Roussel Rahman, RPI Graduate Student

Roussel Rahman, RPI Graduate Student

Although performance averaged over groups of people tends to show a power law relationship between practice and performance, individual performance is more often categorized, not by smooth and continuous improvements, but by long periods of no improvements, periods in which performance even drops, and other periods in which performance shows sudden improvements. I refer to these three periods as “Plateaus, Dips, and Leaps (PDLs)” respectively, and my work focuses on developing methods by which such PDLs can be detected from records of task acquisition. In this work, I propose an information-theoretic approach to detect the PDLs in the game of Space Fortress using Kullback-Leibler Divergence or Relative Entropy. Preliminary analysis of the detected PDLs revealed several aspects of individual strategies and their shifts; for one example, practicing with part-task focus was found to be a general exploration technique recurrently used by the uninstructed players. Additionally, several plateaus in game-generated scores appear to be resulting from the limitations of scoring rules of the game and hide underlying improvement of players. Therefore, I suggest a few characteristics that appropriately sensitive performance metrics should possess, both in the game of Space Fortress and in general.

Share|