People construct simplified mental representations to plan

General intelligence relies on the capacity to flexibly and efficiently plan ahead. Our theory of problem simplification sheds new light on the computational principles that underlie this remarkable ability in humans.

Like Comment
Read the paper

The Puzzle of Human Planning

Much of human agency, creativity, and intelligence relies on planning, the ability to reason about different courses of action and choose accordingly. One enduring mystery about human planning is how it can be so flexible and general, despite our having limited time, memory, and attention. Even simple tasks—such as planning a trip to the grocery store—involve considering a near infinite range of possibilities, and yet people routinely solve such problems quickly and effectively. Understanding how people plan will shed light on the fundamental mechanisms underlying human cognition and also provide insight into how to design planning algorithms that instantiate similar principles.

A Theory of Problem Simplification and Planning

Herbert Simon and Allen Newell at a desk playing chess

In the 60s and 70s, Herbert Simon (left) and Allen Newell (right) developed a theory of planning as search over a problem representation. Their work raised a new question that has long puzzled psychologists and computer scientists—how do people form problem representations?

Our recent paper, now published in Nature, proposes a new approach to studying planning. Starting with the work of Newell and Simon (1972), planning has been conceptualized as a two-stage process: First, a person represents a problem; then, they perform computations over that representation to identify a good plan. For example, in chess, a complete representation of the game includes the board, the different pieces, the rules for how pieces move and how to win, etc. Computing a plan could then involve simulating moves, counter-moves, and outcomes to find a promising sequence of actions to take. Work on planning generally focuses on the second stage of simulating action sequences—for example, studying the heuristics people use to simulate chess moves. In contrast, our proposal addresses a long-standing question about the first stage: How do people form representations of problems?

To shed light on this question, my colleagues and I proposed that people form temporary, simplified models of problems called task construals. If you are a psychologist or philosopher, you may already be familiar with the broader concept of construal. For instance, depending on how you focus your attention, you can construe the image below as a duck or a rabbit. We hypothesized that people can also construe problems in different ways and that they will form construals that facilitate efficient planning. We call this process of strategically conceiving and perceiving problems value-guided construal.

An image that can be interpreted as a rabbit facing right or a duck facing left.
An ambiguous image that can be construed as a rabbit—or a duck.

You can think of value-guided construals in analogy to the difference between photographs and maps. For example, imagine that you’re planning a cross-country trip. You wouldn’t plan your trip using a high-resolution satellite photograph since it would include too many irrelevant details (trees, houses, cars, etc.). What you really want is a map that includes relevant details and excludes irrelevant ones. Of course, relevance depends on other factors—if you’re biking, you need a bike map; if you’re driving, you need a road map. Thus, our account of value-guided construal aims to explain how the human brain flexibly constructs simplified “cognitive maps” that are appropriate for one’s current goals and circumstances.

Maps, unlike photographs, represent a subset of details of the world that are relevant for some context and purpose. Imagery © 2022 Google, Map data 2022.


Studying Value-Guided Construals

To test whether people form value-guided construals, we examined how they navigated 2D mazes and assessed their construals of the mazes using a variety of measures (recall accuracy, confidence, self-reported awareness, and process tracing). Remarkably, across all experimental measures and analyses, we found robust evidence that people formed the value-guided construals predicted by our account.

A cartoon image of a person looking at a maze composed of obstacles and a thought bubble with a construal of the maze in which some obstacles are greyed out and a planned path in the construed version of the maze.
In our experiments, participants planned routes through mazes composed of blue obstacles. We hypothesized that people would plan by forming construals of the mazes that included relevant obstacles (the blue ones in the thought bubble) and ignored irrelevant obstacles (greyed-out obstacles in the thought bubble).

To preview how value-guided construal makes its predictions, consider the following maze where I have labeled three obstacles and am showing a shortest path as a dotted line:

A grid with obstacles and a path from a start state to a goal state. Three of the obstacles have letters beside them.

Value-guided construal tells us which obstacles are relevant to reaching the goal. Put another way, it answers the following question: Which obstacles, if they were removed, would change the best route? For example, if we removed obstacle A, then rather than taking a path through the upper left of the maze, you might instead take a path through the lower right now that it is unblocked. In other words, the presence of obstacle A changes the best route. In contrast, if we removed obstacles B or C, the best route to take would not change at all. Thus, obstacle A would be included in your construal, whereas B and C would not be. 

I have just described a process of dynamically adding and removing problem details while also testing which ones change the best actions to take—this idea is central to our computational theory. Specifically, we propose that people are both computing construals (that is, adding/removing details from their model of a problem) while also computing plans (for example, when testing whether the best actions change). As we discuss in the paper, there are a number of different ways to implement this kind of computation, but the ultimate purpose of this process is to find a construal that is both useful and simple in the sense that it only includes those details necessary to solve the problem at hand. Our account predicts which construals will be useful and simple, while our experimental results show that people form construals that are useful and simple.

A box and arrow diagram. In the center is a large yellow box labeled "Decision-Maker". To the left is a smaller white box that says "Task" and points to a box inside the yellow box that says "Task Construal". Inside of the "Task Construal" box is a smaller box that says "Plan" and points to a box that says "Action" to the right of the big yellow box.
According to our account, people compute both construals and plans to guide their actions.

Closing thoughts

We are excited about this work as it allows us to draw connections between classic approaches in psychology, neuroscience, and computer science. In particular, value-guided construal combines ideas about cognitive control, attention, and structured representations while highlighting new questions about their interaction—for example, how does the need to manipulate construals shape how we learn more basic representations of the world, such as object concepts? Additionally, by developing a general, computational account of the construal process, we can systematically study its role in other domains in psychology (for instance, in insight problem solving or social interaction) as well as apply these ideas to fields such as artificial intelligence and human-computer interaction.

References

Newell, A. & Simon, H. A. Human Problem Solving (Prentice Hall, 1972).

Mark Ho

Postdoctoral Researcher, Princeton University