Interaction-based Authoring for Scalable Co-creative Agents
Mikhail Jacob, Brian Magerko
School of Interactive Computing
Georgia Institute of Technology
Atlanta, GA USA
{mikhail.jacob, magerko}@gatech.edu
Abstract
This article presents a novel approach to authoring cocreative
systems - called interaction-based authoring -
that combines ideas from case-based learning and imitative
learning, while emphasizing its use in open-ended
co-creative application domains. This work suggests an
alternative to manually authoring knowledge for computationally
creative agents that relies on user interaction
“in the wild” as opposed to high-effort manual authoring
beforehand. The Viewpoints AI installation is
described as an instantiation of the interaction-based
authoring approach. Finally, the interaction-based authoring
approach is evaluated within the Viewpoints AI
installation and the results are discussed guiding development
and further evaluation in the future.
Introduction
Within the computational creativity community, our research
has focused on domains that are open-ended, artistically
performative, improvisational, and co-creative between
human and AI agent. co-creative AI agents that can
succeed in these kinds of domains tend to be large-scale
and knowledge-rich since they have to collaborate creatively
on an equal footing with humans. Therefore, one of
the key bottlenecks for developing co-creative agents has
been the knowledge-authoring bottleneck. According to
Csinger et al. (1994), the difficulty, cost, or delay in acquisition
of expert instantial knowledge followed by its subsequent
structuring and storage so as to enable efficient future
utilization is often referred to as the knowledgeauthoring
bottleneck. In fact, the knowledge-authoring
bottleneck has historically been a significant problem for
the intelligent agent community in general and the computational
creativity community in particular.
Many solutions have been proposed in the past to mitigate
the problem. Case-based reasoning (CBR) approaches
and machine learning approaches have utilized online case
acquisition and data mining from corpora as fundamental
methods for dealing with the knowledge-authoring bottleneck.
Data mining approaches have faced a general lack of
corpora for instantial or behavioral content within improvisational
performative domains. Traditional CBR systems
while learning from experience still require instantial content
to be authored in the form of an initial case library.
Learning by demonstration or observation can avoid these
pitfalls, but traditionally require explicit training or teaching
phases before they can be used in the final task.
Within the games research community procedural content
generation (PCG) research has focused on developing
algorithms to generate the instantial content that was once
manually authored by expert designers. This has seen success
with the development of procedurality-centric games
such as Spore and Galactic Arms Race (Hecker et al. 2008;
Hastings et al. 2009). However, PCG systems have yet to
focus on generating behavioral content that is flexible
enough to work in open-ended improvisational domains.
In contrast to the previous authoring approaches mentioned,
this article describes a hybrid knowledge-authoring
paradigm that combines case-based learning with learning
by observation / imitative learning – called interactionbased
authoring. Interaction-based authoring aims to i)
minimize the authoring bottleneck while ii) ensuring that
the subjective experience of interacting with the system is
high quality and that iii) the computer collaborator supports
equal creative agency (the extent to which a creative
collaborator can take decisions, make choices, and affect
co-creation). It proceeds to demonstrate the interactionbased
authoring paradigm within an improvisational interactive
art installation called Viewpoints AI (Jacob et al.
2013a) after comparing the installation to related work in
the field. A brief updated system description is provided
(c.f. Jacob et al. 2013b). Finally, the paper details an initial
attempt to evaluate the interaction-based authoring approach
instantiated within the Viewpoints AI installation
and discusses the results as a guide for iteratively developing
/ refining the installation.
Interaction-Based Authoring
Interaction-based authoring is a hybrid approach to authoring
instantial knowledge and control knowledge for cocreative
interactive systems, combining case-based learning
with imitative learning. While using an interactionbased
authoring approach learning occurs over the lifetime
of the full performance and not during an explicit training
or teaching phase. This is done to boost participant motivation
and engagement encouraging prolonged interaction
Proceedings of the Sixth International Conference on Computational Creativity June 2015 236
with the agent thereby facilitating greater knowledge acquisition.
