control the lynch (and not to determine the roles of other players) and what this says about: the balance of mountainous games, and, the optimal strategies for power roles in certain simple games.

Mountainous
Here we assume that: for P players, each player has a predetermined and publicly known level of skill from most skilled (S=1) to least (S=P), and that players are lynched in reverse order of skill, and that town players are night-killed in order of skill.
Using the notation of players as PX where P is either T(own) or S(cum) and X is the skill of that player, we could write out examples of setups and games respectively as:
SSTTT
S1 T2 T3 T4 S5
The tree of the game being:

Spoiler (click to show/hide)

S1 T2 T3 T4 S5
D1
S1 T2 T3 T4 S5
N1
S1 T2 T3 T4
D2
S1 T3 T4
SCUM WIN

For the setup, there are 20 possibilities. However, it is relatively simple to determine scum's chance of winning (90% SCUM WIN)

Spoiler (click to show/hide)

(T5) = 4x3/20
60% SCUM WIN
(!T5) = 2/5
       (S4) = 1/4
10% TOWN WIN
       (T4) = 3/4
30% SCUM WIN

However, the probability tree becomes far less human-parsable as the size and complexity of the game are increased. In these cases, machine consideration is probably warranted.
Cops
Let us make the following assumptions about cops:
Any cop who is lynched outs their reports
Outed reports on town have their skill changed to that of the cop (if cop unclear and better skill) or become cleared (cop is dead or clear)
Outed reports on scum cause the lynch pool to be reduced to the cop and their outed scum (worse of cop and scum are lynched)
Claimed cops are always night-killed outside of mylo

From these assumptions we can ask: should cops investigate good players or bad ones? When should cops out their reports?

[4maskwolf]

I'll write that program tomorrow, probably, since I don't have work, and let the machine mull over the variety of scenarios available.

[doll]

The strongest mindset in this game is that of lone scum.
Certainly, other players realized this before myself, and told me so in P25 or through their performance.
However, what may not have been seen is the ease by which scum may intuit the roles of players from their position not of knowledge, but of solitude.

Perfect play favors the defender. Furthermore, it is a general truth much seen that while it is easy to do things, it is not easy to have things not done. Ideal play in context therefore depends upon the positioning of imperfect play to enact a desired outcome, the outcome being not outside the strict confines of what can be accepted and the bulwark of the play of the individual backed by the silence of ones own unseen jabbering alone defends the point of contention.

There's no reason to lynch someone just because you think they're scum.

[Tiruin]

The strongest mindset in this game is that of lone scum.
Certainly, other players realized this before myself, and told me so in P25 or through their performance.

I feel like you've realized it a bit before you realized you realized it. But you have linked something that really improved what I've had in insight and used in the background for many a game.

When you are lone scum, all that can defeat you is yourself--and an inspect, barring other direct things (like a kill or redirection). How you take it, and others; this is markedly why I keep advising in the mafiachats I go to if I'm mafia, to play 'lone wolf' unless conditions bar working together being a better possibility--it works with a core principle of being town: The Day Game. To realize that nothing is over until you are lynched or killed, is a far deeper insight than how it initially feels--you can bus a scumbuddy for legitimate reasons, but not so to drive each other off the cliff. You can aim to lynch someone because you think they're scum; do not let this be left as an ungrown reason.

[doll]

So much scurrying around is done to try to give town information, but this is useless.
Why not simply not get killed? Why not dodge the investigation? These things are not difficult.

Also Jim's wrong it's not that hard to read the roles of every player in the game.
It's hard to do so without an empty quickchat, however. Note that he did identify cop early on.

[doll]

While looking for scum in a vanilla setup involves searching for power roles (informed players), this is less-so (though still) the case in more high-energy games like BYORs.

Indeed, I would posit that in a 'crazy' closed setup it should be easier to search for town than for scum. A superb example of this is in Paranormal 25 where quite a few players developed fantastically strong town-reads on myself and Tiruin really quite early in the game.

