Spatial Computing

How Spatial Structures Replace Computational

Effort

Christian Freksa

Abstract At the Advanced Study Institute on Cognitive and Linguistic Aspects of

Geographic Space in Las Navas del Marqués in July 1990, I presented a chapter on
Qualitative Spatial Reasoning. In that chapter, I suggested that spatial inference
engines might provide the basis for rather general cognitive capabilities inside and
outside the spatial domain. In the present chapter, I will follow up on this perspective
and I will illustrate the ways in which research in spatial cognition has progressed
towards understanding spatial reasoning and spatial computing in a more literal
sense: using a spatial substrate. The chapter presents a progression of approaches to
spatial reasoning from purely descriptive to increasingly spatially structured. It
demonstrates how spatial structures are capable of replacing computational pro-
cesses. It discusses how these approaches could be developed and implemented in a
way that may help us to better understand higher-level spatial abilities of cognitive
systems that are frequently attributed to the right cerebral hemisphere in humans.
The chapter concludes by discussing the special role of space and time for cognition
and advocates a thorough overall analysis of the specific problem to be solved to
identify the most suitable approach to computation.

Keywords Spatial cognition

Spatial computing
Symbolic computing



Structure of space High-level spatial abilities

C. Freksa (&)
Spatial Cognition Research Center SFB/TR 8, University of Bremen, Bremen, Germany
e-mail: [email protected]

M. Raubal et al. (eds.), Cognitive and Linguistic Aspects of Geographic Space, 23

Lecture Notes in Geoinformation and Cartography, DOI: 10.1007/978-3-642-34359-9_2,
 Springer-Verlag Berlin Heidelberg 2013
24 C. Freksa

1 Spatial Problems

Let us consider examples of common spatial problems we might encounter. The

problems are to be taken literally; i.e., no additional information is provided;
missing information must be added from knowledge or assumptions about the
environment.
1. Given the triangle ABC with the coordinates A = (1, 3), B = (9, 2), C = (6, 8);
is P = (8, 4) inside or outside the triangle ABC?
2. (How) can I get the piano into my living room?
3. How do I get from here to John’s place?
4. Which is closer: from here to John, or to Mary?
5. Is the tree on my property or on your property?

6.

Problem 1 is a classic high school geometry problem which can be solved

abstractly with linear equations; the correct algebraic solution will locate P on the
line BC; numeric solutions may place P inside or outside the triangle, depending
on the number format and algorithm chosen.
Problem 2 is a form of the classic Piano Movers’ Problem in mathematics
(Schwartz and Sharir 1983); although this problem can be represented geometrically,
in practice it is rarely approached mathematically in the abstract representation
domain but by trial and error in the physical problem domain.
Problem 3 cannot very well be presented in geometric terms; a graph structure
that depicts the location ‘here’, John’s place, and a traversable connection between
them is more appropriate and often times preferable to a solution in the physical
domain, particularly if John’s place is far away.
Problem 4 typically does not require the mathematically correct solution, which
may take a long time to determine. A quickly provided estimate tends to be more
helpful, in practice.
Problem 5 is an example where a formal approach alone may not suffice.
Although the boundaries of the properties will be defined in a legal document in
terms of precise geo-coordinates, the real-world correspondence of the legal
boundary may be too expensive to determine; therefore the correct boundary often
is not known. In addition, it may not be clear where the branches and roots of the
tree start and where they end, in formal terms.
Spatial Computing 25

Problem 6 (related to the Piano Movers’ Problem) is not posed in terms of

words or numbers but in terms of spatial objects (resp. an image thereof). It is a
truly spatial problem presented physically to small children who will try to fit the
small colored objects into the openings of the wooden cube and thus learn about
spatial features like size and shape through physical processes by trial and error.
These examples illustrate that spatial problems may come in different modal-
ities: in terms of numbers, language, or spatial configurations; and in different
domains: abstract mathematical or legal and concrete physical space. Likewise, the
solutions to spatial problems may be required in terms of numbers, language, or
spatial configurations. The solution may or may not be needed in the same
modality or domain as the problem statement. A correct solution may not always
be the best solution, as quickly or cheaply available sub-optimal solutions may be
more useful in certain situations. In other words, we may need to transform
problems and solutions between different modalities and domains, and the gen-
eration of a problem solution may take place in a variety of modalities and
domains. Accordingly, it may be helpful to have approaches available that are
tailored to the respective requirements (cf. Sloman 1985).
This observation raises the issue whether we always have to transform spatial
problems into geometric formalisms to enable computational solutions by means
of sequential interpretation of instructions, or whether we can find ways to directly
process entire spatial configurations, as humans seem to be able to do (Shepard and
Metzler 1971). I will dub the classic computer science approach of sequential
interpretation as left-brain computing, as information processing in the left cere-
bral hemisphere is associated with bottom-up or language-like sequential
processing; I will dub the approach of processing entire spatial configuration as
right-brain computing, as the right cerebral hemisphere in humans is associated
with top-down or holistic processing (cf. Kosslyn 1987).
In the present chapter, I will first review foundations of qualitative temporal and
spatial reasoning. I will then discuss the notion of conceptual neighborhood and
how we can exploit this notion for spatial computing. I will introduce tools for
processing qualitative spatial relations, and then address the transition from spatial
relations to spatial configurations. Finally I will demonstrate and analyze the
notion of spatial computing as contrasted to symbolic computing.

