gym-fantasy-football-auction

Gym environment for a fantasy football auction.

A fantasy football auction involves a set of owners and a list of draftable players. Each owner has certain slots on their roster that they must fill with players who play in certain positions. Owners take turns nominating a player. When a player is nominated, every tick, owners can submit bids which are higher than the current bid. This continues until nobody submits a new bid for a full tick. Play proceeds until all owners' rosters have been filled.

For the purposes of this environment, we assume that every player has an agreed upon value (based upon projections, expert consensus, etc...). The winner of the auction is the player who ends with the highest total value, weighting bench slots lower than non-bench slots.

The reward for this task is simply the agent's final score divided by the score of the winning player.

Action Space

The only action an owner takes in an auction is to nominate a player (giving an initial bid) and to place a bid on a currently nominated player.

Each draftable player (regardless of what has happened in a game) has an index from 0 to number_of_draftable_players - 1.

Think of the action space as a 2-dimensional matrix. The row represents the player index. The column represents the bid on that player - one column for every integer from 0 to the maximum possible bid amount.

For example, the action of nominating player_id 23 with a starting bid of 30 dollars would be represented by the state at row 23 and column 30.

Similarly, assuming player 45 is currently nominated at 22 dollars, the action of raising the bid to 25 dollars would be represented by the state at row 45 and column 23.

Then, our action space is just one simple modification on top of this - we flatten the 2-d matrix into a list.

Observation Space

Let's say we have 200 draftable players, a starting money amount of 100, and 4 owners.

Our observation space is a stack of multiple length 200 layers (length = num_players).

We have these layers whose values change:

• owner (4 layers - one per owner, indicating who owns the player)
• bid value (4 layers one per owner - 0 if not nominated, otherwise, contains integer value representing the most recent bids from all other players)
• max bid (4 layers - one per owner, indicating the max bid this owner can make for any player)
• draftable - 4 layers - one binary layer per owner - set to 1 if the player is draftable by the owner - meaning they have space for it in the roster and the player is not already owned
• nomination - binary where all values are set to 1 on all players if it is nomination time for the owner whose perspective we are taking, otherwise 0.

We have the following fixed input layers, which are always the same:

• value - ECR value of each player, as an integer
• position - one binary layer per possible position - 11 layers total

So the whole state is a 2d array of size (4 + 4 + 4 + 4 + 1 + 1 + 11 = 29 x 200)