The thing which marks town players is that town players have information they want to pass on before they die.
This is absolutely vital and the very thing which scum players have no interest in. Scum don't want to advance the game, they don't want the game to advance after they die, and as Leafsnail & Co. would be keen to point out, scum do not ever intend to be totally 100% honest.
Outside of certain auto-lose situations, scum do not ever want to really fullclaim with all the information they have - even if they have claimed their whole role, they should not outline their plans or wants or desires, or the position of their team.

[TolyK]

So, I've recently been thinking about what kinds of actions there are in Mafia, and I think I've more or less gotten the broad categories down. If I missed something, please mention it. I've been brainstorming these with regards to the engine I'm writing.

• Info-State: Get information about the state of something (the game, a team, player roles, etc.)
• Info-Action: Get information about a target's actions during a period of time.
• Modify-State: Change a target's status (e.g. Alive or Can Vote), their members (i.e. moving from town team to cult, or stealing a player's role), or something about the game (i.e. skipping the next phase)
• Modify-Action: Change how actions are performed (i.e. roleblock, redirect) during a period of time.
• ??: I couldn't think of any (normal) actions that don't fit in the above.

Some examples:

• alignment-cop, role-cop
• watcher (who visited, what actions were used), tracker (visited who, used what ability)
• kill, role-(swap, steal, gift, add-action, votesteal), convert, phase-skip, ...
• roleblock, redirect, nexus, reflect
• (non-causal abilities, i.e. related to time travel, which I'm not going to try modeling :p)

A notable exclusion from the above list are passive (Auto) abilities, which are triggered by certain conditions, and things like protections (doctor/bodyguard) which I don't know how exactly to implement. On one hand, they could set a certain status ("protected") which actions would check before resolving. On the other, they could just modify any kill-actions going after the target. Personally, I'm leaning towards the second one, but I'd like to hear (read) any thoughts on the entire matter.

[notquitethere]

Because you might want to expand a mechanism to encompass other things later, its best to implement things as accurately as to how we think they work.

Protecting doesn't target a kill (the player doesn't even know if there's a kill or not), rather it imparts a temporary status on a target.

And it's not the only thing that does this: there can be abilities which give you one-shot, or permanent kill-immunity, and there are also other player statuses that impact both kills and non-kills, like some reflect powers.

[Mephansteras]

There are also counter-like statuses, such as the Devil's soul counter in Supernatural. This effects both the Devil and everyone who accepted the deal.

[TolyK]

Because you might want to expand a mechanism to encompass other things later, its best to implement things as accurately as to how we think they work.

That's the plan, yep.

Protecting doesn't target a kill (the player doesn't even know if there's a kill or not), rather it imparts a temporary status on a target.

Right, but protecting modifies a kill's behavior, and the "kill" action itself doesn't know what modifies it (while the "protect" action literally only saves from protects).

Technically, the "protect" action would make a temporary listener (that goes away when the phase changes) that, whenever a kill happens, checks if the target is the same as the protect target. If so, it modifies the "kill" action to be "failed" or "blocked" or smth (I'll have to think about terminology), so the kill doesn't go through.

And it's not the only thing that does this: there can be abilities which give you one-shot, or permanent kill-immunity, and there are also other player statuses that impact both kills and non-kills, like some reflect powers.

The one-shots work by just removing themselves when used up. The kill-immunity is basically just adding another (passive/automatic) ability.
They can also be configured to not just affect kills - all targeted actions of any type, for example.

There are also counter-like statuses, such as the Devil's soul counter in Supernatural. This effects both the Devil and everyone who accepted the deal.

Counter as in # / count? That's doable in the current framework.


Are there any other cases I missed other than the 4 I stated? I've looked down the Xylbot role list and haven't (yet) found any that don't fit. Are there any Bay12-specific ones that I forgot about?