2 Qualitative Temporal and Spatial Reasoning

The starting point for much of the research in qualitative temporal and spatial
relations since the late 1980s was the chapter Maintaining knowledge about
temporal intervals by Allen (1983), although the underlying insights had been
published previously (Nicod 1924; Hamblin 1972).
The intriguing result of this research was that thirteen ‘qualitative’ relations
could describe temporal relations between events uniquely and jointly exhaus-
tively (Fig. 1). There was an expectation that the idea of qualitative relations could
26 C. Freksa

Fig. 1 The 13 jointly exhaustive and mutually exclusive qualitative relations between two
temporal intervals

Fig. 2 Upper part (facsimile) of the composition table for the qualitative temporal relations
(without the ‘equals’ relation) from Allen (1983). The relation r1 is composed with the relation r2
to obtain the composite relation found in the table. In most cases, more than one relation may
result from a composition. ‘‘no info’’ means that all 13 relations may result from a given
composition

be extended to one- and higher-dimensional spatial objects that share the

extendedness property of temporal intervals. Initially, researchers had in mind a
single spatial calculus that would compute all-embracing spatial relations between
objects based on information about spatial relations between other objects.
However, it became apparent soon that it would be more effective to develop
specialized calculi that deal with individual aspects of space rather than a com-
prehensive spatial calculus that would integrate multiple aspects of space in a
single formalism. For example, Allen’s interval calculus (see Fig. 2 for the rules of
combining interval relations) can be easily adapted to 1-dimensional oriented
Spatial Computing 27

space (Freksa 1991b; Skiadopoulos and Koubarakis 2004; Liu and Li 2011) or to
three spatial dimensions individually (Guesgen 1989).

3 Conceptual Neighborhood

Interval relations can be described at a finer level of resolution in terms of point

relations, i.e., in terms of relations between the starting points and the ending
points of the intervals. An important feature of physical time and space is that
gradual change in position or size results in small qualitative changes or no
changes at all between the point relations involved. For example, in the transition
from the before relation to the meets relation, only one of the four point relations
between beginnings and endings of the intervals changes: the relation between the
ending of the first interval and the beginning of the second interval changes from
smaller than to equals. Accordingly, spatio-temporal configurations that result
from small physical changes are perceptually and cognitively closely related.
Furthermore, events in close temporal vicinity are related more easily to one
another than events in different epochs. Similarly, nearby spatial locations are
more easily related to one another than locations far apart. This insight is captured
in Tobler’s First Law of Geography: ‘‘Everything is related to everything else, but
near things are more related than distant things’’ (Tobler 1970).
The role of nearness extends from temporal and spatial neighborhood to the
more abstract level of relations: certain relations are closer to one another than
others; in fact, some relations are distinguished only by a single detail. These
relations are called conceptual neighbors (Freksa 1991a). In Fig. 3 the thirteen
interval relations from Fig. 1 are applied to one-dimensional oriented space.
Conceptually neighboring relations are depicted next to each other.

Fig. 3 Left Thirteen qualitative relations for objects (here fishes) in one-dimensional oriented
space. The example classifies positions of objects in the horizontal dimension. The 13 relations
are arranged by conceptual neighborhood. Right The corresponding labels of the qualitative
temporal relations from Allen (1983) depicted in the same spatial arrangement (adapted from
Freksa 1991b)
28 C. Freksa

spatial conceptual

neighborhood between neighborhood between

locations relations

static structure process structure

Fig. 4 Spatial and conceptual neighborhood: The left graph depicts relations between static
spatial locations; directly connected nodes represent spatial neighbors. The right graph depicts
direct transitions between spatial relations due to processes in the domain; edges correspond to
conceptually neighboring relations caused by a minimal spatial change in the domain

The notions of conceptual and spatial neighborhood are closely related:

Whereas two directly connected objects are called spatial neighbors, two by
minimal differences directly connected relations are called conceptual neighbors
(Fig. 4).
Arranging temporal and spatial relations by conceptual neighborhood enables
numerous features for representing spatial knowledge and for spatial reasoning:
• Sets of neighboring relations can be lumped together to define coarse relations
(Freksa 1992a, b);
• Conceptual neighborhoods define hierarchies for representing incomplete
knowledge (Freksa and Barkowsky 1996);
• Qualitative reasoning based on conceptual neighborhoods allows for efficient
non-disjunctive reasoning (Nebel and Bürckert 1995; Balbiani et al. 2000);
• Neighborhood-based incomplete knowledge can be easily augmented as addi-
tional knowledge is gained during successive reasoning (Freksa 1992b);
• Coarse relations based on conceptual neighborhoods frequently exhibit a natural
correspondence to everyday human concepts (Freksa 1992a);
• Spatial and temporal inferences in qualitative reasoning typically result in
conclusions that form conceptual neighborhoods (Freksa 1992a, b);
• Conceptual neighborhoods can be formed on various levels of granularity (cf.
Fig. 5).

