The Tortoise and the Hare: Understanding the Influence of Sequence Length and Variability on Decision-Making in Skilled Performance

Tetris is a complex task which taps into several human skills: among them perceptual learning, planning, motor skills, and sequential decision-making. Following a divide-and-conquer strategy, we adopt a machine modeling approach to isolate the contribution of sequential decision-making from the other three skills. In two studies, we test three sets of 1,771,561 feature-based machine players (MPs) (116, 11 weights for each of 6 features) of Tetris for both long-running (tortoise) and short-running (hare) MPs. Tortoise models run until they die. Hare models are stopped after 506 episodes. For both studies, we select the longest running tortoise model and compare its score and behavior with that of the best scoring hare model. The best tortoise models adopt an endurance strategy which emphasizes single-line over multi-line clears. The best hare models adopt an escalation strategy which stresses multi-line clears. In contrast, our human players tend to adopt the escalation strategy early in their game but switch to the endurance strategy as speed demands increase. Unexpectedly, across three model runs, each with a different random seed, we obtain three different sets of “best fitting” models, that is, the MPs overfit the data even though that data is generated by an essentially infinite random sequence. However, in each model run, the best tortoise adopted the endurance strategy and the best hare adopted escalation.