Interrogating Feature Learning Models to Discover Insights Into the Development of Human Expertise in a Real‐Time, Dynamic Decision‐Making Task

Abstract

Tetris provides a difficult, dynamic task environment within which some people are novices and others, after years of work and practice, become extreme experts. Here we study two core skills; namely, (a) choosing the goal or objective function that will maximize performance and (b)a feature-based analysis of the current game board to determine where to place the currently falling zoid (i.e., Tetris piece) so as to maximize the goal. In Study 1, we build cross-entropy reinforcement learning (CERL) models (Szita & Lorincz, 2006) to determine whether different goals result in different feature weights. Two of these optimization strategies quickly rise to performance plateaus, whereas two others continue toward higher but more jagged (i.e., variable) heights. In Study 2, we compare the zoid placement decisions made by our best CERL models with those made by 67 human players. Across 370,131 human game episodes, two CERL models picked the same zoid placements as our lowest scoring human for 43% of the placements and as our three best scoring experts for 65% of the placements. Our findings suggest that people focus on maximizing points, not number of lines cleared or number of levels reached. They also show that goal choice influences the choice of zoid placements for CERLs and suggest that the same is true of humans. Tetris has a repetitive task structure that makes Tetris more tractable and more like a traditional experimental psychology paradigm than many more complex games or tasks. Hence, although complex, Tetris is not overwhelmingly complex and presents a right-sized challenge to cognitive theories, especially those of integrated cognitive systems.

Game‐XP: Action Games as Experimental Paradigms for Cognitive Science

“…For Game‐XP, the more complex the task, the more subtle the possible differences in the methods discovered and deployed by different players may be. For example, Sibert, Gray, and Lindstedt () compare human data to machine learning classifers that were trained on one of four different “objective functions”—that is, different methods for achieving the goal of gaining the most points possible. Some of these classifiers optimized the number of lines cleared while others optimized points per zoid (a “zoid” is a Tetris piece).…”

“…Only time plus a lot of hard work by many inspired researchers will tell. With the exception of the work by Boot, Sumner, Towne, Rodriguez, and Anders Ericsson (), none of the authors have published more than one or two journal or journal‐quality papers on their paradigm. (As a lower limit, let me confess that the paper published by my lab [Sibert et al., ] is our first on Tetris. ) Unlike chess, action games are not one paradigm but many. Like chess, action games present a natural paradigm in which to study differences in player expertise.…”

Allen Newell's Program of Research: The Video‐Game Test

“…() analyzed the behavior of a single participant. Sibert, Gray, and Lindstedt () apply optimization techniques from machine learning to identify the best combination of feature weights for selecting actions in Tetris. Finally, Reeves, Greiffenhagen, and Laurier () use ethnomethodology and conversation analysis to describe game playing from different viewpoints.…”

“…Sibert et al. () use an optimization algorithm to find the best weights for board features given four different goals; this provides important information for understanding strategies used by players. Finally, Boot et al.…”

“…() and Sibert et al. (), all theories are expressed verbally, where “too much is left unspecified and unconstrained” (Newell, , p. 301). Schrodt et al.…”