Modeling Human Knowledge of Routes: Partial Knowledge and Individual Variation

Benjamin KuipersAAAI 1983

Abstract

This paper explores the computational structure of the "cognitive map," specifically focusing on how humans store and retrieve knowledge of routes. By formalizing route knowledge as a set of associative links between sensory "views" and motor "actions," Benjamin Kuipers explains common cognitive phenomena, such as the ability to navigate a route without being able to describe it verbally. The model accounts for "partial knowledge" states and individual variations by analyzing the working memory costs associated with different types of spatial associations.

Introduction

Navigating a complex environment like a city is a fundamental cognitive task. Rather than relying on a global "bird's-eye" map, humans often utilize a sequence of local observations and decisions. However, human spatial knowledge is frequently incomplete or asymmetrical—for instance, a person might drive to a destination perfectly but fail to describe the turns in a quiet room. Kuipers seeks to define a representation for "sensorimotor procedures" that accounts for these specific types of partial knowledge and individual variations.

Methods: Formalizing Route Knowledge

Kuipers formalizes the sensorimotor world using two primitives:

  • Views ($V$): Sensory images (visual or otherwise) perceived at a specific location.
  • Actions ($A$): Motor operations (e.g., "turn left") that transition the navigator from one view to another.

Route knowledge is stored via associative links, primarily:

  1. $V \rightarrow A$: "When you see this landmark, perform this action."
  2. $(V, A) \rightarrow V'$: "If you perform this action at this landmark, you will arrive at that landmark."

The paper systematically explores the "space of possible variants" by representing links as triples of integers $<V, A, V'>$. Each position indicates whether an element is a retrieval key (2), a retrieved value (1), or not involved (0). This allows for the categorization of nine potential link types, evaluated based on their ability to support route reconstruction and their associated working memory costs.

Table 1: Taxonomy of link combinations and their associated working memory loads.

Key Insights: The Working Memory Constraint

A central insight of the model is the "Learning Theory" regarding these links. Storing a $V \rightarrow A$ link is computationally "cheap" because both the view and the action are present in working memory simultaneously. Conversely, storing a $(V, A) \rightarrow V'$ link is "expensive," as it requires holding the initial view $V$ in memory while the action $A$ is performed until the new view $V'$ appears.

This explains context-dependent navigation: if a person only learns $V \rightarrow A$ links, they can navigate perfectly in the environment (where the environment provides the view $V$), but they cannot describe the route in a lab because they lack the $(V, A) \rightarrow V'$ links needed to predict the next view.

Key Results and Behavioral Phenotypes

Kuipers identifies several minimal combinations of link types that support route reconstruction. The "standard" model (labeled 1+4) consists of $V \rightarrow A$ and $(V, A) \rightarrow V'$. The model predicts different "behavioral phenotypes" based on which links an individual has acquired.

Table 2: Behavioral consequences and recall patterns for different route reconstruction strategies.

Key results include:

  • Formalization of the "I can't tell you how" effect: Modeled as a state where a navigator possesses $V \rightarrow A$ links but lacks $(V, A) \rightarrow V'$ links.
  • Predictive Power: The model explains why certain link combinations (like $2+6$, involving backward recall) are rare in humans due to high computational overhead, while others are common.
  • Graceful Degradation: Spatial knowledge is shown to be robust; even with missing links, encountering a familiar view can trigger a stored action to resume navigation.

Limitations and Future Work

The model focuses strictly on the sensorimotor level. While these routes must eventually integrate into "topological" (connectivity) and "metrical" (distance/direction) maps, the specific mechanism for this transformation is not detailed here. Additionally, the model treats "views" as atomic units, bypassing the computer vision challenge of mapping noisy sensory input to stable identifiers. Future research could explore the "ecology" of individual variation—why certain navigators favor specific link combinations over others.

Summary updated

BibTeX
@inproceedings{kuipers1983modeling,
  author = {Benjamin Kuipers},
  title = {{Modeling Human Knowledge of Routes: Partial Knowledge and Individual Variation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1983},
}
Figure