There are three main aspects to interaction-based authoring.
First, case-based learning is used to index and store
agent experiences in a reusable manner that can be utilized
to drive future behavior or responses in general. Cases can
be stored as input–output pairs (from the agent’s perspective)
with a process to map between inputs and outputs in
order to use them interchangeably.
Second, an imitative learning / learning by observation
system (Tomasello 2000) that can model the way a human
partner responds to the agent’s actions is utilized in order
to interact with other partners in other interaction contexts.
If the new partner’s input action (from the new interaction
context) is similar enough to an input action it has learnt a
model for in the past, it can use that to select an output
action. The agent takes the interactor’s role in that case and
responds, as they (presumably) would have.
Finally, an open-ended co-creative improvisational domain
in which to situate the agent is required so that the
participant or interactor is engaged and therefore motivated
to teach the system for an extended period of time. The
open-ended nature of the domain encourages exploration
of the interaction space, increasing the coverage of the
learning algorithms for future interactions. The co-creative
and improvisational aspects of the domain emphasize the
egalitarian nature of creative decision-making. They also
encourage the user to further explore novel regions of the
interaction space in the event that the system makes a
‘poor’ choice of response, thinking of it as an interesting
offer that they hadn’t considered rather than a mistake. The
interaction-based authoring approach has been instantiated
in an interactive improvisational human–AI art installation
called Viewpoints AI. A description of the installation follows
a brief account of related work.
Related Work
Technology has been contemporarily used to augment performances
and art installations (Reidsma et al. 2006; Latulipe
2011; MacDonald et al. 2015). These pieces use performance
technology as an integral part of their overall
aesthetics and content of the artwork. However, these technologies
have been subservient to human performers, with
shallow knowledge, and / or a lack of clear collaboration
between the machines and humans on stage.
Combining research in arts, AI, cognitive psychology
and philosophy, the field of computational creativity has
focused on many different creative domains (c.f. Boden
2003; c.f. Colton 2012). However, most traditional computationally
creative systems assemble pre-authored content
in novel combinations, without attempting to solve the
knowledge-authoring bottleneck, leading to small systems
with limited scope. In addition, many in the past have ignored
creative collaboration or co-creativity focusing on
systems that do not involve humans except as consumers
or evaluators of the creative artifact or process.
Computationally co-creative systems on the other hand
collaborate with humans in order to participate meaningfully
in the creative process or outcome. Much work has
been done on co-creative agents in the music improvisation
domain (Thom 2000; Hsu 2008; Hoffman and Weinberg
2010). The Digital Improv Project virtual agents that could
perform theatrical improvisation (O’Neill et al. 2011) and
the Computational Play Project virtual agents that could
play pretend with people using toys (Magerko et al. 2014)
are examples of co-creative systems in other domains.
Both however, required extensive pre-authored instantial
content to produce improvisational behavior. The Digital
Apprentice (a virtual collaborator for abstract visual /
sketch art creation; Davis et al. 2014) is a co-creative system
that closely resembles an instantiation of the interaction-based
authoring approach.
The Viewpoints AI Installation
The Viewpoints AI installation is a participatory interactive
installation where a human interactor and a virtual
agent – named VAI – collaborate to improvise movementbased
performance pieces together in real-time. The installation
(see Figure 1) is composed of a large translucent
muslin projection screen that has a human-sized manifestation
of VAI projected onto it from the front and the interactor’s
shadow cast onto it from the rear. This enables an
occlusion-free juxtaposition of the interactor’s shadow
onto the projected virtual agent when their positions overlap.
While the installation is highly participatory in nature
and the experience of improvising is intrinsically tied to it,
an audience can also watch the unfolding performance
from the front of the installation.