4 Neighborhood-Based Reasoning

One important feature of conceptual neighborhood-based abstraction is that

incomplete knowledge can be conceptualized and represented as coarse knowledge
(Fig. 5). By abstracting from missing or unnecessary details, reasoning can be carried
Spatial Computing 29

Fig. 5 Coarse temporal relations forming an abstraction hierarchy. The relation ‘older
contemporary of’ corresponds to the conceptual neighborhood of the three finer relations
‘overlaps’, ‘finished by’, and ‘contains’. The even coarser relation ‘older than’ corresponds to a
larger conceptual neighborhood that additionally includes the two fine relations ‘before’ and
‘meets’

out efficiently. In this way, computationally and conceptually problematic properties

of disjunctive knowledge processing are avoided which are encountered when
incomplete knowledge is represented as a set of completed potential alternatives.
Coarse reasoning does not necessarily yield coarser results than reasoning with
fine relations. But reasoning with coarse relations calls for different inference
procedures than reasoning with fine relations. Conjunctions of partially overlapping
coarse inferences based on imprecise or incomplete knowledge fragments from
different sources result in more precise or fine conclusions if the premises are
appropriately chosen. With this property, the coarse reasoning approach is suited to
model the synergy of multimodal coarse knowledge sources that result in precise
knowledge (cf. distributed representations and coarse coding in biological or artificial
perceptual systems, e.g. Edelman and Intrator 2000). Figure 6 presents a coarsened
version of the Allen composition rules that exploits conceptual neighborhood rela-
tions between fine relations. For example, the relation older contemporary (oc)
corresponds to the union of overlaps (o), finished by (fi), and contains (di).

5 A Multitude of Specialized Calculi and SparQ

A considerable variety of spatial calculi have been developed over the past
20 years (Cohn and Hazarika 2001); these can be classified as
• Measurement calculi, e.g. D-Calculus (Zimmermann 1995);
• Topological calculi, e.g. 4-intersection calculus, 9-intersection calculus, RCC-5,
RCC-8 (Egenhofer and Franzosa 1991; Randell et al. 1992);
• Orientation calculi, e.g. point/line-based: DCC, FlipFlop, QTC, dipole or
extended objects (Freksa 1992b; Ligozat 1993; Van de Weghe et al. 2005;
Moratz et al. 2000);
• Position calculi, e.g. Ternary point configuration calculus (TPCC—Moratz et al.
2003).
30 C. Freksa

Fig. 6 Conceptual neighborhood-based composition table and inferences based on this table.
Spatial pictograms symbolically depict 1D oriented spatial or temporal relations. Each black dot
corresponds to a fine relation; conceptually neighboring relations form lumps of dots that
correspond to coarse relations. Top Conceptually neighboring columns and conceptually
neighboring rows from the original table have been merged. Bottom Two coarse inferences using
this composition table (composition operator is denoted by ). Above the pictograms, the
relations are symbolized in classical logic notation. For an elaborate explanation see (Freksa
1992a)

To simplify and support the use of qualitative spatial calculi for specific
reasoning tasks, various tools have been developed. Prime examples are SparQ
(Wallgruen et al. 2007; Wolter and Wallgruen 2012); GQR (Westphal et al. 2009);
QAT (Condotta et al. 2006); and CLP (QS) (Bhatt et al. 2011). While some
approaches focus at specialized spatial reasoning methods, others aim to integrate
specialized techniques with general knowledge representation methods for logic-
based reasoning.
The toolbox SparQ1 integrates numerous calculi for qualitative spatial rea-
soning and allows for adding arbitrary binary or ternary calculi through the
specification of their base relations and their operations in list notation or through
algebraic specification in metric space. SparQ has a modular architecture and can
easily be extended by new modules (Fig. 7).
SparQ performs a number of operations that are helpful for dealing with spatial
calculi:
• Qualify: quantitatively described configurations are translated into qualitative
relations;

1
www.sfbtr8.spatial-cognition.de/project/r3/sparq/ (accessed: 1 Jan 2012).
Spatial Computing 31

Fig. 7 Modular SparQ architecture. Operations in different qualitative reasoning calculi can be
invoked through standardized commands (from Wolter and Wallgruen 2012)

• Compute-relation generates a qualitative inference for a given calculus based on

the premise relations and the calculus specification;
• Constraint-reasoning allows for the specification of an inference strategy on a
given spatial configuration and returns scenarios that are consistent with the
configuration; if the description of the scenario is inconsistent, SparQ informs
about the inconsistency;
• Neighborhood-reasoning enables conceptually compatible constraint relaxation
and yields semantically meaningful neighboring inferences;
• Quantification generates prototypical ‘general’ pictorial instances of abstract
qualitative descriptions (this is still in an experimental stage).
Although it is helpful to have a variety of calculi available in uniform speci-
fication and interface languages, there is still an issue about which calculus to
select to solve a given problem. Thus, there is a challenge to understand and
describe spatial calculi on the meta-level. The goal is to specify spatial configu-
rations and the type of required problem solution in such a way that the available
calculi can be automatically configured to solve the problem.

