Melioration dominates maximization : stable suboptimal performance despite global feedback
2006, Neth, Hansjörg, Sims, Chris R., Gray, Wayne D.
Situations that present individuals with a conflict between local and global gains often evoke a behavioral pattern known as melioration — a preference for immediate rewards over higher long-term gains. Using a variant of a binary forced- choice paradigm by Tunney & Shanks (2002), we explored the potential role of global feedback as a means to reduce this bias. We hypothesized that frequent explicit feedback about future expected and optimal gains might enable decision makers to overcome the documented tendency to meliorate when choices are rewarded probabilistically. Our results suggest that the human tendency to meliorate is tenacious and even prospective normative feedback is insufficient to reliably overcome inefficient choice allocation. We identify human memory limitations as a potential source of this problem and sketch a reinforcement learning model that mimics the effects of a variable feedback horizon on performance. We conclude that melioration is a powerful explanatory mechanism that can account for a wide range of human behavior.
Juggling multiple tasks : a rational analysis of multitasking in a synthetic task environment
2006, Neth, Hansjörg, Khemlani, Sangeet S., Oppermann, Brittney, Gray, Wayne D.
Tardast (Shakeri 2003; Shakeri & Funk, in press) is a new and intriguing paradigm to investigate human multitasking behavior, complex system management, and supervisory control. We present a replication and extension of the original Tardast study that assesses operators’ learning curve and explains gains in performance in terms of increased sensitivity to task parameters and a superior ability of better operators to prioritize tasks. We then compare human performance profiles to various artificial software agents that provide benchmarks of optimal and baseline performance and illustrate the outcomes of simple heuristic strategies. Whereas it is not surprising that human operators fail to achieve an ideal criterion of performance, we demonstrate that humans also fall short of a principally achievable standard. Despite significant improvements with practice, Tardast operators exhibit stable sub-optimal performance in their time-to-task allocations.
You can't play straight TRACS and win : memory updates in a dynamic Task environment
2004, Neth, Hansjörg, Sims, Chris R., Veksler, Vladislav D., Gray, Wayne D.
To investigate people's ability to update memory in a dynamic task environment we use the experimental card game TRACSTM (Burns, 2001). In many card games card counting is a component of optimal performance. However, for TRACS, Burns (2002a) reported that players exhibited a base- line bias: rather than basing their choices on the actual num- ber of cards remaining in the deck, they chose cards based on the initial composition of the deck. Both a task analysis and computer simulation show that a perfectly executed memory update strategy has minimal value in the original game, suggesting that a baseline strategy is a rational adaptation to the demands of the original game. We then redesign the game to maximize the difference in performance between baseline and update strategies. An empirical study with the new game shows that players perform much better than could be achieved by a baseline strategy. Hence, we conclude that people will adopt a memory update strategy when the benefits outweigh the costs.
Steps towards integrated models of cognitive systems : a levels-of-analysis approach to comparing human performance to model predictions in a complex task environment
2006, Schoelles, Michael J., Neth, Hansjörg, Myers, Christopher W., Gray, Wayne D.
Attempts to model complex task environments can serve as benchmarks that enable us to assess the state of cognitive theory and to identify productive topics for future research. Such models must be accompanied by a thorough examination of their fit to overall performance as well as their detailed fit to the microstructure of performance. We provide an example of this approach in our Argus Prime Model of a complex simulated radar operator task that combines real- time demands on human cognitive, perceptual, and action with a dynamic decision-making task. The generally good fit of the model to overall performance is a mark of the power of contemporary cognitive theory and architectures of cognition. The multiple failures of the model to capture fine-grained details of performance mark the limits of contemporary theory and signal productive areas for future research.