summary: New research reveals ways to use human flexible problem-solving strategies, such as hierarchical and counterfactual reasoning, when faced with complex tasks. The researchers asked participants to predict the hidden ball path through the maze and to make real-time decisions based on sound clues and memories.
Since it is cognitively impossible to track all possible paths, people divided the task into step or revised choices if they suspected an error. The computational model confirmed that strategy choice depends on individual memory strength and task requirements.
Important facts:
- Two important strategies: People use hierarchical reasoning to split tasks into steps and counterfactual reasoning to correct previous choices.
- Memory-driven decision: The shift in strategy depends on how trustworthy participants believe in their memories.
- AI Confirmation: Neural networks adopted a similar strategy when faced with memory and tracking limitations.
sauce: mit
The human brain is very good at solving complex problems. One reason for this is that it can break down problems into subtasks that are easy for humans to solve.
This allows you to complete daily tasks, such as drinking coffee, leaving the office building and moving to a coffee shop, and getting your coffee once. This strategy helps you handle obstacles easily. For example, if the elevator is broken, you can fix how you can get out of the building without changing any other steps.
Although there is much behavioral evidence that demonstrates human skills in these complex tasks, it has been difficult to devise experimental scenarios that allow for accurate characterization of the computational strategies used to solve problems.
In a new study, MIT researchers modeled how people deploy a variety of decision-making strategies to solve complex tasks.
Although the human brain cannot perform this task completely, as it is impossible to track every possible trajectory in parallel, the researchers have found that people can perform reasonably well by flexibly adopting two strategies known as hierarchical reasoning and counterfactual reasoning.
Researchers were also able to determine the circumstances in which people would choose each of those strategies.
“What humans can do is break down the maze into subsections and solve each step using a relatively simple algorithm. If you don't have a way to solve complex problems effectively, you can manage it using a simpler heuristic that gets the job done,” Hughes Medical Institute, and senior author of the study.
Mahdi Ramadan PhD '24 and graduate student Cheng Tang are the leading authors of the papers that appear today. Natural human behavior. Nicholas Watters PhD '25 is also co-author.
Rational strategy
It works very well when you perform simple tasks with clear correct answers, such as humans classifying objects. When tasks become more complicated, such as planning a trip to your favorite cafe, there can clearly be another good answer.
And there are many things that don't work out at each step. In these cases, humans are very good at solutions that complete tasks, even though they are not the best solution.
These solutions often include problem-solving shortcuts or heuristics. Two well-known heuristic humans who are generally dependent on are hierarchical and counterfactual reasoning. Hierarchical reasoning is the process of breaking down a problem into layers, starting with general and going towards detail.
Counterfactual reasoning involves imagining what will happen if you make another choice. These strategies are well known, but scientists don't know much about how the brain determines what it uses in a particular situation.
“This is a really big question in cognitive science. How do you solve problems in a suboptimal way by coming up with clever heuristics that connect them in ways that will get you closer and closer to them?” Jazaelli says.
To overcome this, Jazayeri and his colleagues have devised tasks that are complex enough to require these strategies, but are simple enough to measure the outcome and the calculations that enter them.
This task requires predicting the path of the ball as participants move four possible trajectories in the maze. Once the ball enters the maze, people cannot see which paths move.
At two junctions in the maze, you can hear an audible cues as the ball reaches that point. Predicting the ball's pass is a task that humans cannot solve completely accurately.
“It requires four parallel simulations in your mind, and humans cannot do that. It's similar to having four conversations at once,” says Jazayari.
“This task allows us to leverage this set of algorithms that humans use because they are not optimally resolved.”
The researchers recruited approximately 150 human volunteers to participate in the study. Before each subject began the ball tracking task, the researchers assessed how accurately they could estimate a time pan of several hundred milliseconds.
For each participant, the researchers created computational models that could predict the patterns of errors seen (based on timing skills) for a participant when performing parallel simulations using hierarchical inference only, counterfactual inference only, or a combination of two inference strategies.
The researchers compared subjects' performance with model predictions and found that for all subjects, their performance was most closely related to models that used hierarchical inference, but sometimes switched to counterfactual inference.
It suggests that instead of tracking down all possible passes the ball could take, people broke up the task. First, they chose the direction (left or right). There, I thought the ball would spin at the first junction and continued to track it towards the next turn.
If the timing of the next note they hear is not compatible with the path they chose, they will return and correct the initial prediction, but only for each time.
Returning to the other side, representing the transition to reasoning on the other side, we need to see the memory of the tone that people have heard. However, we found that these memories are not always reliable, and researchers found that people decided whether to return based on how good their memories are.
“People rely on counterfactuals to the extent that it helps,” Jazaeli says.
“People who suffer major performance losses when doing counterfactuals avoid them. But if you're really good at getting information from the recent past, you might go back to the other side.”
Human restrictions
To further validate the results, researchers created a machine learning neural network and trained it to complete the task. Machine learning models trained in this task will accurately track ball passes and make correct predictions each time, unless researchers place limits on performance.
When researchers added cognitive limitations similar to those facing humans, they found that the model changed its strategy. When they eliminated the ability to follow all possible trajectories of the model, it began to adopt hierarchical and counterfactual strategies like humans.
If researchers reduced the memory recall capabilities of models, they began switching to hierarchies, just like humans, if they thought the recall was sufficient to get the correct answer.
“What we found is that networks mimic human behavior when they impose computational constraints they find in human behavior,” says Jazayari.
“This really says that humans are acting rationally under the constraints that they have to function.”
By slightly varying the amount of memory impairment programmed into the model, the researchers also found that strategy switching appears to occur gradually rather than a clear cutoff point. They are currently doing further research to try to determine what is happening in the brain when these shifts in strategy occur.
Funding:
This study was funded by the Lisa K. Yang Icon Fellowship, friends of the student fellowship at the McGovern Institute, the National Science Foundation Graduate Research Fellowship, the Simmons Foundation, Howard Hughes Medical Research Institute, and the McGovern Institute.
About this neuroscience research news
author: Sarah McDonnell
sauce: mit
contact: Sarah McDonnell – MIT
image: This image is credited to Neuroscience News
Original research: Open access.
Mehrdad Jazayeri et al. Natural human behavior
Abstract
Computational infrastructure for hierarchical and counterfactual information processing
Humans use hierarchical and counterfactual strategies to solve complex multi-step decision problems.
Here we designed tasks that would ensure these strategies were addressed, and hypothesis-driven experiments were conducted to identify the computational constraints that would generate them.
Three important constraints were found. Parallel processing bottlenecks that facilitate hierarchical analysis, compensatory but capacity-limited counterfactual processes, and working memory noise that reduces counterfactual fidelity.
To test whether these strategies are computationally rational, i.e., optimal, taking into account such constraints, we systematically trained recurrent neural networks under various limitations.
Only recurrent neural networks exposed to all three constraints replicated human-like behavior.
Further analysis revealed that hierarchical, counterfactual, and tictive strategies are typically considered clear, broadened along the continuum of rational adaptation.
These findings suggest that human decision-making strategies emerge from a shared set of computational limitations and may provide a unified framework for understanding human cognitive flexibility and efficiency.
