It is an interesting idea. What neural net libs are you using?
I would imagine, that an evolutionary algorithm might be most suitable for this type of thing...
Use every two games as a new generation with the team (or teams, if you are tracking multiple ones) in question represented as a singular entity in the model.
Ex: Broncos vs Cowboys & Broncos vs Chargers; each occurrence of 'Broncos' is considered a singular entity.
Each Broncos entity will receive a breeding fitness magnitude based on (just off the top of my head):
- win or loss
- win/lose gap
- if desired team wins with large gap, higher fitness
- if desired team wins with smaller gap, lesser fitness (but still above 1)
- if desired team loses with large gap, very low fitness
- if desired team loses with smaller gap, low fitness (but still below 0)
- time to win or loss
- if shorter the win, the higher the fitness
- if longer the win, lower the fitness (but still above 1)
- if shorter the loss, very low fitness
- if longer the loss, low fitness (but still below 0)
The algorithms for calculating who breeds with who could get pretty complicated, so, I am not going to attempt an example; but the general idea being that those multi-faceted fitness levels provide for multi-dimensional breeding selections.
This kind of system would obviously work best with large data sets. So, historical data with present time (or even real time data too, to see live breeding) data would provide a more mature system to breed your teams with than simply present time data.
You could breed multiple teams and use a simple script to analyze their breeding fitness amongst each generation to produce results of which team is fitter to win in a match.
Would be an interesting experiment...