Korva is a scroll of the internet’s best ideas.
Reward every time to teach
Learning a new trick
Say a dog is learning a new trick. Rewarding every correct response is the smart move, because at the very beginning of the learning curve, a missing reward may look to the dog like a sign it did the wrong thing.
Keeping it up
Once the behavior is learned, reinforcement given only some of the times it occurs can maintain it, on schedules that are predictable or unpredictable. Rewarding takes time away from the behavior's continuation, so paying a worker after each piece assembled on an assembly line simply makes no sense.
Rewarding only some of the time is also cheaper. Not every instance of the desired behavior needs a payoff.
Why must early rewards come immediately?
Edward Thorndike's law of effect, drawn in 1898 from cats escaping puzzle boxes, held that responses followed by satisfying results become more likely. Leo Grice's 1948 experiments with rats found that learning slowed as the delay before the reward grew. Animal trainers bridge the gap with a marker, such as a clicker or a whistle, that sounds at the exact moment of the right response.
How far can rewards be stretched out?
Ferster and Skinner's experiments in the 1950s built up ratio schedules gradually, and their pigeons ended up pecking well over a hundred times for each food delivery. The gradual increase mattered: jumping straight to a high ratio makes responding collapse, a failure called ratio strain. The same experiments showed that responding on a ratio schedule stays steady as long as the rewards keep arriving at some point.
More from Shane
