Continuous vs. Intermittent Reinforcement: Why Less Reinforcement Produces More Persistence

How continuous and intermittent reinforcement differ

Every schedule answers one question: which responses will be reinforced? On continuous reinforcement the answer is all of them (FR 1 is the same thing under another name). A rat whose every lever press produces a pellet, a toddler whose every "please" produces the cookie, a light switch that works every time — each is on CRF, and the learner cannot fail to notice the relation.^[1]

On an intermittent schedule some responses go unreinforced by design, under a rule that counts responses (fixed and variable ratio) or clock time (fixed and variable interval); the schedules page covers those four patterns. Skinner argued that almost all reinforcement outside the laboratory is intermittent — the angler does not catch a fish on every cast — whereas continuous reinforcement is mostly something people arrange on purpose: a trainer with a pouch of treats, a vending machine.^[3]

Two schedules, two jobs

During acquisition — while a behavior is new, weak, or being shaped — reinforce every occurrence; an unreinforced response early on is a lost opportunity. During maintenance — once the behavior is reliable — move to an intermittent schedule, which holds the behavior with fewer reinforcers and makes it robust when reinforcement is occasionally missed.^[4] Ferster and Skinner followed the same rule with pigeons: establish the response on continuous reinforcement, then introduce the schedule.^[1] Three differences drive it.

• Information. On CRF, one unreinforced response is news; on an intermittent schedule it is ordinary. This is why continuous reinforcement teaches fast and extinguishes fast.
• Satiation. Every reinforcer brings satiation closer. A rat on CRF fills up within a short session; on a lean schedule a few pellets sustain a long one.^[1]
• Rate and pattern. Continuous reinforcement produces a steady, moderate rate limited by consumption time. Ratio schedules produce higher rates, and each intermittent schedule has its own pattern of pausing and running.^[1]

The comparison at a glance

Examples of continuous and intermittent reinforcement

In each row the behavior is the same; only the rule for reinforcing it differs.

The dog-training row shows the usual sequence: continuous reinforcement while the sit is being taught, then intermittent treats with praise on every trial as a conditioned reinforcer. The slot-machine row shows a designer who wants persistence rather than learning: no one needs to be taught to press a button, so the machine can start lean and stay lean — Skinner's gambler, held through long droughts by a variable ratio.^[3]

The partial reinforcement extinction effect

When reinforcement is withdrawn, behavior that was reinforced only some of the time keeps going longer than behavior that was reinforced every time. This is the partial reinforcement extinction effect (PRE), and it is paradoxical on its face: the intermittently reinforced behavior received fewer reinforcers and yet is harder to get rid of.

Skinner came to intermittent reinforcement by accident, when a shortage of hand-made pellets led him to reinforce his rats only once a minute rather than for every press.^[5] In The Behavior of Organisms he reported that rats reinforced on this "periodic" schedule produced far larger extinction curves than rats reinforced for every press.^[6] The first experiment designed to isolate the effect was Lloyd Humphreys's in 1939, and it was not operant at all. Humphreys paired a light with a puff of air to the eye until his human subjects blinked at the light; one group received the puff on every trial, another on a random half. When the puff was discontinued, the half-reinforced group went on blinking for far longer.^[2]

Mowrer and Jones extended the finding to operant behavior in 1945. Rats pressed a bar for food on schedules that paid every press, every second, third, or fourth press, or an irregular mixture; the leaner the schedule, the more presses the rats made in extinction. They also noticed that if each rat's "response" was counted as the run of presses its schedule required, the groups extinguished after roughly the same number of runs — the response-unit hypothesis.^[7] The effect has since been reproduced in runways and Skinner boxes, in rats, pigeons, and people, and Mackintosh's 1974 review treated it as one of the most reliable phenomena in the study of extinction.^[8]

Why less reinforcement produces more persistence

The effect took decades to explain, and each explanation captures something the others miss.

The discrimination hypothesis

The oldest account, from Mowrer and Jones, is that extinction has to be noticed before it can take hold. After continuous reinforcement the first unreinforced response is unmistakable evidence that conditions have changed; after intermittent reinforcement a run of unreinforced responses looks like more of the same.^[7] It is correct as far as it goes. But in two 1962 experiments, animals given partial reinforcement, then a block of continuous reinforcement, and only then extinction still showed the effect, although the shift to extinction should have been just as detectable for them as for animals trained on continuous reinforcement alone.^[9][10] Something learned during partial reinforcement survived the intervening block.

Amsel's frustration theory

Abram Amsel argued that nonreward is not a neutral event. When an organism expects a reinforcer and does not get it, the omission produces frustrative nonreward, an aversive state that disrupts behavior and drives the organism away. On a continuous schedule, frustration is first met in extinction, where it does exactly that. On an intermittent schedule the organism meets frustration during training, keeps responding, and is reinforced, so the cues of anticipated frustration become signals for continuing rather than quitting.^[11] This explains why the effect survives a block of continuous reinforcement.