6 From Spatial Relations to Spatial Configurations

Quantitative computation of spatial configurations by means of Euclidean geom-

etry is well understood. For example, in planar geometry, we can compute all
angles, heights, and the area of arbitrary triangles, if the lengths of the edges of the
triangles are given by means of the formulae depicted in Fig. 8.
32 C. Freksa

Fig. 8 Formal abstraction of geometric relations in the Euclidean plane

These formulae are valid for planar spatial configurations independently of

position, orientation, scale, or other influences. Spatial relations in physical envi-
ronments conform to topological and geometric laws that are not affected by con-
textual influences from other modalities. As a consequence, only few constraints
need to be specified and many—or even all—spatial relations are determined.
The principle is well known from high school geometry. For example, on a flat
sheet of paper, we can construct exactly two triangles from the specification of
three line segments, provided the specified lengths conform to the triangle
inequality. In this construction, a compass and ruler are capable of qualitative
representation and they exhibit certain abstraction capabilities: the compass
represents a distance equal to the length of a given line segment and can apply this
distance abstracting from location and orientation. Similarly, the ruler represents a
distance and can apply it to any pair of points, independently of orientation and
location (within practical bounds).

7 Preserving Spatio-Temporal Structure

Although the formal abstraction shown in Fig. 8 is capable of generating arbitrary

spatial relations through abstract computation, the abstraction mechanism does not
preserve spatial structure in the way neighborhood-based representations preserve
Spatial Computing 33

the structure of the represented spatial domain. Structure-preserving representations

exploit structural correspondences between the representation medium and the
represented domain. They have the advantage that essentially the same operations
can be applied to the representation as to the represented domain. For example, on a
geographic map we can navigate much like in the geographic environment with the
advantage that we can maintain an overview more easily and that we do not need to
cover large distances.
As a consequence, structure-preserving representations (Sloman 1971) are
advantageous at least for those situations in which humans use the representations;
this is the case for assistance systems, for example, where spatial and temporal
representations are employed as human–machine interfaces. Humans can carry out
zooming operations by moving towards or away from the representation medium;
at the same time they can perform refinement and coarsening operations; they can
perform perspective transformations by looking at the medium from different
angles; they can aggregate and partition spatial regions by making use of natural
neighborhood structures; they can move across the medium much like in the
represented domain and they can experience spatial and conceptual transitions
while doing so; structure-preserving media also may support shape transformation
operations in similar ways as in the represented domain.
Are there additional reasons for exploring structure-preserving representations
besides the convenience for human users? I believe so. The operations described in the
previous paragraph are helpful not only for human users; they may be useful whenever
• problem statement and problem solution are in the spatial domain;
• there is a single spatial configuration about which we may want to answer many
questions;
• there are agents with spatial perception and locomotion, e.g., mobile robots;
• several agents need to communicate about a given spatial configuration;
• they can save resources by avoiding unnecessary operations.
In other words: structure-preserving representations also may be advantageous
for machine processing. We will come back to this consideration in the next
section.
Geometric-diagrammatic constructions on a piece of paper can serve as
structure-preserving representations of space, since flat paper provides the
universal spatial structure that guarantees the correctness of trigonometric relations
in a planar domain. Figure 9 depicts universal correspondences between geometric
functions in planar spatial structures.
Computation by diagrammatic construction is a form of analogical reasoning
(cf. Gentner 1983): the basis for establishing analogies is given through the universal
spatial interdependencies that justify the comparison between the source domain and
the target domain; the analogies usually concern the abstraction from specific values
in the domain. Nevertheless, geometric constructions are sequential constructions
that are most easily described by classical algorithms and procedures.
34 C. Freksa

Fig. 9 Spatial construction

of trigonometric functions.
The graph depicts
interdependencies of
geometric relations. All
trigonometric functions of an
angle H can be constructed
geometrically in terms of a
unit circle centered at O

Fig. 10 Three line segments

are applied to a spatially
structured domain. Numerous
new entities and relations are
established through the
interaction of these lines and
the constraints of the domain:
nine new line segments, 12
angles, a triangle, its area,
and the spatial relations
between all these entities

8 Space as Computer

In his book Rechnender Raum (‘Computing Cosmos’ or ‘Calculating Space’)

(Zuse 1969), the computer pioneer Konrad Zuse discussed the issue of structure
correspondence between computational representations and the physical domain.
He addressed the issue on the micro-level of discrete versus continuous structures,
maintaining that discrete representations only approximate continuous structures
and mimic random deviations rather than replicating the physical laws of quantum
mechanics.
In this section, I want to discuss the idea of structure correspondence on the
macro-level of spatial configurations and carry the notion of diagrammatic
construction one step further.
Suppose we apply three line segments to a flat surface as shown in Fig. 10.
What do we see in this figure? We can easily identify nine additional line segments
of specific lengths, three line intersections at specific locations, twelve specific
pairwise identical angles, one triangle with a specific area, and numerous relations
between those entities.
Where did all these entities and relations come from as we only placed three simple
straight lines onto the surface? One way to answer this question is: The surface
computed these entities and relations according to the laws of geometry. This would be
Spatial Computing 35

the type of answer we would give if we gave a computer the line equations and the
procedures to generate the mentioned entities and relations. What is the difference
between the computer approach and the ‘flat paper approach’?
The computer algorithm encodes knowledge about the spatial structure of the
surface that enables its interpreter to reconstruct in a sequential procedure step-by-
step certain abstractions of its spatial structure that are constrained by abstract
representations of the lines and their relationships. On the other hand, the flat
surface itself and its spatial structure relate directly and instantly to the lines and
generate the entities and relations without computational procedure by means of
the inherent structural properties. It represents space rather than knowledge about
space. This is why I call this approach spatial computing rather than knowledge
processing.

