predicting sport results - lunds tekniska...
TRANSCRIPT
-
Predicting Sport ResultsBY HANNES JOHANSSON & JAMES HALL
-
Our Project Idea
-
Available Data
-
Artificial Neural Networks• Inspired by biology • Human brains have
~85,000,000,000 neurons
-
Neural Net Model
O1I2
I3
I4
O2
I1
Input Layer Hidden Layer Output Layer
-
Gradient Descent
-
Neural Net Model
O1I2
I3
I4
O2
I1
Input Layer Hidden Layer Output Layer
-
Trained Neural Net
O1I2
I3
I4
O2
I1
Input Layer Hidden Layer Output Layer
-
Programming Environment• Neuroph framework • MySQL database • Java 1.8 • 15 classes across 3 packages
-
What We’re Predicting
• Parallel neural nets • Number of goals • Outputs are combined
-
Our Neural Nets
Neural Net Goals??
?
-
Evaluation
• Number of incorrect goals • Exact predictions • Winner/loser • Threshold
Inconclusive
Team A Team B
-
Our Results
• Accuracy goes from 55% to 60%
• Inconclusive goes from 25% to 80%
• Last-Game-Difference input
%
0
25
50
75
100
Threshold0.5 1.0 2.0
Max Accuracy Inconclusive
-
Possible Improvements
• More inputs • More data • Multiple passes • Other learning methods