Capaldi's sequential theory

E. J. Capaldi's account is about memory rather than emotion. On each trial the organism carries a trace of what happened on the last one, and that trace is part of the situation in which the next response occurs. When a nonrewarded trial is followed by a rewarded one, responding is reinforced in the presence of the memory of nonreward. In extinction that memory is the only kind on offer; a subject whose responding has been reinforced in its presence keeps going, and a subject on continuous reinforcement, which never has, stops.^[12] The theory predicts effects of the sequence of trials, not just their proportion, and it handles effects that appear after only a handful of trials, where frustration has little time to develop.

Nevin's behavioral momentum

John Nevin came at the problem from the other direction. In his research on behavioral momentum, resistance to disruption is measured as the proportional decline from baseline, in the same subject, across situations that differ in rate of reinforcement. Measured this way, richer reinforcement produces more resistance to extinction, not less. Nevin argued that the classic group comparison counts absolute responses without controlling for the different baseline rates the schedules produce, and that what remains of the PRE reflects the size of the change from training to extinction: from continuous reinforcement to none is a large change, from lean reinforcement to none a small one.^[13] In the fuller theory, persistence depends on the rate of reinforcement in a context, and the speed of extinction also on how detectable the change is.^[14]

How to thin reinforcement from continuous to intermittent

The transition is called schedule thinning, and it is where most applied failures happen: the goal is to reinforce fewer responses without the behavior faltering.^[4][15]

1. Establish the behavior on continuous reinforcement. Stay on CRF until the behavior occurs promptly and reliably whenever the occasion arises: the dog sits on the first cue, the student raises a hand without a reminder.
2. Take the smallest step. From every response, go to two of every three, then every other (FR 2), then two of every five — steps small enough that the learner hardly notices. Big jumps are the usual cause of collapse.
3. Make it variable as soon as it is intermittent. Past every-other, vary the requirement around the average rather than fixing it. Variable schedules produce steadier responding with fewer pauses, and they are what the world will impose anyway.^[1]
4. Move only when performance is stable. Raise the requirement after the behavior has held steady at the current step, not on a timetable.
5. Watch for ratio strain and back off. Pausing, slowing, errors, and emotional behavior after a step up are ratio strain: the schedule has been stretched faster than the behavior can bear. Return to the last requirement that worked and stretch again more gradually.^[4]
6. Keep a conditioned reinforcer on every response. The tangible reinforcer becomes intermittent; praise, a click, or a check mark can stay continuous. In a token economy, the tokens thin and the exchange for backup reinforcers is spaced out separately.^[15]
7. Hand the behavior to natural reinforcers. The endpoint is no contrived schedule at all: the sit that produces the walk, the hand-raise that produces the turn to speak; the habits page covers this handoff for your own behavior.

Why intermittently reinforced problem behavior is so hard to extinguish

The partial reinforcement extinction effect has a dark side. Problem behavior in homes, classrooms, and institutions is almost never reinforced continuously: a tantrum is given in to on some occasions and not others; a self-injurious act draws attention from one staff member and is ignored by the next. By the time anyone tries extinction, the behavior has a long history on a lean, variable schedule — the condition that produces the greatest persistence.

Lerman and Iwata's 1996 review of basic and applied extinction research made the point directly: laboratory findings predict that intermittently reinforced behavior will resist extinction, the natural environment supplies intermittent reinforcement almost by default, and applied studies that varied the schedule before extinction were scarce. Their practical conclusions still hold: expect extinction to be slow, expect an extinction burst, and combine extinction with differential reinforcement of an alternative behavior rather than relying on withholding alone.^[16]

Behavioral momentum adds a warning: reinforcing an alternative behavior in the same context enriches it, and Nevin and Grace's theory predicts that a richer context makes all behavior in it more resistant to change, the problem behavior included.^[14] Practitioners therefore deliver the alternative reinforcement in a clearly different situation where they can, and above all they stay consistent: a single give-in during extinction is a reinforcer delivered on the leanest schedule yet.

"Intermittent reinforcement" in relationships

The term has escaped the laboratory. In articles about dating and abuse, "intermittent reinforcement" describes a partner who is warm one day and cold or cruel the next, and the argument is that the unpredictability itself keeps the other person attached, often under the heading of "trauma bonding." It is worth being precise about what the schedule concept supports.

What it supports is a claim about behavior. If someone's texting, apologizing, or effort to please has been reinforced by affection on an unpredictable schedule, that behavior will persist longer when the affection stops than it would have under reliable affection. People are not exempt from the effect; Humphreys's original subjects were people.^[2]

What it does not support is any claim about the bond. The laboratory effect is measured in lever presses and eyeblinks after reinforcement stops; it says nothing about whether unpredictable affection produces stronger attachment than reliable affection, and no schedule experiment has tested that. The "trauma bonding" claims come from clinical and popular writing, not schedule research, and the evidence for them as operant findings is thin. Real relationships also contain much that a schedule analysis leaves out: relief when tension breaks (negative reinforcement), fear, money, children, and the practical difficulty of leaving. What the term fairly says is that behavior reinforced unpredictably will be hard to stop, and will get worse before it fades.

