Source
- Bogacz et al. (2006), LIT-569, read in NOTE-461: optimal thresholds and starting points (Figs. 12–16), and the closing section “Optimization and Cognitive Control”.
- Ratcliff & McKoon (2008), LIT-576, read in NOTE-445: Experiments 1–3 and §8 on individual differences.
- Schurger, Sitt & Dehaene (2012), LIT-605, read in NOTE-462: the model’s urgency term and the Discussion.
What was actually shown
The proposal, stated by Bogacz et al. The closing section of LIT-569 says: “Within the framework of the DDM, we can think of control as the adjustment of these parameters to optimize performance”, drift for attention, starting point for expectancy, threshold for the speed–accuracy trade-off. The paper makes this precise where it can. For reward rate the optimal threshold is unique and depends on the delays only through their sum. With unequal priors the optimal starting point is proportional to the log prior odds, and the data of Laming, Link and Van Zandt lie closer to that rule than to the alternative (significantly so for one data set). In their own dot-motion experiment, one reported participant’s fitted threshold rose from 0.16 to 0.26 as the delay between trials grew.
Selective influence, from Ratcliff and McKoon. In LIT-576’s three experiments (14 to 17 participants each), a change of goal moved one parameter and the stimulus moved another:
- speed versus accuracy instructions changed median RT by 120–200 ms and were fitted by boundary separation alone (0.109 against 0.152), and a separate non-decision time for the two instructions differed by 6 ms;
- a 75:25 stimulus proportion moved the starting point about a third of the way toward the likelier boundary, and letting the drift criterion vary as well improved the fit by 1%;
- coherence, which is the evidence and not a goal, was fitted by drift alone.
This is what could have come out otherwise. If instructions moved drift, the parameters would not separate what the person wants from what the stimulus gives. They did not.
When to act. Schurger et al. (LIT-605) model Libet’s task, where the instruction is to move at no particular time, with a constant urgency input to a leaky accumulator. Because of the leak, urgency does not ramp the trace to threshold. It “simply moves the baseline level of activity closer to the threshold so that a crossing is very likely to happen soon”. They say of the goal: it “is effected by setting up circumstances (moving baseline premotor activation up closer to threshold) that favor the spontaneous initiation of a movement … However, the precise moment is not directly decided by a goal-directed operation”. The fitted model reproduces waiting times and the readiness potential (THEORY-059).
The synthesis is the record’s. Each source states its own part. None states the general claim that goals act through parameters for both which and when.
What this does not say
- Not who sets the parameters, or how. Ratcliff and McKoon say the model is silent on how criteria are set, and object that Bogacz et al.’s learning account cannot explain calibration from a verbal instruction in one trial (§7.1). The account is about where goals act, not about the process that puts them there. THEORY-070 is one proposal for that process.
- Not that every goal is a parameter. It is scoped to tasks the accumulation models fit: fast decisions among a fixed set of options. Choosing what the options are, or forming a plan, is outside it.
- Not optimality. That a parameter is what moves does not mean it moves to the optimum. Bogacz et al. predict that learners will set thresholds too high, and offer this as the explanation of people’s apparent bias toward accuracy.
Connections
- The unity it is about.
behavioral-integration: the interface between a goal and the decision machinery. It says nothing about self-governance; a goal here is whatever sets a parameter, endorsed or not (ADR-024). - THEORY-068 says what comparison the parameters are parameters of, and when it is lossless.
- THEORY-070. The EVC account’s control signals have an identity and an intensity, and its list of identities includes thresholds and attention. So it is compatible with this one and adds a decider. The two are not rivals.
- The supervisory-attention and control-theory traditions. Norman and Shallice’s account of willed control (LIT-600) and Carver and Scheier’s feedback-control framework (LIT-604) are in the record unread. As they are usually described, the supervisory system acts by biasing the activation of competing schemas, which would be this account’s picture. That is not checked, and nothing here rests on it.