PhD: Orienting Schemata in Wayfinding

Abstract (verbatim)

The purpose of the following studies is to assess the development of generic spatial knowledge (i.e., orienting schemata) and its impact on wayfinding in novel large-scale built environments. Although contemporary cognitive mapping theories have focused on explaining how cognitive maps are acquired and utilized within a given environment, very little effort has been placed in assessing the influence of generic spatial knowledge (i.e. orienting schemata) on the cognitive mapping process and its relation to wayfinding. The current set of studies addresses several important issues about orienting schemata including how generic cognitive maps are represented in memory, the process by which this information is acquired, and their influence on wayfinding behavior. To assess these issues, a new paradigm was introduced in which participants explored a series of virtual environments with the goal of finding a target in each environment. The virtual environments presented to participants were unique in that they were built according to a set of design rules, which defined what features (i.e., landmarks) made up an environment as well as the spatial relations (i.e., topological relations) between those different features. These rules were based on both artificially determined construction rules and on floor plans of real homes. Participants' performance was measured by counting the cumulative number of places visited before finding the target. The findings from these studies indicate that through repeated exploration, participants successfully abstracted the invariant relations from multiple instances of a particular type of environmental setting. Importantly, the acquisition of such abstract knowledge structures facilitated subsequent searches in environments that conformed to these rules of design. Participants who explored environments in which systematic violations were introduced such as changing the identities of objects or changing the topological relations between objects in rooms resulted in a significant increase in the number of visited before finding the target. The results from these studies provide valuable insight into how orienting schemata develop and their subsequent influence on spatial behavior (i.e., wayfinding) in novel large-scale built environments.

The gap

Cognitive mapping research had a blind spot for decades. Theories from Siegel and White (1975) onward described spatial learning as an incremental, bottom-up process — the wayfinder enters a new place with no relevant prior knowledge, and gradually accumulates landmark knowledge → route knowledge → survey or configurational knowledge through direct exploration. The accuracy and completeness of a wayfinder's spatial knowledge was assumed to be entirely determined by experience with that particular place (O'Neill, 1991a).

But that description is not entirely complete. People bring with them, from prior experiences, a rich body of background knowledge about the structure of built environments which they utilize in the exploration and learning of new places they have never previously explored. We walk into a supermarket we've never been in and know the dairy is along the back wall — produce and dairy along the side and back walls, non-perishable items central in the store. We enter a stranger's house and find the bathroom without asking. Real-world places are highly structured: objects and their spatial relations co-vary in predictable, invariant ways across instances of the same kind of environment (Chun and Jiang, 1999).

Although researchers had suggested spatial knowledge can be represented as schemata (e.g., Evans, 1980; Kuipers, 1978), ambiguity existed over what "schema" meant in this context. Kuipers's (1978) usage referred to low-level sensorimotor production rules — [view, action, view]. The dissertation used the term to mean stereotypical or generic spatial knowledge that people possess for a set of similar environments. Very little empirical work (Gross and Zimring, 1990, 1992) had assessed the exact role that such schemata play in the acquisition, representation, and utilization of environmental information.

The question

Can people abstract structural regularity across multiple instances of an environment type? And once abstracted, does that generic knowledge measurably improve wayfinding in novel environments built from the same regularity?

The work used the term orienting schemata (after Neisser, 1976; Scholl, 1987) for these abstract spatial representations — cognitive structures specialized in directing perceptual and motoric exploration of the environment, distinct from the cognitive map of any specific place. Three studies were conducted to address three claims in sequence:

1. Orienting schemata can be acquired through induction, without explicit instruction.
2. They encode both spatial structure (topology) and object identity — not just one or the other.
3. The acquisition mechanism is not an artifact of artificially-constructed deterministic rules; it operates on the probabilistic statistical regularity that real-world architecture actually exhibits.

