Where does baseline collide?

In behavior analysis and single-subject experimental design, the baseline is the phase in which a target behaviour is measured repeatedly before any intervention, establishing the naturally occurring rate against which a treatment effect is later judged. Baer, Wolf, and Risley's 1968 Some current dimensions of applied behavior analysis set out the logic by which a stable baseline makes a subsequent change interpretable as an effect of the intervention. The referent is a measured stretch of behaviour-over-time serving as a within-subject control condition, and it has nothing to do with a reference line on the ground or a term inside a learning rule beyond the shared idea of a fixed point of comparison.[1]

Unit

The stable rate or level of behaviour recorded before any intervention

Recovery means

A return-to-baseline (the reversal in an ABA design) tests whether behaviour reverts once the intervention is withdrawn

Measured by

Recorded by repeated measurement of the target behaviour across sessions until a stable pattern appears

In reinforcement learning, a baseline is a reference value subtracted from the observed return in a policy-gradient update, so that only better- or worse-than-expected outcomes push the policy; because the baseline does not depend on the action it leaves the gradient's expectation unchanged while sharply reducing its variance. Williams's 1992 Simple statistical gradient-following algorithms for connectionist reinforcement learning introduced the REINFORCE family in which such a reinforcement baseline appears. The referent is a scalar inside a learning rule, unrelated to a geodetic reference line or a behavioural control phase apart from the common idea of a comparison point.[2]

Unit

A scalar reference value subtracted from the return in the update

Recovery means

Not applicable

Measured by

Chosen as a running estimate of expected return (a learned value function or average) and tuned to minimise gradient variance