Maze Learning

Abstract

Maze learning is a type of spatial learning, and one of the oldest experimental paradigms in learning research: the process by which an animal acquires the correct route to, or location of, a goal within a maze in order to obtain reinforcement or escape. Introduced to psychology when Small adapted the Hampton Court hedge maze for the laboratory rat, it became the workhorse method through which learning was quantified as the fall in errors and time across trials. Maze studies furnished the evidence for latent learning and the cognitive map, and modern variants — the radial-arm maze and the water maze foremost among them — remain the standard assays of rodent spatial memory and its hippocampal substrate. This article covers the paradigm's origin, its major forms, the theoretical debates it settled, and how learning in a maze is measured.

Keywords: maze learning, cognitive map, latent learning, water maze, radial-arm maze

Maze learning is where the experimental psychology of learning began to take a measurable form. A maze reduces the problem of learning to a clean question — can the animal reach the goal, and how efficiently — and it yields two numbers, the errors committed and the time taken, whose decline across trials is a direct index of what has been learned. From Small's first rats in 1901 the apparatus spread through comparative psychology so thoroughly that for the first half of the twentieth century the study of animal learning was, in large part, the study of maze behavior (Small, 1901). The paradigm's importance is not merely historical: the questions it forced — whether a rat learns a sequence of movements or a map of space, whether learning requires reward — produced the concepts of latent learning and the cognitive map, and its descendants are the instruments on which contemporary spatial neuroscience depends (Tolman, 1948; Morris, 1984).

What Maze Learning Is

Maze learning is the process by which an organism, through repeated exposure, comes to traverse a maze efficiently — reaching a goal box, a hidden platform, or a baited arm while committing progressively fewer errors. A maze is any apparatus that offers choices between correct and incorrect paths, from a simple T with a single decision point to the multiple-alley labyrinth Small modeled on the Hampton Court hedge maze (Small, 1901). What unites them is that the animal must learn something about the structure of the space, and that its learning is legible in behavior: an error is an entry into a blind alley or a turn away from the goal, and the trial-by-trial fall in errors and in traversal time is the learning curve.

The paradigm is defined by its method rather than by a single underlying process, and this is the source of both its power and its difficulty. Performance in a maze can improve for several reasons — the animal may be building an allocentric map of the room, memorizing a chain of body-turns, learning to swim more efficiently, or simply becoming less fearful of the apparatus — and a well-designed maze study is one that arranges conditions so that only the process of interest can account for the improvement (D'Hooge & De Deyn, 2001). Because maze learning is the experimental instantiation of spatial learning, the central theoretical question it raises is exactly the one spatial learning poses: does the animal acquire a flexible representation of where the goal is, or a rigid habit of how it once got there?

That maze behavior can express a maplike knowledge of space, rather than only a memorized route, was the conclusion Tolman drew from experiments in which rats behaved, when a familiar maze was altered, as though consulting a representation of its layout (Tolman, 1948). The maze thus became not merely a tool for measuring how fast an animal learns but an arena for asking what it learns.

Types of Maze Learning

In the MeSH classification, maze learning sits directly beneath spatial learning and is subdivided by the specific apparatus used to elicit it. The three narrower descriptors below name standardized behavioral tests rather than distinct kinds of learning; each is a particular maze whose geometry emphasizes a different facet of behavior, and the same animal in any of them may deploy either a place or a response strategy. MeSH is an indexing vocabulary for the biomedical literature, so the list reflects how studies are catalogued rather than a theory of how maze behavior decomposes; a single study frequently uses more than one of these tests.

Table 1

Direct Subtypes of Maze Learning in the MeSH Classification (tree F02.463.425.874.500)

Note. The three descriptors are the narrower terms of maze learning in the current MeSH tree; none is yet a separate article and each is shown here as plain text. Two of them (the elevated plus maze and open field) principally index emotional and exploratory behavior rather than spatial learning, illustrating that subtypes in an indexing vocabulary need not be mutually exclusive or share a common process.

The Classic Maze Paradigms

The earliest mazes were complex labyrinths meant to mimic a natural foraging problem. Small's apparatus, and the many multiple-alley mazes that followed, presented a rat with a long series of choice points between the start and a food goal, and learning was scored as the gradual elimination of entries into blind alleys (Small, 1901). These mazes were rich but analytically blunt: because they confounded many decisions, they were well suited to measuring whether an animal learned but poorly suited to asking what it learned. The history of the paradigm is in large part a progressive simplification of the apparatus so that a single, interpretable question could be posed.