Schema theory the dissertation rests on

Five properties of schemata, drawn from the classical literature (Bartlett, 1932; Minsky, 1975; Shank and Abelson, 1977; Mandler, 1984; Brewer and Nakamura, 1984; Rumelhart, Smolensky, McClelland, and Hinton, 1986):

1. Molar. Schemata are large, structured knowledge units — not built up from smaller atoms each time.
2. Pre-existing. Formed from past experience and retrieved at the time of need, rather than constructed dynamically (which is what mental models do).
3. Hierarchically organized. A "toaster" schema sits inside "kitchen" inside "house"; higher nodes encode more general categories than lower nodes.
4. Generic. Schemata represent stereotypical information, abstracted across episodes — not tied to any single time-and-place experience.
5. Variable slots with defaults. Some elements are fixed (a room has a ceiling). Others are slots filled with optional values within a restricted range (a bedroom's "bed" slot defaults to a mattress + box spring; a "furniture" slot accepts chair, desk, table — but not "car").

The dissertation also acknowledged the criticism of classical schema theory (Alba and Hasher, 1983; Rumelhart et al., 1986) — that fixed scripts can't handle the diverse, dynamic, spontaneous situations humans actually encounter — and accepted the parallel-distributed-processing (PDP) reinterpretation: schemata don't explicitly exist but emerge from the interaction of simpler elements working in concert, with schema coherence dependent on the strength of connection weights among units in a network. This reinterpretation preserved the useful properties of schemata (generic, generative, abstract) while addressing the rigidity problem.

The paradigm

Three sets of virtual environments, built in a custom VR system. Each environment was a multi-room layout the subject explored in first-person view via mouse and keyboard. The target was always a flag in one specific room. The dependent measure was the cumulative number of rooms visited before finding the target — a behavioral, ecologically grounded measure of search efficiency, rather than the room-label recall or sketch-map tasks that prior schema research (Arbuckle et al., 1994; Tversky and Hemenway, 1983) had used.

What made the design work was the generative grammar: each VE was built from an explicit set of construction rules — which room types could connect to which, which objects appeared in which room types, and how the layout was permitted to vary. The same rules generated many different specific environments. Subjects saw a sequence of them and (in the right conditions) gradually learned the underlying structure even though no two environments were ever identical. Subjects were never told there was a rule.

Wayfinding difficulty across stimuli was equated using O'Neill's (1991b) ICD (InterConnection Density) plan-complexity measure, so observed differences in search performance couldn't be attributed to one environment being intrinsically harder than another.

Study 1 — Acquisition

Design. 16 unique VEs generated from a single deterministic rule set, presented in 8 blocks of 2 VEs each. 161 3D objects in the pool. Start position varied within each block: subjects first appeared 3 rooms from the target, then 4 rooms from the target. Presentation order was counter-balanced. Practice VE was a 5-room cross-shaped layout for control familiarization. 5-minute time limit per trial; data from any subject who failed to reach the target inside the limit was excluded (2 of N trials).

Method. Two-way repeated-measures ANOVA on cumulative rooms-visited, with within-subjects factors of Presentation Order (Block 1 through 8) and Start Position (3 or 4 rooms). Greenhouse-Geisser correction applied for sphericity violations.

Result. Only the Presentation Order main effect was significant: F(2.87, 37.3) = 5.20, p = .005. Neither Start Position, F(1, 13) = .383, p = .547, nor the interaction, F(1.83, 23.8) = 1.62, p = .219, mattered. Holm-Bonferroni-corrected paired t-tests confirmed Block 1 vs. Block 5: t(15) = 3.78, p = .002; Block 1 vs. Block 8: t(15) = 2.79, p = .014; Block 5 vs. Block 8 not significantly different, t(15) = .358, p = .725.

Translated: rooms-visited dropped from ~12 in Block 1 to ~6 by Block 5 and stayed there. Schema acquisition asymptoted at about 5 blocks (10 trials). Learning was implicit — subjects in debriefing said they "got the hang of it" but couldn't articulate the structure they'd internalized.

Why the Start Position null result matters. The original intent of the variable was to add more potential start locations (violin, helicopter, light bulb, or clock room) so subjects wouldn't memorize a single starting view. With only a one-room difference between conditions, the absence of an effect was unsurprising in hindsight — but it ruled out a trivial "subjects just memorized the starting room" explanation.

Study 2 — Content and Organization

Question. Did Study 1's improvement reflect a genuine orienting schema, or just practice using mouse-and-keyboard navigation in any VE? And if a schema was being formed, what did it encode — spatial topology, object identity, or both?

