Shaping: Differential Reinforcement of Successive Approximations — A Practitioner's Guide

Successive approximations

Shaping reinforces successive approximations toward a terminal target response, withholding reinforcement from the current approximation once a closer one reliably occurs and reinforcing only the new, closer form. The core skill is selecting approximation steps close enough together that the learner can bridge each one without extinguishing, a step too large stalls the whole program, but not so small that progress crawls and the learner satiates on a marginally different response for sessions on end (Ghaemmaghami et al., 2018; Hodges et al., 2017).

Differential reinforcement of the response class

Underneath every shaping program is a moving differential reinforcement contingency: at each stage, reinforcement is delivered for members of the current approximation's response class and withheld from the previous, now-obsolete approximation. This is the same logic used elsewhere in the differential reinforcement family, applied to a response class whose criterion changes over time rather than staying fixed.

Percentile schedules

A percentile schedule reinforces any instance of behavior that falls above a criterion computed from the learner's own recent performance, commonly a rolling percentile (the current response needs to beat, say, 70 percent of the last several responses) rather than a fixed, hand-set step (Kwak et al., 2022). Because the criterion recalculates itself as the learner's baseline shifts, a percentile schedule removes the guesswork of picking the next fixed approximation and self-adjusts to the learner's actual pace.

Shaping versus chaining

Shaping and chaining both build complex behavior gradually, but they operate on different units. Shaping changes the topography or dimension (force, duration, accuracy, latency) of a single response class across successive approximations until it matches a terminal target that did not previously exist in that form. Chaining links already-existing component responses, each already at or near criterion on its own, into a sequenced complex behavior, using forward, backward, or total-task chaining, without changing any single component's topography. A practical rule: if the learner cannot yet emit any recognizable version of the target response, that's a shaping problem; if the learner can already emit all the pieces but not in a connected sequence, that's a chaining problem.

Shaping versus prompting

Shaping and prompting-hierarchy procedures are often combined rather than chosen between. A prompting hierarchy assumes the target response, or something close enough to it, already exists in the repertoire and can be evoked with enough antecedent support, then faded. Shaping assumes the terminal response does not yet exist in any usable form and has to be built by reinforcing successive approximations, because no amount of prompting can evoke a response the learner has never emitted. In practice, many acquisition programs use both in sequence: shape a rough approximation first, then apply prompt fading to sharpen and generalize the final form (Aravamudhan et al., 2020).

Reach for shaping specifically when nothing in the repertoire is close enough for a prompt to evoke or a chain to link, per the distinctions above. Starting a full shaping program from approximation zero when a rough, prompted version already occurs costs time you don't need to spend.

Setting the initial criterion is where most shaping programs succeed or fail before the first session even runs. Take a baseline of the learner's current best approximation to the target, not the target itself, and set the first reinforced step just beyond that baseline, close enough that the learner can reasonably contact reinforcement within the first several trials. A first step set too close to the terminal target produces an extinction-heavy session with little to reinforce; a first step set too close to baseline wastes sessions reinforcing something the learner could already do.

For step size across the program, a fixed, hand-picked sequence of approximations works when you have a clear picture of the terminal topography and reasonable confidence in the intermediate steps, common in verbal-behavior shaping (Leigland, 1996) and articulation work (Aravamudhan et al., 2020). Reach for a percentile schedule instead when the terminal criterion is more about degree than topography (on-task duration, work rate, response force), or when a fixed sequence keeps stalling. One demonstration using a percentile schedule on recent on-task performance doubled the target behavior without adding materials or breaks, direct support for reaching for it over a fixed sequence once the fixed steps stop moving the learner (Kwak et al., 2022).

Once the initial step is set, the ongoing work is pacing the criterion advance correctly, session by session.

Build criterion shifts around consecutive-session stability, not a single good session. One demonstration teaching complex functional communication responses advanced the criterion only after two consecutive sessions at the current step, moving from simple to complex responses across three to four defined steps without triggering resurgence of the earlier, simpler form (Ghaemmaghami et al., 2018). That two-session stability rule is a concrete, exportable heuristic for any shaping program: if you're advancing criteria off a single strong session, you are moving faster than the data support, and watch for resurgence of the earlier approximation as the tell that you did.

Watch response topography, not just approximation count, in domains like feeding. Shaping food acceptance with a large, rotating set of target foods builds variety faster, but if a learner stalls at merely interacting with food without eating, downsizing to a small, constant set is the correct correction, not simply reinforcing more interaction trials (Turner et al., 2020).

Shaping pairs naturally with prompting and chaining rather than standing alone. Applied articulation programs that combine shaping with prompted trials and immediate feedback get faster acquisition than shaping run in isolation (Aravamudhan et al., 2020). A shaping program written without a paired prompt or feedback component is likely leaving speed on the table.

Two errors show up more than others. The first is a shaping hierarchy with fixed step sizes and no explicit rule for when to move the criterion, which leaves staff to advance too fast, extinguishing the earlier approximation, or too slow, inviting satiation and a plateaued session. The second is a criterion that quietly drifts because staff reinforce both the current and the prior approximation inconsistently, which stalls the program the same way integrity lapses stall any other differential-reinforcement plan. Naming the advance rule in the written plan itself, session count at criterion, percentile threshold, or both, and defining the current approximation precisely enough that two staff would score it the same way, is what turns a shaping idea into a shaping program (Parnell et al., 2017).