9 The Notion of Spatial Computing

Much of what we do in artificial intelligence and computer science takes place on

the knowledge level (Newell 1982). The hype of general purpose computing in the
1960s was based on the insight that we can express everything we can think and
talk about in terms of physical symbols and that we can manipulate these symbols
in a computer similarly to the way we reason and talk about arbitrary domains. In
this way we can use our knowledge and understanding of these domains to answer
questions about them and to solve problems. In the generality of this insight we
may have lost sight of the fact that the domains of space and time are omnipresent
not only in the worlds we talk about but also within our physical symbol
manipulation systems. Considering this fact, couldn’t we make use of the spatial
and temporal properties of these physical systems on the object level rather than
reason about them on the meta-level? For certain tasks in the spatial and temporal
domains we would simply act in space and time and see what happens rather than
process knowledge about space and time and know what happens.
Figure 11 schematically depicts the relation between the meta-level of formal
and computational reasoning and the object level of spatial configurations. The
formal reasoning approach to computing spatial relations is shown in the upper
part of the figure; the approach that applies spatial structures directly is shown in
the lower part of the figure. In the classical formal reasoning approach, the task
either must be given in formal terms or it must be formalized from object-level
configurations before formal reasoning processes can be invoked. The formal
result can be presented on a formal level or be transformed to an object-level
configuration by instantiation.
The spatial computing paradigm takes place on the object level of spatial
configurations. The task is directly presented as spatial configuration (e.g., the
configuration of Fig. 10), or a spatial configuration is generated through instan-
tiation from a formal specification. If the task is to answer questions about spatial
properties and relations of the configuration, the result is available instantly due to
36 C. Freksa

Fig. 11 Two approaches to generating spatial entities and relations: in the upper part of the
figure, a classical sequential symbolic computing approach transforms a formal specification of a
spatial configuration by means of formal reasoning into a formal result on the meta-level. In the
lower part a spatially structured substrate on the object level guarantees compliance with spatial
constraints and instantly makes available all spatial implications of the configuration in spatial
form. Specific relations can be read-off directly from the configuration. Transformations between
the object level and the meta-level can be carried out at the task stage or the solution stage

the intrinsic constraints of the spatial substrate on the object level. If the task
involves physical operations on spatial configurations, these operations will be
subject to spatial constraints of changing physical configurations and require
processing time; but all spatial implications of the operations will be available
instantly as the operations are performed. Depending on the form in which the
result of the ‘computation’ is needed, we need processes that extract the desired
result from the configuration (e.g., a line segment or an angle) as input for the next
spatial computing task; if the result is needed on the formal level, the object-level
entity needs to be formalized, e.g., for use in a classical computation process.
The spatial computing approach involves a paradigm shift that makes it difficult
to compare with symbolic approaches by purely computational measures.
The reason is that in the symbolic realm, we assume that problems are given in
formalized form and that the results will be needed in formalized form, as well.
Perceptual operations necessary to formalize spatial knowledge are not taken into
account in symbolic approaches. Thus, computational cost is restricted to symbol
manipulation processes. However, real-world spatial situations may be different as
the example problems in the introduction suggest: for some of them the formal-
ization task may be too time-consuming or expensive; a direct object-level
mapping to a spatial substrate may become more feasible, particularly, as sensor
technology continues to develop. Perception processes that had to be performed by
humans in the process of formalizing spatial knowledge become a part of the
spatial computing paradigm; however, they will not be discussed in this chapter.
Spatial Computing 37

The term ‘spatial computing’ is used by various researchers in interesting ways

that are related to the topic of this chapter: In a Dagstuhl Seminar on computing
media and languages for space-oriented computation the organizers state that ‘…
it is important to make space not an issue to abstract away, but a first-order effect
that we optimize. The distinguishing feature of spatial computing then is that
computation is performed distributed in space and topology define the computa-
tion’ (DeHon et al. 2007). Similar goals are stated in a proposal for spatial cloud
computing for use in the Geospatial Sciences (Yang et al. 2011). A student of the
MIT School of Architecture and Planning posted the manuscript for a master thesis
(Greenwold 2003) in which he defines ‘Spatial computing is human interaction
with a machine in which the machine retains and manipulates referents to real
objects and spaces.’ The artist Albert Hwang presents a video film series (Hwang
2012) in which he demonstrates simulations of augmented reality technology for
interaction with spatial environments to illustrate the power of spatial computing
technology that is yet to be developed.
Whereas the former two projects share technical goals—using spatial substrates
for computation—with the approach presented here, the latter two projects share
some of the motivation for our approach: interaction with spatial environments
through perception and action. The main motivation for our approach, however, is
to understand cognitive functionality by conceiving and implementing suitable
representations, processes, and technical realizations.