Design. 48 undergraduates (24 female, 24 male) from UMN Intro Psych for course credit. 6 subjects withdrew with motion sickness and were replaced. Training session: 10 schema-consistent (S-C) VEs from the same rule set as Study 1. Test session immediately followed, with no forewarning. Test environments came in four flavors:

• S-C (Schema-Consistent): same rule set, same objects.
• S-IC (Change Pattern): same objects, but topological connections between rooms randomly re-arranged. Four such layouts generated, each with a different pattern of room-room connections. ICD complexity matched the S-C set.
• S-IC (Change Objects): same topology, but the conjunction of objects defining a room category replaced with a randomly drawn set from the same 161-object pool. Snodgrass and Vanderwart (1980) familiarity ratings were matched across S-C and S-IC sets (mean ~3.56 vs ~3.57). Objects sampled without replacement so the same item never appeared twice across settings.
• Control: different topology AND different objects — schema-inapplicable in both dimensions, used to rule out raw practice effects.

Method and result. A two-way mixed ANOVA tested Presentation Order × Environment Type. Training-session learning replicated Study 1: within-subjects effect F(3.84, 176.39) = 16.6, p < .001. In the test session, both topology violations AND object violations produced substantial performance drops back toward Block-1-equivalent search efficiency. The schemas formed in training were brittle to either form of disruption.

What this argued. The orienting schema encodes both the topological skeleton AND the object-content slots that fill it. Either failure was enough to break the schema's predictive power. This was incompatible with simpler accounts that proposed schemas as purely topological abstractions, and consistent with classical schema theory's slot-with-restricted-values machinery.

Study 3 — Real-world generalization

The ecological move. Studies 1 and 2 used artificially generated rule sets where the connections between rooms were strictly fixed. Real environments aren't built from rigid grammars — they vary, but they vary within statistical bounds. If the schema-acquisition mechanism is real and useful, it has to work on probabilistic regularity, not just deterministic regularity.

Method. Sampled real house floor plans, computed a transition probability matrix between room types (treating the floor plan as a Markov chain), and generated VEs that conformed (or didn't) to those statistics. A sample of 100 floor plans was used to compute the underlying distribution; 20 of those houses provided the specific stimuli. Layout typicality of each candidate VE was measured by decomposing it into routes, computing each route's overall Markov probability across the sample, and averaging.

The numbers are striking: across the 100-plan sample, getting from a garage to a front porch admitted 52 unique possible paths — but within the 20-house subset that defined the stimulus typicality, only 12 unique routes were observed. Certain micro-routes were frequent: garage → utility → hallway → foyer → porch appeared 6 times in the 20-house sample; garage → utility → kitchen → living → foyer → porch appeared 2 times; many others appeared exactly once. Even when no single route was strongly determined, room-pair transitions were highly stereotyped.

Training and test. 48 subjects (24F/24M) again. 7 blocks of training in schema-consistent (S-C) VEs, then 3 blocks of test trials in either S-C or S-IC environments. The S-IC condition randomized object locations while holding topology constant — disrupting the learned association between object identity and likely location.

Result. Training-session improvement was clean: room count dropped from ~12 in Block 1 to ~4 by Block 7, asymptoting around Block 4. F-test on the training Presentation Order main effect: F(3.84, 176.39) = 16.6, p < .001. In the test session, only the between-subjects effect of Environment Type was significant: F(1, 46) = 73.27, p < .001. S-C subjects continued at ~4 rooms; S-IC subjects shot up to ~10 rooms.

A direct comparison between the S-IC condition's Block 1 (the start of training, fully naive) and their Block 10 (the test trials after 7 training blocks): t(22) = 1.16, p = .257. Not significant. After all that training, the S-IC subjects searched no better than they had on their first ever trial — when the object/location associations were randomized, the schema gave them no advantage.

Why this study matters most. The deterministic constraints of Studies 1 and 2 were artificial. Study 3 said the schema-acquisition machinery handles probabilistic regularity — and that the size of the wayfinding benefit, when the schema applies, is large enough to be observable in straightforward behavioral measures.

