“`html
The human mind excels at addressing intricate challenges. One factor contributing to this is that people can dissect problems into manageable subtasks that can be resolved individually.
This enables us to accomplish a daily activity such as going out for coffee by dividing it into steps: exiting our office building, making our way to the coffee shop, and, upon arrival, acquiring the coffee. This approach allows us to navigate hindrances effortlessly. For instance, if the elevator is out of order, we can modify our exit strategy while keeping the other steps unchanged.
Although there is a significant amount of behavioral evidence showcasing humans’ proficiency in handling these complex tasks, it has been challenging to create experimental setups that facilitate an accurate depiction of the computational methodologies we utilize to tackle problems.
In a recent investigation, MIT scholars have effectively modeled how individuals employ various decision-making techniques to navigate a challenging task—specifically, predicting how a ball will traverse a maze when the ball is not visible. The human brain cannot execute this task flawlessly because it cannot simultaneously track all potential paths, but the researchers discovered that individuals can function quite well by adaptable use of two methods known as hierarchical reasoning and counterfactual reasoning.
The researchers were also capable of identifying the scenarios in which individuals opt for each of those methods.
“What people are able to do is segment the maze into sections and then resolve each aspect using relatively straightforward algorithms. Essentially, when we lack the ability to tackle a complex issue, we cope by employing simpler heuristics that achieve the objective,” states Mehrdad Jazayeri, a professor of brain and cognitive sciences, a member of MIT’s McGovern Institute for Brain Research, a researcher at the Howard Hughes Medical Institute, and the senior author of the study.
Mahdi Ramadan PhD ’24 and graduate student Cheng Tang are the leading authors of the paper, which publishes today in Nature Human Behavior. Nicholas Watters PhD ’25 is also a co-author.
Logical strategies
When individuals engage in straightforward tasks that have a clear correct solution, such as classifying items, they perform remarkably well. As tasks grow more complicated, like planning a visit to your favorite cafe, there may no longer be one distinctly superior answer. Additionally, at each phase, numerous obstacles could arise. In these scenarios, individuals excel at formulating a solution that accomplishes the task, even if it isn’t the optimal one.
These solutions often incorporate problem-solving shortcuts, or heuristics. Two notable heuristics that people frequently depend on are hierarchical and counterfactual reasoning. Hierarchical reasoning entails decomposing a problem into layers, beginning with the general and advancing toward the specific. Counterfactual reasoning consists of imagining what might have occurred had you made a different choice. Although these approaches are well recognized, scientists remain uncertain about how the brain determines which one to apply in a specific context.
“This represents a significant question within cognitive science: How do we problem-solve in a suboptimal manner, by devising clever heuristics that we string together in a way that ultimately brings us closer to a solution?” Jazayeri remarks.
To address this, Jazayeri and his colleagues created a task that is just complicated enough to necessitate these strategies while remaining straightforward enough for the outcomes and calculations involved to be quantifiable.
The task requires participants to forecast the course of a ball as it travels through four potential paths in a maze. Once the ball enters the maze, individuals cannot see which route it takes. At two intersections in the maze, they receive an auditory signal when the ball arrives at that position. Anticipating the ball’s trajectory is a challenge that is impossible for humans to solve with complete accuracy.
“It necessitates four simultaneous simulations in your mind, and no human can accomplish that. It’s akin to conducting four conversations at once,” Jazayeri explains. “The task permits us to access this collection of algorithms that humans utilize, as it can’t be optimally solved.”
The researchers enlisted around 150 human volunteers to partake in the study. Before each participant began the ball-tracking task, the researchers assessed how accurately they could estimate durations of several hundred milliseconds, approximately the time it takes for the ball to traverse one arm of the maze.
For each participant, the researchers developed computational models capable of predicting the error patterns that would manifest for that individual (based on their timing accuracy) if they were executing parallel simulations, employing only hierarchical reasoning, only counterfactual reasoning, or various combinations of the two reasoning methods.
The researchers compared the participants’ performance against the models’ predictions and discovered that for every participant, their performance correlated most closely with a model that utilized hierarchical reasoning but occasionally shifted to counterfactual reasoning.
This suggests that instead of monitoring all possible paths that the ball might take, individuals divided the task. Initially, they chose a direction (left or right) where they believed the ball turned at the first junction, and continued to monitor the ball as it approached the next turn. If the timing of the subsequent sound they heard did not align with the path they had selected, they would go back and modify their initial prediction—but only occasionally.
Reverting to the other side, which signifies a move to counterfactual reasoning, requires individuals to recall the sounds they heard. However, it appears that these memories are not always dependable, and the researchers found that individuals decided whether to revert based on their perceived reliability of memory.
“Individuals depend on counterfactuals to the extent that it is advantageous,” Jazayeri notes. “Those who experience a significant performance decline when employing counterfactuals tend to avoid them. However, if you’re someone who excels at recalling information from the recent past, you might revert to the alternate side.”
Human constraints
To further substantiate their findings, the researchers devised a machine-learning neural network and trained it to execute the task. A machine-learning model trained on this task will accurately track the ball’s path and make the correct prediction every time, unless the researchers impose restrictions on its performance.
When the researchers introduced cognitive constraints similar to those encountered by humans, they observed that the model adjusted its strategies. When they restricted the model’s ability to follow all possible paths, it began employing hierarchical and counterfactual strategies akin to those used by humans. If the researchers limited the model’s memory recall capacity, it would switch to hierarchical reasoning only if it believed its recall was adequate to yield the correct answer—mirroring human behavior.
“What we discovered is that networks replicate human behavior when we impose on them those computational limitations that we observe in human functioning,” Jazayeri states. “This truly indicates that humans act rationally within the constraints they must operate under.”
By subtly altering the extent of memory impairment programmed into the models, the researchers also discovered indications that the transition between strategies seems to happen gradually, rather than at a distinct threshold. They are currently conducting additional research to ascertain what transpires in the brain as these strategic shifts take place.
The research was supported by a Lisa K. Yang ICoN Fellowship, a Friends of the McGovern Institute Student Fellowship, a National Science Foundation Graduate Research Fellowship, the Simons Foundation, the Howard Hughes Medical Institute, and the McGovern Institute.
“`