4 ms·

I'm really struggling with making sense of this. Assuming that you're assigning exactly 50/50 results in each case: Half the heads, having been praised were s

I'm really struggling with making sense of this. Assuming that you're assigning exactly 50/50 results in each case: Half the heads, having been praised were successful again. That's implying that praise was 50% effective. Half the tails, having been scolded, were suddenly successful. That's implying that scolding was 50% effective. How is scolding more effective than praising?

contravariant 9y ago
The point of this example is that there is (obviously) no difference in effectiveness. However compare this simplified example with the following citation from the article: >“On many occasions I have praised flight cadets for clean execution of some aerobatic maneuver, and in general when they try it again, they do worse. On the other hand, I have often screamed at cadets for bad execution, and in general they do better the next time. So please don't tell us that reinforcement works and punishment does not, because the opposite is the case.”
dragonwriter 9y ago
With praise, half got worse and half stayed the same. With scolding, half got better and half stayed the same. Scolding is obviously better, in terms of effect on outcomes. Though it's probably a better parallel to the flight instructor case if you have a six-sided die, and you praise on 5-6, and scold on 1-2: most of the praised get worse, most of the scolded get better, those who get neither scolding or praise have mixed results, and a few of those scolded get worse and likewise a few praised get better. The coin flip example, while in one sense more clearly showing the problem, simplifies to the point where the connection with realistic scenarios is harder to see, IMO.
- DenisM 9y ago
  In other words Gaussian distributions will reliably revert to the mean, whereas feedback will always push to the right. Those praised, already on the right, will move against the desired direction, left, towards the mean. Those scolded, already at the left will move in the desired direction right, also towards the mean.
dagss 9y ago
I think they summed up multiple coin tosses (say 10). Then "scold" or "praise" the losers/winners with lowest/highest sum of 10 tosses. (So you get a probably distribution and regression towards the mean.)

contravariant 9y ago

The point of this example is that there is (obviously) no difference in effectiveness. However compare this simplified example with the following citation from the article: >“On many occasions I have praised flight cadets for clean execution of some aerobatic maneuver, and in general when they try it again, they do worse. On the other hand, I have often screamed at cadets for bad execution, and in general they do better the next time. So please don't tell us that reinforcement works and punishment does not, because the opposite is the case.”

dragonwriter 9y ago

With praise, half got worse and half stayed the same. With scolding, half got better and half stayed the same. Scolding is obviously better, in terms of effect on outcomes. Though it's probably a better parallel to the flight instructor case if you have a six-sided die, and you praise on 5-6, and scold on 1-2: most of the praised get worse, most of the scolded get better, those who get neither scolding or praise have mixed results, and a few of those scolded get worse and likewise a few praised get better. The coin flip example, while in one sense more clearly showing the problem, simplifies to the point where the connection with realistic scenarios is harder to see, IMO.

DenisM 9y ago
In other words Gaussian distributions will reliably revert to the mean, whereas feedback will always push to the right. Those praised, already on the right, will move against the desired direction, left, towards the mean. Those scolded, already at the left will move in the desired direction right, also towards the mean.

dagss 9y ago

I think they summed up multiple coin tosses (say 10). Then "scold" or "praise" the losers/winners with lowest/highest sum of 10 tosses. (So you get a probably distribution and regression towards the mean.)