oliver reimer matthew crites brian jones. determine the winner of a nfl game between two teams. ...
TRANSCRIPT
NFL Game SimulatorOliver Reimer
Matthew CritesBrian Jones
Determine the winner of a NFL game between two teams.
How?◦ What aspects of a team are most important in a
game’s outcome?◦ How to represent match-up exploitations?
(i.e. a top-tier defensive team vs. a mediocre offensive ball club)
The Problem
Given 2 teams, teamH and teamA based upon the comparison between offensive and
defensive factors, O and H◦ OH , DH
◦ OA , DA
Each team’s “edge” in the matchup will be computed as
The team with the larger edge value is the predicted winner
The Solution
A
H
H
D
OE
H
A
A
D
OE
Given a set of n team normalized, offensive attributes, P= {0, 1, … n-1}
And a set of n team normalized, defensive attributes Q= {0, 1, … n-1}
And a set of n weights W = {0, 1, … n-1} The offensive factor is calculated
The defensive factor is calculated
The Factors
ii WPn
i
1
0
ii WQn
i
1
0
Statistics were taken as a per game average over the 16 games of the 2014 NFL season.
Offensive statistics were normalized
Defensive statistics were normalized
Normalization
avgleague
statteam
_
_
statteam
avgleague
_
_
Passing Yards per game Rushing Yards per game Points per game First Downs per game Fumbles (lost and recovered) Interceptions
Statistics Considered
What set of weights will give optimal results?
What team attributes are more important in the game outcome?
The nature of these questions support a genetic approach
Weights
Four step process◦ Determine fitness of the members of the
population◦ Generate offspring◦ Mutate◦ Update the population
Genetic
Start with a population of n randomly generated solutions◦ Weights had values ranging from 0 – 1 exclusive
Determine Fitness Generate Offspring from the most fit Mutate Update Population Repeat for m generations
Genetic
Four Solution Fitness categories ◦ Edge differences ascending◦ Edge differences descending
◦ Simulation ascending◦ Simulation descending
The four categories produce vastly different results
Fitness
|| AEEH
Comparison
PPG
FLPG
FDPG
PASS_YPG
INTPG
RUSH_YPG
PAPG
FRPG
FDAPG
PASS_YAPG
INTPG_D
RUSH_YAPG
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
SIMADESASCSIMD
Accuracy in predicting the 2014 NFL Season◦ Edge Difference Ascending
50.4 %◦ Edge Difference Descending
70.7 %◦ Simulation Ascending
70.7 %◦ Simulation Descending
48.8 %
Results
Most Important◦ Forcing Turnovers
Fumbles Interceptions
Least Important◦ Yardage Allowed
Rushing Passing
◦ First Downs Allowed
Results
Hot Streaks Past Match-up Experience Injuries Home-field/Weather advantage
Future Work
Questions?