Canabalt is a simple but addictive Flash game where the goal is to keep your character alive for as long as possible as he runs parkour-style from rooftop to rooftop using just one action: jump. The iPhone version of the game tweets your score each time you play (or is it only when you beat your highest score? It's not clear), so a simple Twitter search yields some interesting data about the distribution of scores. Princeton graduate student John Myles White has collected the data, and used the ggplot2 package in R to plot the scores as a sample distribution:
Is it possible that score over 60,000 is legit? It's plausible that someone could fake a high score by posting to Twitter directly (mimicking the game's automatic tweet). Perhaps Extreme Value Theory could offer some evidence. Anyone up to the challenge?
John Myles White: Canabalt
The scores are like a mix of distributions - many poor players, combined with decreasing numbers of better, more experienced players; there's also dependence (in that players tend to improve with time).
IIRC, usually EVT is based on assumptions of independence and sampling from a common distribution, neither of which would hold true here.
Posted by: GB | November 19, 2009 at 21:29