10 Basic Entities of Cognitive Processing

In geometry, the spatial world can be described in terms of infinitesimally small

points; lines are defined in terms of points; areas in terms of lines; etc. In contrast,
in cognition, basic entities usually are not infinitesimally small points; they may be

geometry cognition
Composition
Aggregation

object configurations
configurations
Composition

objects
Aggregation

objects
Decomposition
Refinement

parts
areas ‘basic
level’
lines

points

Fig. 12 Two ways to conceptualize physical objects. Left In geometry, we aggregate arbitrarily
complex structures from atomic point entities. Right In cognition, basic entities may be
geometrically complex, meaningful entities. Through cognitive effort, basic entities can be
decomposed into more elementary entities or aggregated into more complex configurations;
cognition on the level of basic cognitive entities is possible without invoking elementary
constituents
38 C. Freksa

entire physical objects like books or chairs (Fig. 12). Basic cognitive entities carry
meaning related to their use and function and we perceive and conceptualize them
in their entirety even if certain details are not accessible to our perception. We may
not even know about the composition of basic cognitive entities; still, we are able
to talk about them and to use them for daily activities. The cognitive apparatus
appears to be flexible as to which level in a huge lattice of part-whole relations to
select as ‘basic level’ (cf. Rosch 1978); it also appears to be able to focus either on
the relation between an object and a configuration of objects, or alternatively, on
the relation between an object and its parts. Both transitions involve cognitive
effort, while the mere consideration of the basic level appears almost effortless.
It is known that we can apply simple mental operations, e.g., mental rotation, to
entire spatial objects at once (Shepard and Metzler 1971). In spatial computing, we
would like to implement processes that have comparable capabilities: manipulating
entire objects without manipulating all their constituent parts. We would expect to
obtain only a coarse result of cognitive operations on basic cognitive entities with
little effort; to resolve details, we would have to invest cognitive power. This pro-
cessing approach would be in contrast to geometric spatial processing where we
would expect to know the details before we know the complex structure.

11 Conclusions and Outlook

Let me return to the spatial problems that I used in the beginning of the chapter to
introduce various perspectives on spatial challenges. The main message of this
exercise is that spatial problem solving consists of more than solving equations.
First of all, a spatial problem needs to be perceived as one. Second, it needs to be
represented as one. Third, the representation needs to be processed. Fourth, the
result needs to be interpreted in spatial terms.
With regards to the representation of spatial knowledge for problem solving we
have lots of options, as there are many ways to conceive of space. For example,
space may be conceived of as empty space—‘‘what is there when nothing is
there’’—or as the space spanned by physical objects. Space can be described in
terms of a multitude of reference systems as becomes evident if we look at the
many spatial representation systems and calculi we can develop. All the different
representations have advantages and disadvantages, depending on the problems we
want to tackle or the situations we want to describe. Some problems can be solved
directly on the object level; others are facilitated by suitable abstractions.
Nevertheless, spatial structures—and to a similar extent temporal structures—play
special roles in everyday actions and problem solving. Many other dimensions seem to
dominate our lives: monetary values, quality assessments, efficiency criteria, emotions,
social structures, etc.—but do they play comparable roles with respect to cognitive rep-
resentation and processing? I do not think so. I propose that the special role of space and
time has to do with the fact, that internal representations may be a-modal, but they cannot
be ‘‘a-structural’’. In other words: cognitive representations and processes depend on a
Spatial Computing 39

spatio-temporal substrate; without such a substrate, they cannot exist. But they may not
depend on a specific spatio-temporal substrate: a multitude of structures may do the job.
Different abstractions from physical space may be advantageous in different situations.
Space and time provide fundamental structures for many tasks that cognitive agents
must perform and for many aspects of the world that they can reason about. Main-
taining these structures as a foundation simplifies cognitive tasks tremendously,
including perceiving, memorizing, retrieving, reasoning, and acting. This is well
known from everyday experiences, such as using geographic maps for wayfinding.
For other domains it is helpful to create spatially structured foundations to support and
simplify orientation; for example, spatial structure is the basis for diagrams that help
us reason about many spatial and non-spatial domains.
A conceptually simple implementation of a truly spatial computer could be a
robot system that manipulates physical objects in a spatial domain and perceives
and represents these objects, the configurations constructed from these objects, and
the parts of the objects as well as their relations from various orientations and
perspectives. A more sophisticated approach would involve the construction of a
spatial working memory, perhaps visually accessed, whose basic entities are entire
objects rather than their constituents. Spatial operations like translation, rotation,
and distortion would globally modify configurations. Perception operators extract
qualitative spatial relations from these representations. The development of this
implementation can be guided by our knowledge about working memory capa-
bilities and limitations as well as by our knowledge about spatial representations in
the human mind (Schultheis and Barkowsky 2011).
As technological materials become more sophisticated, the connection between
spatial substrates and digital computer technology will become successively
stronger. Sensor technology will be integrated into spatially structured materials in
the years to come. A vision of such materials of the future that would support the
concept of spatial computing on the substrate level can be found in a recent special
issue in sensors and actuators (Lang et al. 2011).