Participants interact with the virtual agent behind the
muslin screen while a Microsoft Kinect depth camera senses
and records their movements. Recorded movements are
analyzed systematically using a formal version of the
Viewpoints framework, as described by Bogart and Landau
(2005). Viewpoints is used in theatrical movement and the
staging of scenes to focus on the physicality of action and
analyze performance in terms of the physical Viewpoints
of time (tempo, duration, kinesthetic response, and repetition)
and space (shape, gesture, spatial relationship, topography,
and architecture), as well as the vocal Viewpoints
(pitch, dynamics, acceleration, silence, and timbre). The
Figure 1: The Viewpoints AI Installation
Proceedings of the Sixth International Conference on Computational Creativity June 2015 237
participant’s movements are interpreted through a subset of
the Viewpoints framework and are then responded to by
the agent. The formalization of Viewpoints is thus used as
a framework to represent and reason about movement.
The Viewpoints AI installation uses contrasting visual
elements of light and shadow to showcase how the human
participant and the virtual agent arrive at this liminal interaction
space from two very different worlds. Visually, VAI
is a glowing anthropomorphic character composed of a
playful cloud of fireflies. The participant’s crisp shadowed
form is transported to the ephemeral 2D space between the
two worlds through the medium of shadow theatre.
System Description
The Viewpoints AI installation is powered by an agent
architecture that is conceptually composed of three modules
– perception, reasoning, and action. Earlier versions
of the system are described in Jacob et al. (2013a; 2013b).
The following sections describe the agent architecture
briefly, going into more detail where necessary to illustrate
updated aspects of the system.
Perception
The Viewpoints AI agent architecture receives input from
the depth camera as a frame of joint positions in continuous
3D space at a certain frame rate to get “joint space”
gestures. It then discretizes the joint space gestures and
derives additional information about them in real-time using
a formalization of the Viewpoints framework to get
discrete “predicate space” gestures. These two types of
gestures are then sent along to the reasoning module.
Parsing Viewpoints Predicates The Viewpoints predicates
that have been formalized to date make up a subset of
the physical Viewpoints, including tempo, duration, and
repetition, as well as parts of spatial relationship, topography,
shape, and gesture. The current version of the installation
has a general purpose machine learning toolkit (Hall et
al. 2009) integrated within the agent architecture that classifies
Viewpoints predicates using classifiers trained using
supervised learning on expert movement-practitioner /
dancer data. Adding new predicates to the system is as
straightforward as training new classifiers with more data
demonstrating or exemplifying that particular attribute or
aspect of the Viewpoints framework. Emotional content of
the performance portrayed through gestures are also classified
through this supervised learning process.
The current movement analysis pipeline employs modular
feature detectors for motion-based features (eg. vertical
knee velocity, tangential knee acceleration, etc.) of the
joint space gestures. These are then used to feed classifiers
(with the specific classification algorithm chosen empirically
according to classification performance). Training
data for classification is obtained by collaborating with
expert local movement-practitioners and dancers.
Turn-taking Model Turn-taking refers here to the process
of naturalistically timing the use of the shared performance
space so as to coordinate each other’s (potentially overlapping
or simultaneous) movements. This can be decomposed
into the problems of how to best time the agent’s
movement turns coordinating with the interactor and how
to segment a user movement turn or gesture. The first
problem is solved by the interaction convention that the
agent moves whenever the interactor does, either mirroring
them (when they move arrhythmically) or improvising an
original response to their movements (when they perform
rhythmic repeated movements). The second problem is
discussed below.
In the current version of the installation the agent tries to
detect a beat to the interactor’s movement (helped by playing
dance music during the interaction) and segments their
gestures using the detected beat. It does this by creating a
set of 1D motion vector-based local beat detectors for each
moving joint. These report possible joint-level candidate
beats by looking at the half period of the joint motion.
When candidate beats are repeated multiple times, they are
confirmed and reported to a global tracker. The global
tracker then chooses a candidate local beat as the global
beat, which is then used to segment the movement at the
start and end of the beat duration. Additional trimming of
the segment is done so that the start and end are the same.
