Gunslinger

A solution for a puzzle game that uses a variety of learning techniques, by Ryder Moody and Enrique Cruz.

The Rules

It is the wild west, and you find yourself among a circle of N outlaws all bearing grudges and loyalties from past encounters. In particular, each outlaw has a list of exactly f friends and a list of e enemies (the two lists are disjoint), where e + f < N. Each outlaw has a different list. The friend relationship is symmetric: if A is a friend of B then B is a friend of A. However, if A is an enemy of B, B is not necessarily an enemy of A; enmity could have arisen secretly. Further, a friend of a friend could well be an enemy!

There is a showdown, each of you ready to draw your guns at the slightest provocation. Each gunslinger chooses another gunslinger and fires a shot; alternatively, a gunslinger can elect not to shoot. All shots happen at once, with the bullets in flight simultaneously. Each gunslinger has a shield that can protect against one shot, but not against more than one. A player is eliminated if, on a single turn, she is shot by more than one gunslinger. Otherwise the player remains in the game for the next round.

Players have complete information about who tried to shoot whom on previous rounds. From this information, they can try to infer friendship and enmity relationships. Players do not have direct access to other players' relationships, and do not even know which group's code corresponds to which player.

The showdown ends when nobody has died for ten consecutive rounds. This might happen if all remaining players consistently choose not to shoot. It will also happen if all but one player consistently choose not to shoot, since it takes two bullets to eliminate a player.

At the end of the showdown, players (including eliminated players) score a point for each of the following:

Surviving at the end.
Having a friend that survived.
Having an enemy that was eliminated.

So the maximum score on a round is e+f+1.

For the tournament, we will run many instances of the game with the same collection of players, but with new randomized lists for each showdown. We'll try different settings for e, f, and N and see how they influence strategy.

Running

$ cd [this dir]
$ javac gunslinger.sim.Gunslinger
$ java gunslinger.sim.Gunslinger gunslinger/players.list [options given in
Gunslinger README]

Changelog

v0.0.1

Added initial given player and sim
Copied the player into a new base player class in gunslinger/g1/Player.java to work from

v0.0.2

If you have enemies, shoot a random one. If you don't have any enemies, shoot a non friend. If you dont have any non friends, don't shoot.
Simplest actual strategy, with 1 g1 and 8 dumb players, the single g1 player wins about 50% of the games.

v0.1.0

First real take on strategy: use a feature vector with hard coded weights.
Features in Vector: "enemy", "friend", "foe" (has shot me), "friend's foe", and "no relationship".
Feature vector starts as best guess for definitions of players, updates over time. Player with highest score is shot.
Weights picked with common sense. Will use learning techniques to optimize these weights in future.

v0.1.1

Add new feature to tell if somebody who shot your friend actually killed them, meaning we should not target this person since there is no point
Add initial relationship tracking (not yet in use)
Add choice to shoot or not based on being provoked: don't shot unless somebody shoots you or your friend

rymo4 / gunslinger