The decisive experiments came from Tolman's Berkeley laboratory and turned on what a maze-running rat actually acquires. In the latent-learning design, rats allowed to explore a maze without reward showed little apparent learning — until food was introduced, whereupon their error scores dropped abruptly to the level of animals that had been rewarded throughout, revealing that the unrewarded exploration had built a usable representation all along (Tolman & Honzik, 1930). The finding was a direct challenge to the view that reinforcement is necessary for learning, and it forced a distinction between learning and performance: knowledge can be acquired without reward and expressed only when there is reason to. Tolman synthesized this and related results into the claim that the rat learns a cognitive map of the maze — an internal representation of its spatial layout — rather than a chain of stimulus-response connections (Tolman, 1948).

What the Maze Reveals About the Brain

The maze was also the tool of the first sustained attempt to localize memory in the brain. Lashley trained rats on mazes, removed varying amounts of cortex, and found that the learning deficit scaled with the quantity of tissue destroyed rather than its location — results he summarized as the principles of mass action and equipotentiality, and which led him to doubt that a memory has a fixed cortical seat (Lashley, 1929). That negative conclusion framed the search that followed: the maplike learning the maze demands turned out not to be diffuse across the cortex but to depend on a specific structure, and the evidence came from a new generation of recording and lesion studies.

The theoretical claims born in the maze acquired a neural substrate when single-unit recording became possible. O'Keefe and Dostrovsky found that individual hippocampal neurons fire only when a rat occupies a particular part of its environment — place cells — and proposed that their population constitutes the physical cognitive map that maze behavior had implied (O'Keefe & Dostrovsky, 1971; O'Keefe & Nadel, 1978). The maze thereby became the behavioral bridge between an abstract construct and recordable cells: the map Tolman inferred from error curves could now be observed as patterned firing while the animal solved the task.

That the hippocampus is necessary for one kind of maze learning, and not merely active during it, was established with the water maze. Morris and colleagues showed that rats with hippocampal lesions were severely impaired at finding a platform hidden beneath opaque water, whose position could be recovered only from distal cues, yet learned normally to swim to a platform made visible — isolating the maplike component of maze learning as the hippocampus-dependent one (Morris et al., 1982). What stores that learning is activity-dependent synaptic plasticity: blocking the hippocampal NMDA receptor, which gates long-term potentiation, prevented rats from learning the hidden platform while leaving cued learning intact, tying acquisition of a new spatial map to the same plasticity mechanism implicated in memory generally (Morris et al., 1986).

These maze-derived findings generalize across species. The same hippocampal-entorhinal machinery that supports place learning in the water maze underlies human spatial memory, and the representation of space it builds is now understood as a general format for structured knowledge rather than a navigation-specific device (Grieves & Jeffery, 2017).

Measuring Learning in the Maze

Two modern mazes dominate because each isolates a process the classic labyrinth confounded. The radial-arm maze separates two kinds of spatial memory within one apparatus: arms radiate from a central platform, each baited once, so that an efficient forager visits each arm exactly once. Re-entering an already-emptied arm is a working-memory error — forgetting where one has been within the current trial — whereas entering an arm that is never baited is a reference-memory error, a failure to learn the stable rule across trials. Olton and Samuelson used this dissociation to show that rats hold an accurate, list-like memory for places already visited (Olton & Samuelson, 1976).

The Morris water maze measures place learning through the falling escape latency to a hidden platform and, on a probe trial with the platform removed, through the proportion of time spent searching its former location. Because the platform is invisible and the water erases local cues, the task cannot be solved non-spatially, which made it the standard assay of hippocampus-dependent learning (Morris, 1984). Its very ubiquity created a methodological problem: small differences in pool size, cue arrangement, and trial spacing produce large differences in results, which is why standardized protocols were developed to make water-maze scores comparable across laboratories (Vorhees & Williams, 2006). The task has since been adapted well beyond basic research — for example, as a sensitive behavioral endpoint in translational models of traumatic brain injury, where the shape of the latency curve indexes the severity of an acquired spatial-memory deficit (Tucker et al., 2018).

Worked Example

Consider how errors decline as a rat learns a maze. The canonical learning curve is negatively accelerated and well described by the decaying exponential E(n) = E_∞ + (E₀ − E_∞)·e^−n/τ, where n is the trial number, E₀ the initial error count, E_∞ the asymptotic floor of irreducible errors, and τ the learning time constant in trials.

Suppose a rat begins at E₀ = 18 errors, has a floor of E_∞ = 2 errors, and learns with τ = 4 trials. Each trial multiplies the remaining excess above the floor by e^−1/4 = 0.7788, so the error count falls 18 → 14.46 → 11.70 → 9.56 → 7.89 across the first four trials. To find when the rat first reaches a criterion of E ≤ 4 errors, solve 2 + 16·e^−n/4 ≤ 4, i.e. e^−n/4 ≤ 2/16 = 0.125, giving n ≥ −4·ln(0.125) = 8.32. The criterion is therefore first met on trial 9: at trial 8 the count is still 4.17 errors, just above criterion, and at trial 9 it is 3.69.