12 A Final Note

Although we talk about spatial cognition, spatial reasoning, and spatial computing,
we frequently fail to characterize the type of solution to spatial problems that we
want to achieve. Our repertoire of approaches yields results on different levels of
sophistication: some approaches only yield solutions to spatial problems; others
yield some sort of explanations along with the solutions or instead of a solution.
Accompanying explanations may be: ‘this is the only solution’; ‘this is one of
possibly several solutions’; ‘these are all solutions’; or ‘there is no solution’.
Why is sophistication an issue? For highly abstract, formal approaches the
quality of a solution is not obvious. Formal proofs or explanations (or both) are
required to characterize the type of solution. In the more concrete, spatially
structured solutions, the results are more easily perceptible, more obvious in that
40 C. Freksa

proofs may not be required—cognition and commonsense reasoning seem to

operate without formal proofs, for the most part. On the other hand, can we be sure
that we found the best solutions, the only solution, or all solutions? This is an old
debate that calls into mind the discussion on the validity of constructive geometry
to find solutions or to prove correctness.
There are different domains in which we can ground our knowledge: perceptual
experience about spatial and temporal environments that does not require proofs
and formal logics that does not require empirical justification. Both domains are
important for human intellect and human reasoning. It does not make much sense
to say one is superior over the other; they are two rather different realms. They
may become particularly powerful when they are engaged jointly, one to carry out
spatio-temporal perception and action and the other to reason about them on the
meta-level and to explain what is going on in an overarching theory.
It is interesting to note that artificial intelligence research on commonsense
reasoning so far has been restricted to characterizing commonsense reasoning on a
descriptive level. Almost no AI work exists that emulates spatial cognitive abilities
in a similar way as constructive geometry reflects spatial laws in the physical
world or as artificial neural networks reflect topological structures akin to those of
biological systems.

Acknowledgments I am grateful for valuable and detailed comments on earlier versions of this
chapter from Thomas Barkowsky, Mehul Bhatt, Stefano Borgo, Holger Schultheis, Thora Ten-
brink, Diedrich Wolter, several anonymous reviewers, and the editors of this book. Generous
support from the German Research Foundation to the Spatial Cognition Research Center SFB/TR
8 Bremen and Freiburg is gratefully acknowledged.

References

Allen JF (1983) Maintaining knowledge about temporal intervals. CACM 26(11):832–843

Balbiani P, Condotta J-F, Ligozat G (2000) Reasoning about generalized intervals; preconvex
relations and tractability. In: Proceedings of TIME-2000 conference, pp 23–30
Bhatt M, Lee JH, Schultz C (2011) CLP(QS): a declarative spatial reasoning framework. In:
Egenhofer MJ, Giudice NA, Moratz R, Worboys MF (eds) COSIT’11, LNCS 6899. Springer,
Heidelberg, pp 210–230
Cohn AG, Hazarika SM (2001) Qualitative spatial representation and reasoning: an overview.
Fundamenta Informaticae 43:2–32
Condotta JF, Saade M, Ligozat G (2006) A generic toolkit for n-ary qualitative temporal and
spatial calculi. In: TIME’06. IEEE Computer Society, Washington, pp 78–86 ISBN 0-7695-
2617-9
DeHon A, Giavitto J-L, Gruau F (2007) 06361 Executive report — computing media languages
for space-oriented computation. In: DeHon A, Giavitto J-L, Gruau F (eds) 06361 Abstracts
collection—computing media languages for space-oriented computation, Schloss Dagstuhl.
http://drops.dagstuhl.de/opus/volltexte/2007/1026
Edelman S, Intrator N (2000) Coarse coding of shape fragments ? retinotopy * representation
of structure. Spat Vis 13:255–264
Egenhofer MJ, Franzosa RD (1991) Point set topological relations. Int J Geogr Inf Syst 5:
161–174
Spatial Computing 41