What the orienting schema actually is

Two candidate interpretations sit on the results, and the dissertation was careful to not force a choice between them:

Interpretation A — Abstract topological network. Consistent with Gross and Zimring's (1990, 1992) schema-based cognitive mapping theory and Thinus-Blanc's (1996) Spatial Integration System (SIS). The schema is a generic, abstracted graph: nodes are room types, edges are typical connections, and slots are filled with default objects. Wayfinders extract spatial invariants across exposures and form an allocentric representation that feeds back into action planning. Strongly supported by the topology-violation effect in Study 2 (changing the network broke performance).

Interpretation B — Micro-routes. Subjects didn't form an integrated topological network at all. They learned small two- or three-room combinations — garage-utility, living-foyer-porch — and recombined them dynamically. Under this account, efficient search in a novel environment doesn't require an integrated set of design rules; it just requires the right local fragments to chunk together. Consistent with the fragmentary nature of search paths observed in the protocols, and with the probabilistic regularity that Study 3 found in real houses (most variation is local).

Both interpretations preserve abstraction: in both cases, what's learned generalizes beyond any single environment. The two differ in granularity. The current studies' design couldn't adjudicate between them — distinguishing the two would require testing transfer to entirely novel target objects whose locations weren't part of training, which a follow-up study would handle.

The theoretical contribution

Three claims survive Study 3:

1. Generic spatial knowledge is a real, empirically measurable thing. Prior to this work, schema-based cognitive mapping had been mostly theoretical (Gross and Zimring) or anecdotal. The VE paradigm gave it a quantitative behavioral signature — rooms-visited-before-target, varying as a function of schema applicability.

2. It's acquired by induction. No explicit instruction, no metacognitive verbalization, no slow consolidation. Subjects' search performance asymptoted in 6 to 10 trials and the full study finished in about 30 minutes. This timescale rules out the Complementary Learning System view (McClelland, McNaughton, and O'Reilly, 1995) — under CLS, semantic generalization across episodes requires neocortical consolidation over much longer time spans. The current results look more like rapid hippocampal-style learning of statistical structure (cf. Fiser and Aslin, 2001, 2002 on conditional-probability learning in visual scenes).

3. Cognitive mapping research had been measuring the wrong thing. Studies that put a subject in a single environment and tracked their learning curve necessarily held prior knowledge constant — making schemata invisible by construction. To see schemata, you have to vary the environment across trials in a structured way. That's what the VE paradigm was built to do, and what it succeeded in doing.

Methodological caveats the dissertation flags

The dissertation explicitly identifies limitations and follow-ups, rather than overclaiming:

• The Markov-chain typicality measure used in Study 3 captures only sequential statistical regularity. For spatially-dependent (not time-dependent) information, a Markov Random Field formulation would be more appropriate, and was suggested as the right next-step analytic tool.
• Training blocks presented same-category environments contiguously. Real-world exposure interleaves environment categories across the day. A mixed-category training schedule would test whether the rapid induction observed here is partly an artifact of the blocked design.
• The fixed-target paradigm doesn't distinguish "abstract topological network" from "memorized micro-routes." Testing novel target locations not present in training would.

Personal context

Filed at the University of Minnesota — Twin Cities in May 2004. Defended under Charles R. Fletcher (cognitive psychology / discourse comprehension). Committee members spanned cognitive psychology, vision science (Gordon Legge), spatial cognition (Herb Pick), and computational vision / Bayesian inference (Paul Schrater) — the spread of expertise reflects the dissertation's reach across schema theory, behavioral wayfinding measurement, and statistical-structure learning. Funded by UMN's Center for Cognitive Sciences. Editorial and emotional support, per the acknowledgements, from Sandra Virtue.

After defending, Alan moved out of academia and into industry UX, starting at Motorola Mobility (Razr era) in 2004. The through-line back to the thesis is visible in the work that followed: in-vehicle voice UX at Telenav (Scout Launcher, Voice Assistant, Novo Ride, Ford Navigation) is, in effect, applied wayfinding research — drivers extracting spatial expectations from one interface to apply to the next, in environments far less constrained than house floor plans. The schema-acquisition machinery the dissertation characterized is the same machinery that determines whether a new car's infotainment system feels intuitive on first drive or requires explicit instruction.

• cognitive psychology
• wayfinding
• spatial cognition
• virtual environments
• schema theory
• PhD
• University of Minnesota
• orienting schemata
• Charles Fletcher