Reasoning
Segmented gestures in both joint and predicate space are
sent to the reasoning module for the agent to determine an
appropriate response gesture. Joint space gestures are then
stored in a gesture library in exemplar clusters, each cluster
having a universally unique identifier number (UUID).
These clusters are produced through an approximate gesture
recognition algorithm using a content vector of aggregated
versions of the same set of motion-based features
used earlier in Viewpoints predicate classification. This is
done in order to find patterns in interactions and cluster
similar gestures together. It is a simplification of the hard
problem of online matching in real-time of an input gesture
to one (or potentially none) out of a potentially unbounded
set of gestures without prior training of any sort. The corresponding
predicate space gesture is then sent to a Soar
agent (Laird 2012) for further processing in order to
choose a response gesture. This case-based learning is a
key mechanism within the Viewpoints AI installation that
helps it instantiate interaction-based authoring.
Response Strategies The Soar agent has a set of strategies
for selecting responses to the input gesture that are then
output to the action module. These strategies are selected
amongst using pragmatic and aesthetic rules for agent behavior.
The response strategies were chosen using an analogy
to methods that jazz improvisers use to respond to
offers from fellow musicians. For example, repetition is
important for establishing a motif, signaling understanding
or acceptance of a communicative intention, signaling
which performer is being lead by another, etc.
The most important response strategy, which forms the
lynchpin of the interaction-based authoring approach, is the
Proceedings of the Sixth International Conference on Computational Creativity June 2015 238
application of observationally learned input–response gesture
pairs. The agent observes the collaborator’s response
to its action and builds an association with parameters to
control its application. The use of observed action response
association leverages the collaborator’s more advanced
reasoning faculties in order to respond to some other interactor
in another context.
For example, when the agent learns to associate “waving”
gesture inputs to “bowing” gesture responses by
watching the collaborator execute “bowing” responses to
its own “waving” gestures, it can respond using the learned
association of “waving” and “bowing” gestures when a
new interactor “waves” at the agent. Of course in this example,
“wave” and “bow” gestures are actually clusters of
gestures with corresponding IDs to which the actual input
gestures match approximately (as mentioned earlier) – no
semantics of the words “wave” or “bow” are implied to be
understood by the agent.
A key assumption that the input response association is
based on is that the interactor’s response is always related
to the previous gesture from the agent and that there is always
some reason behind it. Both of these could well be
false, if the interactor gets bored and tries something completely
new for example. However, associations that are
seen more often are given positive reinforcement helping
to weed out weaker associations. This role-taking process
forms the key mechanism for learning by observation and
imitative learning within the Viewpoints AI installation
that helps it instantiate interaction-based authoring.
Another response strategy is the selection of emotional
reflex reactions to emotional content portrayed in the gestures.
There is an “emotional algebra” authored in the system
that responds according to a commonsensical set of
rules (e.g. responding to angry input gestures with angry or
fearful responses). This emotional algebra is rigid and uncomplicated
yet enables a simple short-circuit reflex response
system to quickly respond to portrayed emotionally
salient content within gestures.
An important response strategy is for the system to mimic
the interactor’s input gesture back to them. Mimicry /
repetition is important in facilitating smoother interactions
between people (Behrends et al. 2012). In contrast, a (trivial)
response strategy involves performing no response at
all, though this promotes a sense of uncertainty and is thus
discouraged unless as a last resort.
Another response strategy is for the agent to consider an
existing gesture and transform it. This creates a variation of
that gesture using dimensions or aspects of the Viewpoints
framework (eg. faster in tempo, smoother in movement,
adding repetitions, etc.). In addition, the system can use
acontextual functional transforms to add variety in the enacted
form of the gesture, such as reversing the direction of
movement, changing the limb in which movement takes
place, etc. See Jacob et al. (2013a; 2013b) for more detail.