Freksa C (1991a) Conceptual neighborhood and its role in temporal and spatial reasoning. In:
Singh M, Travé-Massuyès L (eds) Decision support systems and qualitative reasoning. North-
Holland, Amsterdam, pp 181–187
Freksa C (1991b) Qualitative spatial reasoning. In: Mark DM, Frank AU (eds) Cognitive and
linguistic aspects of geographic space. Kluwer, Dordrecht, pp 361–372
Freksa C (1992a) Temporal reasoning based on semi-intervals. Artif Intell 54:199–227
Freksa C (1992b) Using orientation information for qualitative spatial reasoning. In: Frank AU,
Campari I, Formentini U (eds) Theories and methods of spatio-temporal reasoning in
geographic space, LNCS 639. Springer, Berlin, pp 162–178
Freksa C, Barkowsky T (1996) On the relation between spatial concepts and geographic objects.
In: Burrough P, Frank A (eds) Geographic objects with indeterminate boundaries. Taylor and
Francis, London, pp 109–121
Gentner D (1983) Structure-mapping: a theoretical framework for analogy. Cogn Sci 7:155–170
Greenwold S (2003) Spatial computing. Manuscript submitted to school of architecture and
planning, Massachusetts Institute of Technology, Cambridge. http://acg.media.mit.edu/
people/simong/. Accessed 1 Jan 2012
Guesgen HW (1989) Spatial reasoning based on Allen’s temporal logic. ICSI TR-89-049.
International Computer Science Institute, Berkeley
Hamblin CL (1972) Instants and intervals. In: Fraser JT, Haber FC, Müller GH (eds) The study of
time. Springer, Berlin, pp 324–331
Hwang A (2012) http://albert-hwang.com/projects/spatial-computing/. Accessed 1 Jan 2012)
Kosslyn SM (1987) Seeing and imagining in the cerebral hemispheres: a computational approach.
Psychol Rev 94:148–175
Lang W, Lehmhus D, van der Zwaag S, Dorey R (2011) Sensorial materials—a vision about
where progress in sensor integration may lead to. Sens Actuators, A 171(1):1–2
Ligozat G (1993) Qualitative triangulation for spatial reasoning. In: Campari I, Frank AU (eds)
COSIT 1993 LNCS 716. Springer, Heidelberg, pp 54–68
Liu W, Li S (2011) Reasoning about cardinal directions between extended objects: the NP-
hardness result. Artif Intell 175:2155–2169
Moratz R, Renz J, Wolter D (2000) Qualitative spatial reasoning about line segments. In:
Proceedings of ECAI 2000, pp 234–238
Moratz R, Tenbrink T, Fischer F, Bateman J (2003) Spatial knowledge representation for human-
robot interaction. In: Freksa C, Brauer W, Habel C, Wender KF (eds) Spatial cognition III –
Routes and navigation, human memory and learning, spatial representation and spatial
reasoning, LNAI 2685. Springer, Heidelberg, pp 263–286
Nebel B, Bürckert HJ (1995) Reasoning about temporal relations: a maximal tractable subclass of
Allen’s interval algebra. JACM 42(1):43–66
Newell A (1982) The knowledge level. Artif Intell 18(1):87–127
Nicod J (1924) Geometry in the sensible world. Doctoral thesis, Sorbonne, English translation in
Geometry and Induction, Routledge and Kegan Paul, 1969
Randell DA, Cui Z, Cohn AG (1992) A spatial logic based on regions and connection. In:
Proceedings of 3rd international conference on knowledge representation and reasoning.
Morgan Kaufman, Los Altos, pp 55–66
Rosch E (1978) Principles of categorization. In: Rosch E, Lloyd BB (eds) Cognition and
categorization. Erlbaum, Hillsdale
Schultheis H, Barkowsky T (2011) Casimir: an architecture for mental spatial knowledge
processing. Top Cogn Sci 3:778–795
Schwartz JT, Sharir M (1983) On the ‘‘piano movers’’ problem I: the case of a two-dimensional
rigid polygonal body moving amidst polygonal barriers. Commun Pure Appl Math 36:
345–398
Shepard RN, Metzler J (1971) Mental rotation of three-dimensional objects. Science 171:
701–703
Skiadopoulos S, Koubarakis M (2004) Composing cardinal direction relations. Artif Intell
152:143–171
42 C. Freksa

Sloman A (1971) Interactions between philosophy and artificial intelligence: the role of intuition
and non-logical reasoning in intelligence. Artif Intell 2:209–225
Sloman A (1985) Why we need many knowledge representation formalisms. In: Bramer M (ed)
Research and development in expert systems. Proceedings of BCS expert systems conference
1984. Cambridge University Press, Cambridge, pp 163–183
Tobler W (1970) A computer movie simulating urban growth in the Detroit region. Econ Geogr
46(2):234–240
Van de Weghe N, Kuijpers B, Bogaert P, De Maeyer P (2005) A qualitative trajectory calculus
and the composition of its relations. In: Proceedings of geospatial semantics, vol 3799,
pp 60–76
Wallgruen JO, Frommberger L, Wolter D, Dylla F, Freksa C (2007) Qualitative spatial
representation and reasoning in the SparQ toolbox. In: Barkowsky T, Knauff M, Ligozat G,
Montello D (eds.) Spatial cognition V: reasoning, action, interaction, LNAI 4387. Springer,
Heidelberg, pp 39–58
Westphal M, Woelfl S, Gantner Z (2009) GQR: a fast solver for binary qualitative constraint
networks. In: AAAI spring symposium on benchmarking of qualitative spatial and temporal
reasoning systems, Stanford
Wolter D, Wallgruen JO (2012) Qualitative spatial reasoning for applications: New challenges and
the SparQ toolbox, In: Hazarika SM (ed) Qualitative spatio-temporal representation and
reasoning: Trends and future directions. IGI Global, Hershey. doi: 10.4018/978-1-61692-868-1
Yang C, Goodchild M, Huang Q, Nebert D, Raskin R, Xu Y, Bambacus M, Fay D (2011) Spatial
cloud computing: how can the geospatial sciences use and help shape cloud computing? Int J
Digit Earth 4:4
Zimmermann K (1995). Measuring without measures. The D-Calculus. In: Frank AU, Kuhn W
(eds) COSIT 1995, LNCS 988. Springer, Heidelberg, pp 59–67
Zuse K (1969) Rechnender Raum: Schriften zur Datenverarbeitung. Vieweg, Braunschweig