By using means-ends search guided by goals, the search
space of possible actions to be added to the behavior of the
gadget is pruned, resulting in improved average-case
efficiency of the algorithm. Each iteration of the algorithm
tries to satisfy a goal in the blend space, and will thus only
consider actions that can achieve that goal. While the total
number of actions in both input spaces may be large,
usually only a small fraction of them can be projected to
achieve any given goal. In contrast, a naïve selective
projection algorithm will attempt to project all actions in
any input space using all possible projection methods.
Although the goal-driven best-first search in the worst case
has to consider all possibilities in the search space, in an
average case it only considers a small portion of the total
possibilities (Weld 1994). A goal-driven blending
algorithm also has the added benefit of ensuring that any
result of the algorithm is guaranteed to meet all of the
acceptability requirements.
In summary, the gadget generation system implements a
goal-driven model of blending, including relatively
efficient selective projection (due to pruning of the search
space) and elaboration procedures. These procedures are
driven by the gadget’s goal, the particular story that serves
as the context, and the general domain of storytelling.
However, this system has not fully investigated the
selection of input spaces, which will be illustrated in the
next case study.
A Virtual Character for Pretend Play
Our second case study illustrates the use of goals in
blending with a specific focus on the selection of
appropriate input spaces. In this case, a real-world object is
selected to represent an object from a fantasy world, as
required in children’s pretend play. A goal specifies means
to appropriately prune potential input spaces and select one
option based on contextual constraints of similarity. Below
we describe how our pretend play system can be viewed
through the lens of conceptual blending; full details on the
system are presented in (Zook, Riedl, and Magerko 2011).
In pretend play, children construct and enact story
scripts and roles with real-life objects (Nourot 1998).
Examples of pretend play include lightsaber duels with
cardboard tubes, holding pretend tea parties with stuffed
animal guests and acting as a group of pirates sailing on a
couch. When enacting these scripts, pretenders have goals
of using particular objects from a fictional world, but are
limited to using the real-world objects that are ready at
hand. Pretend players imaginatively overlay the fictional
object onto the real-world object to create a blend, i.e. a
pretend object existing in both the fictional world and the
real world. Pretend play research has found children
project traits of the fictional object on the real object. As an
example, children engaged in a lightsaber duel with
cardboard tubes may make buzzing noises when they
swing the tube. In general, the pretending process involves
identifying real-world objects as stand-ins for the fictional
object, and selectively projecting traits of the fictional
object onto the real object. The objects are input spaces to
the blend. The construction process must account for how
input spaces relate to a larger context of a target pretend
play activity, pruning the options considered.
Building computational systems that can engage in
pretend requires the capacity to construct the objects used
in these scripts (Zook, Riedl, and Magerko 2011). To
formalize the problem, the play activity (lightsaber duel)
provides a structuring situation such that a pretender—
human or agent—selects a real world object (cardboard
tube) as a presentation space for a given reference space of
a fictional object (lightsaber). That is, the goal is to find a
presentation input space that most closely matches the
reference input space. Once the input spaces are selected,
the blending process takes the most relevant aspects of the
fictional object for the activity (buzzing), which are
imposed onto the real object in the blend space for use in
play (swinging a cardboard tube while buzzing). This
process starts with a situation in the fictional world and a
specific fictional object (the lightsaber I am using in the
duel), and seeks a presentation object in the real world to
effectively manifest the fictional object.
To reason about numerous objects in the fictional world
and the real world, we need a computational representation
of objects and their attributes. Lakoff and Johnson (1980)
proposed that the salient perceptual, motor-activity, and
purposive features of objects affect how humans interact
with them. We model objects in both the fictional and real
domains using selected attributes in these categories.
Following prototype theory (Rosch 1978), these attributes
are assigned fuzzy values to represent a real-valued ([0, 1])
range of degree of membership (DOM). As an example, a
lightsaber may have a 0.8 DOM value for the perceptual
feature of being blue (very blue, except for the handle), 0.9
DOM value for the motor-activity feature of ease of
handling (very easy to hold and swing), and 0.1 DOM
value for the purpose of supporting weights (unsuitable for
propping up heavy objects).
Iconic attributes are salient attributes of an object that
distinguish it from similar objects within the same
category. These attributes help to resolve the potential
ambiguity of which fictional object is being represented by
a given real world object. For example, if a pretender grabs
an object and begins making buzzing noises, it may be
unclear if they are signaling that they are holding a buzzing
lightsaber or shooting a laser pistol. An iconic posture of
handling, however, makes this difference clear. Iconic
attributes help participants in pretend play interpret other
players’ behaviors and intentions.
The computational play system algorithm has three steps
for context and goal driven blending: (1) select a real-
world object based on the pretending goal and context;
(2) select the set of fictional object attributes to project;
and (3) project these attribute values into the blend. The
first step is the selection of the input space—the real-world
object—to be blended with the fictional object.
Selecting an input space uses the pretending goal to first
prune impossible input spaces and then search for the
optimal input space among those that remain. Conceptually