At any one time, most people are pursuing multiple goals: answering email, doing the laundry, choosing how to invest retirement savings, helping a child with their math homework. When confronted with such variety, how do we choose which goals we will attend to at any given time? We obviously can't do our laundry if we're on an airplane, but stripping away these kinds of circumstances and examining persistence toward goals in general, what factors lead us to continue to work toward a specific goal, and which make us more likely to switch to a different goal, at least temporarily?
Caltech grad student Sneha Aenugu and John O'Doherty, the Fletcher Jones Professor of Decision Neuroscience, addressed this question in an online gameplay experiment. It was already known among social scientists that people tend to over-persist in pursuit of a long-term goal even when it might be more favorable to switch to another goal. But Aenugu and O'Doherty wanted to quantify this tendency and see how it is affected by unpredictable changes in circumstances.
"There are a lot of things we could be doing at any given instant. How do we decide, 'OK, this is the thing I want to do right now'?" Aenugu, a third-year graduate student in social and decision neuroscience , asks. "Games are a great vehicle for this because throughout gameplay you have to decide what goal you want to pursue at any given moment. Furthermore, in deciding what goals to pursue you have to be sensitive to a changing environment. What goals are your adversaries pursuing? Are they blocking you from achieving your goal, or is the environment in general blocking you at this point? Do you switch over to an alternate goal and then come back to the other when it's more favorable to do so, or do you persist with your initial goal?"
The game Aenugu and O'Doherty designed to test human goal-persistence behavior asked players to collect cards in three suits: cat, hat, and car. Each suit contained two types of cards (for example, the suit "car" has two types of cards, one a key and the other a piece of luggage), and the game rewarded players with points when they completed seven cards in a suit (or in a later iteration of the game, four, six, or eight cards depending on the suit). Gameplay consisted of separate blocks of online play; in each block, the odds of obtaining a card in a particular suit change. In one block, there may be an 80 percent chance of getting a cat card, for example, while in another block, the chance of getting a cat card may be 40 percent or less. In some versions of the game, players were told that the odds would change in different blocks and that one suit would turn up more frequently than the others, but they were not told how much the odds would change. In other versions of the game, players were explicitly told what the odds would be (80, 75, 70, 60, 55, or 50 percent) but they were not told which suits would actually dominate.
The results? As expected, players persisted with collecting cards in a suit they were already collecting even when this was not the optimal choice. However, the over-persistence displayed by individual players differed widely. "Some people are good at putting off immediate rewards and waiting for future rewards, while other people are not," Aenugu says. "Because people tend to prefer immediate rewards, they may want to stick with a particular suit because it's closer to completion-that's when they get points in the game-even if that suit is not favored in a given block of play," Aenugu says.
This over-persistence is the hallmark of a retrospective approach to the decisions required in gameplay-that is, one in which players look back to see how they have progressed in order to decide how to move forward. To contrast the choices of these types of players with optimal gameplay, Aenugu, whose background is in engineering and computer science, created two additional players, both computer algorithms, that played prospectively-basing choices on immediate results and future predictions. One of the algorithms was tailored to adapt to the apparent odds in any given block and select cards from suits that seemed to be performing well regardless of their past performance or if they belonged to a suit that was nearing completion. The second algorithm also played prospectively but, in addition, exercised a preference for completing suits where possible rather than only choosing cards in suits that seemed to be performing best. This type of behavior is called discounting.
Human players persisted more than either of these computer-generated agents, indicating that some factor in addition to discounting is at work among players. Aenugu uses a metaphor drawn from classical physics-momentum-to describe players' tendency to over-persist. Momentum, Aenugu explains, "is a product of progress itself and also the rate of progress. We have mathematically shown that momentum, understood in this way, gives a good approximation of the time it takes players to complete the goal."
Aenugu notes that a calculation of goal-persistence based on momentum is not quite as successful at optimizing performance as the prospective model. "But performing the same calculations looking only at momentum-that is, at current progress and rate of progress-is nearly as good as a prospective model, and it is very cheap to do this computation. That is, it is expensive in terms of the amount of thinking involved to rely on prospective reasoning. Besides, the world is uncertain. You don't know when things are changing. So you can't rely too much on having a perfect prospective model. Sometimes having a cheap model is all you really need."
"It's not like we are doomed to persist," Aenugu adds. "If we know that we have a tendency to over-persist, and if we have more details about the environment in which we are making choices, maybe we can shift our strategies to be more effective. Even in our experiment, giving players instructions about the odds of having a dominant suit in any particular block shifted their behavior."
The researchers plan to examine how these findings about over-persistence, and, in particular, its variability across different players, may have a role to play in computational psychiatry: incorporating neuroscientific computations into clinical psychiatric practice. "We think that understanding the nature of the variation across individuals in how goals are selected might help us gain insight into certain disorders such as depression, anxiety, ADHD, or OCD," O'Doherty says.
The paper, titled " Building momentum: A computational account of persistence toward long-term goals, " was published in the February 2025 issue of PLOS Computational Biology.