The time constant governs the whole trajectory. A faster learner with τ = 2 reaches the same 4-error criterion by trial 5 — the count is 4.17 at trial 4 and first drops below criterion, to 3.31, at trial 5 — whereas a hippocampally impaired animal with a much larger τ may barely separate from its starting error rate within a session, the pattern lesion studies report. This is why the shape of the curve, summarized by τ, rather than performance on any single trial, is the measure of maze learning (Morris, 1984; Morris et al., 1982).

Discussion

Maze learning holds a singular place in the history of psychology because it made learning measurable and, in doing so, framed the questions the field would spend a century answering. The apparatus turned a vague notion — that an animal “learns” — into a falling curve of errors and latencies, and the effort to interpret that curve produced the discipline's enduring distinctions: between learning and performance, exposed by latent learning; and between knowing where and knowing how, exposed by the contest between place and response strategies (Tolman & Honzik, 1930; Tolman, 1948). The maze was thus not just a method but the setting in which cognitive and stimulus-response accounts of learning were joined in direct empirical combat.

Its interpretive difficulty is the counterpart of its power. Because a maze score aggregates motivation, locomotion, emotionality, and memory, no single number means anything without the controls that hold the other factors constant — the visible-platform control in the water maze, the never-baited arms of the radial maze, the open-field baseline for activity (Morris et al., 1982; Olton & Samuelson, 1976). The modern refinement of maze methodology is essentially the story of building those controls into the apparatus itself, so that the falling curve can be read as learning of a specified kind. That discipline is what allows a hundred-year-old paradigm to remain the front-line assay of spatial memory and its neural mechanism.

Current Directions

Contemporary work pulls maze learning in two directions at once. The first is standardization and translation: because water-maze and radial-maze scores are notoriously sensitive to procedural detail, effort has gone into protocols that make results reproducible across laboratories and into adapting the tasks as quantitative behavioral endpoints in disease models, where a deficit's magnitude and time course carry clinical meaning (Vorhees & Williams, 2006; Tucker et al., 2018). Here the maze is a precision instrument, and the research question is how to read it cleanly.

The second direction is theoretical and reaches back to the paradigm's founding question. The cognitive map that maze behavior first revealed is now understood to extend beyond physical space: the hippocampal-entorhinal system that codes maze location also organizes conceptual and relational knowledge, and a lively debate concerns whether that representation is best described as a metric, Euclidean map or as a graph of remembered places and the transitions between them (Grieves & Jeffery, 2017; Peer et al., 2021). The distinction matters for maze research because a graph-based representation predicts systematic distortions in how animals and humans judge distance and direction, and it reframes the maze not as a scale model of a room but as a network of learned routes and choice points — closer, in fact, to how the earliest labyrinth learners may have solved Small's maze.

Common Misconceptions

A rat learns a maze by memorizing a fixed sequence of turns.

That is only the egocentric, response-based way to solve a maze. Tolman's latent-learning and place-learning experiments showed that rats also acquire a maplike representation of the maze's layout, enabling them to take novel shortcuts and to exploit knowledge gained without reward (Tolman & Honzik, 1930; Tolman, 1948).

A faster escape or fewer errors always means better spatial learning.

Maze performance also reflects motivation, swimming or running ability, and fear of the apparatus. A shorter latency can arise from faster swimming rather than better place memory, which is why controls such as the visible-platform condition are essential to attribute a change to learning specifically (Morris et al., 1982; Vorhees & Williams, 2006).

All mazes measure the same thing.

Different mazes isolate different processes. The radial-arm maze separates working from reference memory, the water maze isolates hippocampus-dependent place learning, and the elevated plus maze indexes anxiety-like behavior rather than spatial learning at all, despite being classified as a maze (Olton & Samuelson, 1976; D'Hooge & De Deyn, 2001).

Glossary

Blind alley.

A dead-end path in a maze; entry into one is scored as an error, and the elimination of blind-alley entries across trials is the classic measure of maze learning.

Cognitive map.

An internal, maplike representation of spatial layout that supports flexible navigation; inferred by Tolman from the maze behavior of rats and later localized to the hippocampus.

Escape latency.

The time taken to reach the hidden platform in the water maze; its decline across trials is the primary index of place learning.

Hippocampus.

A medial temporal lobe structure required for hippocampus-dependent maze learning, such as finding a hidden platform from distal cues; it houses place cells.

Latent learning.

Learning that occurs without reinforcement and remains unexpressed until an incentive is introduced; demonstrated in mazes by Tolman and Honzik and taken as evidence that learning and performance are distinct.

Learning curve.

