1

I am trying to train a basic Neural Network to predict Football final scores based on:

i) Time in the match

ii) Current Score

iii) Parameters representing strength of home and away team.

In order to generate my training set, I sampled a number of games at regular times throughout the match (i.e. every 5 minutes).

When I trained my network based on these samples, the performance wasn't particularly good (for one it did not learn that at the end of the game the score never changes).

I believe one of the problems is that the training samples are very highly correlated as they are only separated in time by 5 minute intervals. What would be a good tactic to train the network given these highly correlated training samples?

rwolst
  • 612
  • 6
  • 12
  • What is your desired result from the nn? the score? Because one thing that may help the network is to try to estimate the variation of the score based on the current score and not the exact score. I mean, if current score is 1-0, than a 0,0 is impossible, but if you allow the network to determine the exact final score than it is still viable in the nn perspective – ASantosRibeiro Nov 07 '14 at 16:00
  • My final output layer is a softmax, so I have 36 outputs representing the scores: 0-0, 0-1, 0-2, ..., 0-over5, 1-0, 1-1, ... and the probabilities assigned to each. – rwolst Nov 07 '14 at 16:03
  • that just seem too many output nodes, which will make the training harder. why not a linear layer and 2 nodes? and the desired result just final_score-curren_score ? remember that large networks usually require large amounts of data to train efficiently – ASantosRibeiro Nov 08 '14 at 17:29
  • Right, but my goal is to estimate the pdf of the score, so I need the relatively large number of output nodes. I can't think of another way to do it with less output nodes. – rwolst Nov 08 '14 at 17:39

0 Answers0