How does the Reversal Learning Game work?
You pick one of two cards, 80 times. One card is lucky: it wins 8 times in every 10 you pick it. The other wins only 2 times in 10. Once you have picked the lucky card 8 to 12 times in a row, the two swap without warning, and the card that was paying off mostly stops.
The cards change sides at random between picks, so a side can't stand in for a card. A run counts picks, not wins: picking the lucky card and losing still counts, because the pick was right. How long a run has to be is drawn afresh after every swap, so you can't count your way to it. Wins and losses are dealt from shuffled sets of 10, so everyone gets the same amount of bad luck, in a different order.
How is the Reversal Learning Game scored?
Your result is the number of swaps: how many times you found the lucky card and stuck with it long enough for the two to change places. Two habits cost swaps. Leaving the lucky card after an unlucky loss breaks your run, and staying with a card after it has swapped wastes picks while the new run waits.
Beside it, “followed a swap after” is how many times on average you picked the old card after a swap before moving to the new one, and “stayed after bad luck” is how often you picked the lucky card again straight after it lost. The two pull against each other: the quicker you leave a card that loses, the faster you follow a real swap, and the more runs you break for nothing. How a few fixed ways of playing do over 80 picks, 1,000 simulated games each:
| Way of playing | Swaps, on average |
|---|---|
| Stay until the card loses twice in a row | 5.3 |
| The same, but 1 pick in 10 at random | 3.9 |
| Stay until it loses three times in a row | 5.2 |
| Leave the card after every loss | 0.9 |
| Pick at random | 0.1 |
What is reversal learning?
Reversal learning is noticing that a choice which used to pay off no longer does, and switching to the one that now does. In the probabilistic form used here, even the right choice sometimes fails, so the hard part is telling an unlucky loss from a real change. One loss proves nothing; several in a row from a card that used to win are worth acting on.
Roshan Cools, Luke Clark, Adrian Owen and Trevor Robbins (2002) scanned 13 healthy volunteers during a probabilistic reversal task. Activity in the right ventrolateral prefrontal cortex rose on the trial when a person stopped choosing the old stimulus and switched, not when a right choice was simply met with a loss. Swainson and colleagues (2000) found reversal deficits in patients with frontal or temporal lobe lesions, and in medicated patients with Parkinson's disease. This game is a recreational version with its own rules, and it gives no diagnosis.
What are win-stay and lose-shift in reversal learning?
Win-stay is picking the same card again after it wins; lose-shift is leaving it after it loses. Lose-shift sounds sensible, but here the lucky card loses 2 times in 10, so leaving after every loss breaks run after run: in the table above it averages 0.9 swaps, against 5.3 for waiting for two losses in a row.
Hanneke den Ouden and colleagues (2013) studied probabilistic reversal learning in nearly 700 people. A variant of the serotonin transporter gene went with more lose-shifting but not with sticking to the old choice after a reversal; a variant of the dopamine transporter gene went with that sticking, and not with lose-shifting.
What can affect your Reversal Learning Game result?
Luck, still: every set of 10 has its 2 losses, but two of them can fall together just as a run is building. Then attention, because a swap is easy to miss after a long, comfortable run, and patience, because a winning streak on the wrong card can feel like a pattern. For rules that change without warning but never lie about right and wrong, try the Card Sorting Test; to see how predictable your own choices are, How Random Are You?
What doesn't the Reversal Learning Game measure?
It isn't a test of intelligence or a clinical assessment. Research versions differ in their win rates, run lengths and number of trials, so a score here can't be set beside one from a study. The points are imaginary, and how you play a game with nothing at stake may not be how you choose when it matters.