Suppose we have $n$ agents with hidden "ratings" $r_1, r_2, \ldots, r_n$, and player $A$ wins a match against player $B$ with probability $1 / (1 + \exp (r_B - r_A))$.
We can use the outcomes of past matches to select who plays in the next match.
How should we schedule games in order to most quickly determine the best player? How does this change if we can schedule multiple games at once?
Deliverables
A tool which takes in a list of players, as well as a function which reports the outcome of a game between them, and returns the player with the highest estimated rating.
In the parallel case, the provided function may be a Ray task.
Description
Suppose we have $n$ agents with hidden "ratings" $r_1, r_2, \ldots, r_n$, and player $A$ wins a match against player $B$ with probability $1 / (1 + \exp (r_B - r_A))$.
We can use the outcomes of past matches to select who plays in the next match.
How should we schedule games in order to most quickly determine the best player? How does this change if we can schedule multiple games at once?
Deliverables
A tool which takes in a list of players, as well as a function which reports the outcome of a game between them, and returns the player with the highest estimated rating.
In the parallel case, the provided function may be a Ray task.