Hi everyone, i know next to nothing about AI so excuse me if i say something that doesn''t make sense. Here is my problem, i have a game where 100 teams can play against each other. Teams can only play one match home and one match away against an opponent so every team will make 198 matches by season. I want to be able to predict what will be the result of one match between 2 teams(eg. 45 - 78). Some important variables that i know about each team: I know that each team made X points in the season, x victories away and x victories home, x defeats away and x defeats home and finally x draws away and x draws home. I also know the total of home + away results if that maters. Another variables that might be important: I know the last 10 seasons game results from when TEAM A played with TEAM B. Lets assume TEAM A won 9 games and loose only one against TEAM B. However TEAM A, in this particular season only won 3 games against opponents and loose the other 48. TEAM B won 40 games and loose only 11. As you can see if i only take in account the history of matches played between TEAM A and TEAM B in the last 10 seasons the algorithm will probably predict that TEAM A will win, however this season TEAM A is been doing poorly so the algorithm should probably attribute the victory to TEAM B. This is what i want to do, predict the results between two teams. This should be done taking in account all the variables collected in the past and give priority to the one(s) who will have more impact. Am i the responsible for selecting which variables have more influence in matches or can by any mean a particular algorithm do that by itself? Which algorithms/techniques should i choose to tackle this problem? Thanks for your input.

Ultimately, as the analyst, you are responsible for such things. Some learning algorithms, like tree induction, will select variables on their own, from (and this is important) teh candidate variables you provide. If you provide useless variables, no modeling algorithm will do other than provide a useless model or collapse. If you provide enough crazy inputs, many modeling systems will produce useless models with confidence.

Assuming that one does not abuse the process, there are a number of relatively automatic processes which can serve as tools. Search online for things like "subset selection", "forward selection" or "feature selection"- this should be enough to get you started. As a start, see my article "Modeling Methodology 2: Model Input Selection", which was published in the Jan/Feb, 1998 issue of "PC AI" magazine. You can find it at:

http://will.dwinnell.com/will/willTechnicalPublications.html

-Predictor
http://will.dwinnell.com