A repository of code for a comparison of Ratcliff's Drift-Diffusion Model to an ACT-R model of choice. The ACT-R model implements a simple version of a one-stimulus, two-alternative forced choice (2AFC) task.
In DDM, decisions are modeled as drifting processes that proceed towards one of two possible thresholds (corresponding to the two options). The model is controlled by three parameters:
- A drift parameter v
- A threshold or boundary parameter a, which represents the distance from either of the response thresholds.
- A non-decision time T, which represents response and encoding times.
In ACT-R, decisions can be made through several mechanisms, procedural and declarative. Here, we describe a general enough mechanisms that comprises both, and is based on the following steps:
- The model holds in mind an internal correctness criterion in the working memory (WM) buffer, and maintains the two options in the visual buffer.
- The model retrieves the S/R mappings from LTM.
- The model proceeds through a check stage, during which
it compares the retrieved S/R against the internal
correctness criterion.
- If a match is found, the model responds.
- If not, the model either...
- Restarts the search with probability p
- Responds anyway with probability (1 - p).
The main parameters in the ACT-R implementation are:
- W: the spreading activation from the internal correctness criterion. This corresponds to WM in ACT-R and should correspond to v in DDM.
- S: The noise in declarative memory. This mimics the resistance to spreading activation.
- U: The perceived utility of restarting the search process, defined in terms of RL expected value V. This determines the boundary parameter a.
The ACT-R device implements a minimal implementation of a 2AFC task. Specifically, the task contains only one stimulus, and two responses (left and right) are possible. By default, left is assumed to be the correct response. The stimulus presented is also the only visual object presented on the screen. The stimulus consists of a visual label, which can be either "correct" or "incorrect". The label itself is unimportant, as the correct response is always the same (left). This matches the DDM methods, where one boundary is assumed to be the correct response for all trials.
The difference between "correct" and "incorrect" trials (or, if one
prefers, the "correct", "neutral", and "incorrect" trials) has no
effect until one turns on the visual-activation
parameter in ACT-R,
which enables preading activation. This would create competition for
various responses.