The trial-by-trial trajectory of errors or latency; in maze learning it is negatively accelerated, falling fastest early and flattening toward a floor.

Maze.

An apparatus offering choices between correct and incorrect paths to a goal; forms range from the simple T-maze to multiple-alley labyrinths, the radial-arm maze, and the water maze.

Morris water maze.

A pool with a submerged platform hidden from view and located from distal cues; the standard assay of place learning and its hippocampal dependence.

Place strategy.

Solving a maze by reference to an allocentric map anchored to distal cues, so that the goal is found from any start point; contrasted with a response strategy.

Probe trial.

A test trial in the water maze with the platform removed, on which the proportion of time spent searching its former location indexes the strength of the learned spatial memory.

Radial-arm maze.

An apparatus of arms radiating from a center, each baited once, that dissociates working-memory errors (re-entering a visited arm) from reference-memory errors.

Reference memory.

Memory for the stable features of a task that hold across trials, such as which maze arms are ever baited; dissociable from trial-specific working memory.

Response strategy.

Solving a maze by repeating a learned body-turn or movement sequence rather than by consulting a map; the egocentric alternative to a place strategy.

Spatial learning.

The acquisition of knowledge about locations and spatial relations; maze learning is its principal experimental paradigm and MeSH subtype.

Working memory (spatial).

Memory for locations visited within the current trial, such as which maze arms have already been entered; dissociable from cross-trial reference memory.

Key Researchers

Willard S. Small (1870-1943). Clark University; he introduced the maze into experimental psychology in 1901, adapting the Hampton Court hedge-maze design to study learning in the laboratory rat and founding the paradigm. Wikipedia

Edward C. Tolman (1886-1959). University of California, Berkeley; his latent-learning and place-learning experiments in mazes produced the concept of the cognitive map, the argument that maze-running rats learn spatial relations rather than chains of responses. Wikipedia

Richard G. M. Morris. University of Edinburgh; he devised the Morris water maze in 1984, the dominant assay of rodent spatial learning, and used it to show that hippocampal lesions and NMDA-receptor blockade impair place navigation while sparing cued approach. ORCID - Wikipedia

John O'Keefe. University College London; he discovered hippocampal place cells in 1971, giving the cognitive map that maze behavior had implied a recordable neural substrate, and shared the 2014 Nobel Prize in Physiology or Medicine. ORCID - Wikipedia

Charles V. Vorhees. Cincinnati Children's Hospital Medical Center and the University of Cincinnati; his standardized Morris water maze protocol became the field's methodological reference for assessing spatial learning and memory across laboratories. Faculty page

Frequently Asked Questions

What is maze learning?
Maze learning is the process by which an animal acquires the correct route to, or location of, a goal within a maze, measured by the fall in errors and time across trials. It is the founding experimental paradigm of animal learning research, introduced by Small in 1901 (Small, 1901).

Who invented the maze in psychology?
Willard S. Small introduced the maze to experimental psychology in 1901 at Clark University, adapting the design of the Hampton Court hedge maze to study learning in the laboratory rat (Small, 1901).

What is latent learning?
Latent learning is learning that occurs without reinforcement and stays hidden until an incentive appears. Tolman and Honzik showed that rats allowed to explore a maze unrewarded dropped abruptly to rewarded-group performance once food was introduced, revealing knowledge acquired all along (Tolman & Honzik, 1930).

How is learning measured in a maze?
Learning is read from the decline across trials in errors — wrong turns or blind-alley entries — and in the time or latency to reach the goal. In the water maze, escape latency and time spent searching the correct location on a probe trial are the standard measures (Morris, 1984).

What is the difference between the radial-arm maze and the water maze?
The radial-arm maze dissociates working-memory errors (re-entering a visited arm) from reference-memory errors, while the Morris water maze isolates hippocampus-dependent place learning by requiring escape to a platform hidden beneath opaque water (Olton & Samuelson, 1976; Morris, 1984).

Why is the hippocampus important for maze learning?
The hippocampus is required for the maplike, place-based form of maze learning. Rats with hippocampal lesions cannot find a platform defined by distal cues yet swim directly to a visible one, and NMDA-receptor blockade in the hippocampus prevents new place learning (Morris et al., 1982; Morris et al., 1986).

Does maze performance measure only memory?
No. A maze score also reflects motivation, locomotor and swimming ability, and emotionality, so controls such as the visible-platform condition and the open-field baseline are needed to attribute a change specifically to learning (D'Hooge & De Deyn, 2001; Vorhees & Williams, 2006).

Is maze learning still used in modern neuroscience?
Yes. Standardized water-maze and radial-maze protocols remain front-line assays of spatial memory and are widely used as behavioral endpoints in disease models, including translational research on traumatic brain injury (Vorhees & Williams, 2006; Tucker et al., 2018).

References