Yoked criteria shifts in decision system adaptation: Computational and behavioral investigations

Yoked Criteria Shifts in Decision System Adaptation: Computational and Behavioral Investigations Blair C. Armstrong (blairarm@andrew.cmu.edu) Department of Psychology and the Center for the Neural Basis of Cognition, Carnegie Mellon University 5000 Forbes Avenue, Pittsburgh, PA 15213 USA Steve Joordens (joordens@psych.utoronto.ca) Department of Psychology and the Centre for Computational Cognitive Neuroscience, University of Toronto Scarborough 1265 Military Trail, Toronto, ON M1C 1A4 Canada David C. Plaut (plaut@cmu.edu) Department of Psychology and the Center for the Neural Basis of Cognition, Carnegie Mellon University 5000 Forbes Avenue, Pittsburgh, PA 15213 USA Abstract We describe a theory of decision system adaptation in which yoked criteria shifts serve as a simple but powerful mechanism for rapidly minimizing errors without sacrificing speed. To support our theory, we implemented a connectionist model of lexical decision, wherein the state of a word perception network was “read” by a pair of decision units. The response criteria for these decision units were then subjected to yoked shifts to examine how, in the face of perceived errors, such a response mechanism might adjust performance. We also present the results of a lexical decision experiment that manipulated the truthfulness of the feedback participants received so as to trigger the error correction mechanism while keeping other task parameters constant. The results of the experiment largely parallel those of the simulation, suggesting that yoked decision shifts make an important contribution to error minimization in decision system adaptation. Keywords: decision making, decision system adaptation, yoked criteria shifts, lexical decision, connectionist modeling. An individual’s ability to rapidly and correctly decide between two alternatives is critical to their survival and wellbeing. For example, a new driver may learn to brake or accelerate when faced with a yellow light under dry road conditions. However, if one day it snows their established decision behavior will need to be adapted to accommodate this fact. Thus, the driver must be capable both of deriving an initial calibration of their decision system, and of rapidly adapting this system in face of change. The work reported in this paper focuses on how the updating of a previously well-calibrated decision system is accomplished. Motivating our work is an interesting pattern of effects reported by Gomez, Ratcliff, and Perea (2007), who varied the task participants completed (two-choice lexical decision vs. go/no go lexical decision) in a within- subjects design. After fitting their data with a diffusion model, the authors determined that changes of the decision criteria were key in accounting for performance differences across task blocks, with adjustments to most other parameters only having a modest effect. Within the context of decision system adaptation, Gomez et al.’s (2007) findings might also suggest that rather than re-configure the system de novo when faced with different task demands, participants may adapt to the new task primarily by shifting their decision criteria. In many cases, such a shift may provide a rapid means of accommodating modest (and perhaps not so modest) changes without necessitating a potentially costly re-derivation of all of the parameters in the decision system. If we assume that decision system adaptation occurs via shifts of decision criteria, this raises the question of how the criteria should be shifted. We examined this issue within the context of an abstract forced choice task wherein participants must make rapid and accurate A and B responses indicating the presence of stimuli a and b. Imagine that after becoming proficient at the task, some property of the task changes such that participants’ find themselves incorrectly responding A to b items (e.g., all the b’s become more a-like). One logical adaptation would be to shift the A decision criterion so as to require the additional accumulation of evidence that an a was presented before making an A decision. However, such an adaptation would have the result of slowing overall reaction time (RT) because the average amount of evidence to be collected before making a particular decision will have increased―an effect which may not be adaptive if there are benefits associated with being fast. The issue of overall increases in RT can be avoided, however, if both criteria are shifted in unison, such that when more evidence is required to make an A response, less evidence is required to make a B response. By keeping the average response criteria constant, overall accuracy and RT should remain (approximately) constant. Furthermore, this yoked shift should lead to 1) an increase in accuracy for B responses as less evidence must be accumulated to reach the B criterion, and by corollary that 2) RTs for b stimuli should decrease as accumulating less evidence should take less time; the converse―namely decreased accuracy and increased RT―would be predicted for A responses. It is worth noting that data compatible with these speed-accuracy relationships were reported by Wagenmakers, Ratcliff, Gomez, and McKoon (2008) in a lexical decision