Does not consistently providing a reward strengthen operant conditioning?

When learning about Operant Conditioning, I remember being taught that not consistently rewarding the desired behavior could (seemingly counterintuitively) actually increase the strength of learning the conditioned behavior. Is this correct?

If this is correct, is it more effective to begin providing a reward at or near 100% of the time the conditioned behavior is performed, and then slowly reduce the amount to a lower percentage? Or is it more effective to simply start with a lower percentage?

-
The first part of this answer might also be helpful. –  Artem Kaznatcheev Jun 11 '12 at 12:11

What you're thinking of is called an Intermittent Reinforcement Schedule. Different Reinforcement schedules have different effects, variable ratio in particular seems to be what you're getting at:

a reinforcement schedule in which the number of responses necessary to produce reinforcement varies from trial to trial

What's important is the behavior is reinforced intermittently but regularly enough to be "worth" it. This produces a powerful effect; I don't immediately have any sources but I do believe B.F. Skinner ran a number of such trials. where the subject will repeatedly perform the behavior in an attempt to get more rewards.

An explanation I've heard is that with a predictable reward schedule (fixed ratio) your participant knows they can except a reward, so there's no reason to mindlessly perform the task over and over; compare a Slot Machine (variable ratio) to a machine that gives out money on a fixed number of tries (fixed ratio). A subject knows with the fixed ratio machine they can always get their reward at any time. With the slot machine you have to keep playing and playing to get that big next reward.

Remember that for proper Shaping you should start with a high percentage (100%) initially to condition the response otherwise the participant might not be conditioned at all. Going back to the gambling example, it's much more effective if you give your participant a taste of winning when they play the slots or they'll quickly quit as it seems there's no reward to be had at all.

Shaping involves immediate rewarding behaviors approaching the desired behavior, it's certainly more important to start with a high % of rewarding, especially with non-human subjects when getting them to perform the wanted task can be problematic to begin with.

-
Just as a tack-on, providing a reward at near 100% of the time could interfere with operant conditioning. A classic example is one where children like to, say, ask parents if they could have ice cream without having finished dinner. A parent may usually refuse, but if he/she gives in every so often - or on occasion - then a child will continue with repeating the behavior, knowing full well that the chance it's parent(s) will give in is, well, there. –  PheonixEnder Jun 11 '12 at 11:00
psycnet.apa.org/journals/bul/47/3/193 might be a relevant reference. –  Ofri Raviv Aug 18 '12 at 20:59

'Reinforcement' is anything that increases the chances that an organism with repeat a behavior. When you are teaching a behavior, in the beginning, eliciting behavior takes a lot of reinforcement. So, for example, if you want to teach you dog to come to you every time you go to the back door, you feed him a cookie when you go to the back door. After a while this behavior becomes REALLY strong. He expects a cookie. Now you can work on another principle of operant conditioning called 'extinction'. It works like this. If you stop reinforcing a behavior, the dog in this case, gets frustrated, and trys really hard to get that reinforcement to continue. There are two possible outcomes. Either A: extinction, where the behavior that was being reinforced vanishes, or B: The behavior escalates until you reinforce it, and you get a new level of behavior. (this is how bad habits get worse, and good habits get better)

So, if you want to get a really strong behavior you need both kinds of reinforcement schedules. First you start off with a fixed reinforcement schedule to build the expectation of reinforcement, then you switch to a reinforcement schedule that causes some level of frustration to elicit the extinction response, and you intentionally 'fail' extinction to get a stronger behavior.

A completely random reinforcement schedule is not nearly as effective as a trainer that is tuned into the emotions of the animal that is being trained. A really good trainer can cause just enough frustration to get a little more behavior, without causing so much that the animal just gives up.

-