Diffusion Decision Model: Current Issues and History

(Ratcliff, Smith, Brown, & McKoon, 2016)

The article begins with a description of two types of everyday decision making, namely, decisions made at a low cognitive level, and those made at a higher cognitive level with prolonged deliberation.  The article describes diffusion models which address the former type of decisions and these are described as “rapid matching of a perceptual representation to stored knowledge in memory, which allows us to identify things in our immediate surroundings and determine how we should respond to them” (Ratcliff et al., 2016, p. 260). In a laboratory, these decisions are typically based on simple, two-choice tasks with the important measurements being response times (RT) and the probabilities of making the two choices with researchers’ interest being focused on speed versus accuracy.

Research has often been centred around sequential sampling models. The authors provide a useful diagrammatic breakdown of these models.

In addition they explain what diffusion models are and terminology used in relation to these models. Simply put, the model represents the rate of accumulation of evidence in a two-choice scenario and the setting of boundaries (or thresholds). The models describe how, from a starting point (where evidence about a stimulus from perception or memory begins) evidence accumulates to reach a boundary or threshold (as this is two choice decisions, there are two boundaries) resulting in a decision. Boundaries are defined as the amount of evidence that has to be accumulated before a decision is made and the average rate of accumulation is called the ‘drift rate’.

In the two-choice diffusion model, drift rates are determined according to the quality of evidence, speed-accuracy trade off and the effects of bias towards one or the other boundaries. With respect to sequential effects, the authors suggest that with easy stimuli and rapid presentation of the sequences, sequential effects can be found over several trials and that the diffusion model can model sequential effects.

When discussing Across-trial variability in model components, the authors suggest that random walk models (discrete precursors of diffusion models), predicted that RT did not differ for either correct or incorrect responses and mention research that suggest that there is cross-trial variability and that across-trial variability can be accounted for by collapsing boundary models. This has been supported by research using EEG.

The authors then discuss research which indicates implicit boundaries and is called the response signal task where stimulus is presented and then after some time, a signal is given. It is suggested that a mixture of processes is responsible, namely a decision that is terminated at the boundary, and a decision that has not terminated at the boundary. Conflict tasks also seem to require dynamic changes in diffusion model parameters over time.

Besides RT, optimality has also been investigated through the use of the sequential probability ratio test. The authors however caution that these tests have less appeal to real-life decision-making scenarios. They discuss an alternative to optimality which is reward rate maximisation (maximising the number of correct decisions, and the associated reward, per unit time).

Linking optimality theory and neural decision making has led to collapsing bounds theories. Less evidence is required to create a decision as time passes – in other words, boundaries collapse over time, coming closer to the centre thus reducing time taken to reach the boundary.

On average, incorrect decisions are generally slower on average than correct decisions. Collapsing boundaries and increasing urgency signals may account for this.

In discussing expanded judgement tasks where a noisy sequence of stimulus elements has to be integrated to make a decision, the authors point out that that although using external stimulus noise and internal noise is an attractive idea, the equivalence of these two stimuli is questionable. Expanded judgement tasks need a new representation of every element in the sequence, which must be integrated with the memory representation of the elements that precede it.

MY NOTES:

1. A very useful glossary of terms is given in the article.
2. Good break down of what falls under sequential sampling…

——————————————————————————————————–

Under the heading of Sequential sampling models the following occurs:

1. RANDOM WALK/ DIFFUSION MODELS (relative evidence criteria)
   1. Diffusion Models (Continuous time continuous evidence)
      1. Ornstein Uhlenbeck (OU) Model
      2. Standard Wiener Diffusion Model (constant drift)
   2. Random Walk Models (discrete time continuous evidence)
2. ACCUMULATOR MODELS (absolute evidence criteria)
   1. Hybrid accumulator/diffusion models (absolute evidence criteria)
      1. Dual diffusion model (Independent evidence totals with or without decay in drift)
      2. Leaky competing accumulator (LCA) model (inhibition between evidence totals, decay in drift)
   2. Recruitment Model (discrete time discrete evidence)
   3. Accumulator Model (Discrete time continuous evidence)
   4. Poisson Counter Model (Continuous time, discrete evidence)
   5. Linear ballistic accumulator model (continuous time, continuous evidence: non-stochastic

—————————————————————————————————

3. See p 275 under the heading “brief stimulus presentation” where they talk about stimulus information over time.
4. Also see p 277 for “outstanding questions” ie. possible research related to this.

 

Ratcliff, R., Smith, P. L., Brown, S. D., & McKoon, G. (2016). Diffusion Decision Model: Current Issues and History. Trends in Cognitive Sciences, 20(4), 260–281. https://doi.org/10.1016/j.tics.2016.01.007