Decision making QT

•How do we pick speed-accuracy tradeoffs in simple decisions?
•Is there a generic circuit structure in the brain that is involved in all decisions?
•Does the way we pick speed-accuracy tradeoffs in simple decision making tasks generalize to more complex decisions?

In order to understand the physical basis of human and animal decision making, I focus on developing and testing mathematical/computational decision making models that are extremely simple in terms of an analog, neural circuit-level description. These models are nevertheless capable of implementing ideal, decision-theoretic algorithms that optimize some well-defined measure of performance when appropriately parameterized (e.g., maximizing the rate of rewards earned in a decision-making task with many trials). With less fine-tuning on the part of the circuit designer and more homeostatic feedback control on the part of the circuit, they can also approximate optimal performance heuristically, employing minimally complex circuit structures and diminished parameter sensitivity. Optimized, feedback-controlled versions of these models have explained interesting patterns of reaction times and choice biases by humans that I and my colleagues have predicted and then observed in perceptual/economic decision making, and in patterns of electroencephalographic (EEG) and functional magnetic resonance imaging (fMRI) data in tasks similar to those studied in non-human animals.

These models, which consist of networks of stochastic temporal integration devices (i.e., noisy capacitors, or diffusion processes), make clear predictions that can be tested with simple empirical methods – e.g., classic signal detection and two-alternative forced choice experiments. These models have also clearly shown their ability to emulate, and thereby unify, a variety of existing psychological models of operant/instrumental conditioning (melioration/matching; Simen & Cohen, 2009), interval timing (scalar expectancy theory; Simen, Balci, Cohen & Holmes, in review), decision making (sequential probability ratio tests; Balci, Simen, Niyogi, Saxe, Hughes, Holmes & Cohen, in review; Bogacz, Brown, Moehlis, Holmes & Cohen, 2006; Simen, Cohen & Holmes, 2006; Simen, Contreras, Buck, Hu, Holmes & Cohen, 2009) and even features of production system models of complex behavior (SOAR, ACT; Polk, Simen, Lewis & Freedman, 2002; Simen & Polk, 2010).

I am currently focusing on extending these stochastic differential equation-based models of two-alternative tasks to decisions with more decision alternatives (McMillen, Simen & Behseta, in review; Simen, McMillen and Behseta, 2010), and to decisions involving computations on experienced durations of time (Balci, Freestone & Simen, in preparation).

Home

Below is a QuickTime movie showing a two-alternative perceptual decision making circuit in action (click here for a Flash animation). Panel A plots a drift-diffusion process (blue) that accumulates evidence in favor of one choice over the other; after each choice, the evidence level, or belief state, decays back toward 0; green dashed line depicts the reward-maximizing threshold. Panel B plots a reward rate estimate computed with a leaky integrator; green dashed line depicts the maximum possible expected reward rate. This estimate determines the decision boundaries, or thresholds, applied to evidence accumulation in panel A (red). The signal presented on any trial is shown in panel C in green (upward steps represent continuously presented evidence for choice 1; downward represents evidence for choice 2; a signal of 0 indicates an inter-trial period). Responses are shown as red dots. Errors are plotted in panel D. Response times are plotted in each trial in panel E. Decision times and accuracy converge to the reward-maximizing speed-accuracy tradeoff for this task (Simen, Cohen & Holmes, 2006).