Common mistakes

• Starting intermittent. Reinforcing "every few" sits before the puppy knows what a sit is. A behavior that has not been acquired cannot be maintained.
• Thinning too fast. Jumping from every response to one in ten and concluding that the learner "lost motivation." The behavior did not fail; the schedule did.
• Never thinning. The learner satiates, the behavior comes to depend on the trainer's presence, and it collapses the first time the reinforcer is unavailable.
• Confusing intermittent with inconsistent. Intermittent reinforcement is a rule applied on purpose to an established behavior. Inconsistency is an accidental schedule applied to whatever happens to be occurring, which is often the behavior nobody wanted.
• Confusing intermittent with noncontingent. A treat that arrives whether or not the dog sat is not a schedule of reinforcement for sitting; intermittent means some responses pay, not that reinforcers arrive at random.
• Reading the effect as "intermittent is stronger." It is more resistant to extinction, which is not the same thing. On Nevin's measures, richly reinforced behavior is more resistant to other disruptions, such as satiation and free reinforcers, than leanly reinforced behavior.^[13]

What the evidence does not show

• It does not show that intermittent reinforcement produces better learning. Acquisition is slower, and a behavior started on a lean schedule may never be acquired. The advantage is durability once reinforcement stops.
• It does not show a single, unconditional law. The size of the effect, and in some analyses its direction, depends on how resistance is measured — absolute responses across groups or proportional decline within a subject — and on the training history.^[13][8]
• It does not show that intermittently reinforced behavior cannot be extinguished. Extinction is slower, sometimes by a great deal; it is not prevented. Skinner's periodically reinforced rats stopped pressing in the end.^[6]
• It does not identify an optimal ratio or a recipe for thinning. The applied literature recommends small steps and stability at each step; the numbers depend on the learner, the behavior, and the reinforcer.^[4][15]
• It does not say much about clinical populations directly. Lerman and Iwata found few applied studies that manipulated the schedule before extinction; the clinical rule of thumb is an extrapolation from the laboratory.^[16]
• It does not explain attachment. Claims that unpredictable affection creates a distinctive bond are not schedule findings.

Key takeaways

• Continuous reinforcement follows every response with the reinforcer; intermittent reinforcement follows only some. The first teaches a new behavior fastest, and the second maintains an established one with fewer reinforcers and less satiation.
• Behavior built on intermittent reinforcement persists far longer when reinforcement stops, even though it received fewer reinforcers: the partial reinforcement extinction effect, shown by Humphreys in 1939 and extended to rats by Mowrer and Jones in 1945.
• The explanations — a harder-to-detect change, counterconditioned frustration (Amsel), responding conditioned to the memory of nonreward (Capaldi), a smaller stimulus change (Nevin) — agree that an organism which has never met an unreinforced response has no defense against the first one.
• To thin a schedule, establish the behavior on continuous reinforcement, take the smallest step, go variable as soon as the schedule is intermittent, move only when performance is stable, and back off at the first sign of ratio strain.
• Problem behavior is almost always reinforced intermittently by accident, which is why it resists extinction in clinics and homes; expect a slow extinction and a burst, and pair extinction with reinforcement of an alternative behavior.
• The schedule concept describes the persistence of a behavior, not the strength of a bond; the popular use of "intermittent reinforcement" to explain relationships holds only as a claim about behavior.

Check yourself

A new owner starts teaching a puppy to sit and gives a treat for roughly every third sit "so the behavior will be strong." Two weeks later the puppy still does not sit on cue. What went wrong?

The schedule was thinned before the behavior existed. On an intermittent schedule most sits go unreinforced, so the puppy rarely contacts the relation between sitting and the treat and the behavior never gets established. Intermittent reinforcement maintains a behavior that has already been acquired; it does not build one. Start on continuous reinforcement, and thin only once the sit is prompt and reliable.

A teacher has kept a class working quietly for a month by praising them after every good five-minute block. She decides that is too much praise and switches to praising once per lesson. Within a week the class is noisy again. What happened, and what should she do?

She stretched the schedule far too fast, from every block to roughly one in ten, and the behavior showed ratio strain and then broke down. The class did not lose motivation; the schedule did not hold. She should return to a requirement close to the one that worked, hold there until quiet work is stable, then thin in small, variable steps, keeping a brief acknowledgment on most blocks as a conditioned reinforcer while the tangible reinforcer thins.

A friend says her partner's unpredictable affection has "intermittently reinforced" her into staying. Is the schedule concept being used correctly?

Partly. The concept applies to behavior: if her efforts to reach out have been reinforced by affection only unpredictably, those efforts will persist longer than they would have under reliable affection, and there is no reason to think people are exempt. But the laboratory effect says nothing about attachment or the strength of a bond, and staying in a relationship involves negative reinforcement, fear, and practical constraints that a schedule analysis leaves out. The fair conclusion is only that the behavior will be hard to stop and will get worse before it fades.

Want more? The 20-question quiz covers every page on the site with instant explanations.