Table 1 A list of the learning models fitted to judgments of automation reliability, and associated learning parameters
From: How do humans learn about the reliability of automation?
Model | Description | Learning parameters |
|---|---|---|
Bayesian | Bayesian learning of automation reliability assuming a single true state | \(p, q\) |
Delta | Judgements of automation reliability are updated based upon the prediction error (delta) between the previous reliability estimate and the current automation accuracy | \(r_{0}\), \(\alpha\) |
Two-Kernel Delta | Two simultaneous delta-rule learners track automation reliability. Estimates are taken from the slower learner unless the difference between the two processes is above a threshold, signalling a shift in the environment, in which case the fast delta learner is used | \(r_{0}\), αfast, αslow, T |
Sampling (proportional to delta weights) | A single previous memory is sampled to inform the current estimate of automation reliability. Previous experiences are sampled proportionately to their weights under a delta-rule updating process | r0sampling_recency, αsampling |
Sampling (last/average) | Samples either the most recent experience with automation, or the average reliability of all previous experiences | r0sampling_last_average, weightr0, probt |
Contingent Sampling | Automation reliability is assumed to be sensitive to the history of automation accuracy over the recent m contacts. Thus, the reliability estimate is based on previous cases where the history of automation accuracy m contacts back matches the history m contacts back in the current instance | r0sampling_contingent, m |
IIAB | Estimates of automation reliability are updated in a stepwise manner when a “change point” is identified. Sometimes, the model has “second thoughts” and expunges or updates a previous change point | T1, T2, \(p, q, p_{{\text{change point}}} , q_{{\text{change point}}}\) |
No updating | No learning process | \(r_{0}\) |