A final response mode is for the agent to consider past
experiences from its episodic memory and choose a similar
gesture to bring into the new interaction context. This is
achieved using Soar’s episodic memory partial graph
matching capabilities in order to approximately match the
Viewpoints predicates of gestures and / or the direct predicate
space representation of their movements from other
interaction contexts to the current interaction context. This
is valuable to inject novelty into the current interaction
context. It uses the lower dimensionality (and higher level
of abstraction) of the Viewpoints predicate space to pick a
gesture that is roughly midway on a scale of novelty (from
completely identical to absolutely novel). This episodic
retrieval process is a key mechanism for case-based learning
within the Viewpoints AI installation that helps it instantiate
interaction-based authoring. Viewpoints predicates
form the index vocabulary for the case-based retrieval.
It should also be noted that this particular response
strategy introduces novelty to the creative experience, balancing
the predictability of other response strategies such
as the application of observationally learned patterns.
Action
The action module receives both predicate and joint space
gestures from the reasoning module and proceeds to create
the suitably transformed and rendered virtual agent embodiment
procedurally. The Viewpoints predicates associated
with the gesture being performed directly affect the visualization
(e.g. the energy of the agent’s movements control
the colors of the agent). The visual embodiment of VAI is
an anthropomorphic figure with a body composed of glowing
particles that keep to the bounds of the figure while
flying around probabilistically. In the current version of the
installation, the agent also has a region around the chest of
a corresponding interactor that glows with a diffuse red
colour in time to a rhythm if the agent has detected the user
moving to a beat. This has the visual effect of a glowing
heart beat that rises and falls with the interactor’s movements.
This was also designed to serve as a subtle form of
coordination between the two collaborators. For more detail
see Jacob et al. (2013a; 2013b).
Interaction-Based Authoring Beyond the
Viewpoints AI Installation
The Viewpoints AI installation instantiates the interactionbased
authoring approach to acquire knowledge from interactors
while attempting to provide a high quality subjective
experience to the interactors and support their creative
agency. It does this through knowledge acquisition of two
kinds. Firstly the case-based learning component stores all
gesture content it has seen or experienced in episodic
memory. Secondly it learns how to use these gestures to
respond to people by learning interaction patterns or pairs
of gestures from observing people and then imitating their
actions in a novel context. Finally the installation is situated
in a co-creative performative domain so that there is a
low bar for meaningful collaboration as well as to encourage
exploration of the interaction space due to player engagement
and acquire more knowledge as a result. The
approach differs from others by attempting to provide a
full-fledged co-creative experience right from the outset
Proceedings of the Sixth International Conference on Computational Creativity June 2015 239
without requiring explicit training or teaching phases.
The approach can be extended beyond the movementimprovisation
domain to increase the scalability of other
co-creative agents as well. The instantial gesture content
that the system learns using case-based learning can be
generalized to other types of response content, for example
strokes on a canvas or notes played on a synthesizer. Imitation
learning in turn can also be used to learn more general
response control knowledge. For example, the system
could learn patterns of strokes on a canvas or sequences of
notes. Currently the Viewpoints AI installation only does a
first order pairwise learning of gestures, however that
could be extended to higher order sequences of patterns.
Evaluation Methodology
The following sections describe an initial effort to evaluate
the success of the installation in addressing three main research
questions. RQ1: Can the interaction-based authoring
approach minimize the authoring bottleneck? RQ2:
Can usage of the interaction-based authoring approach
create high quality subjective experiences using the system?
RQ3: Can systems built with the interaction-based
authoring approach support collaboration with equal creative
agency (the extent to which a creative collaborator can
take meaningful decisions, make meaningful choices, and
affect the co-creative process or product)?
RQ1 was evaluated with formal analysis of authorial
leverage (Chen et al. 2009) as an initial attempt. More detailed,
pragmatic testing is required next. For the analysis,
three cases were compared: 1) a purely mirroring version
of the installation where the agent would only mirror the
movements of the interactor but not respond in any other
way, 2) a version of the installation with a pre-authored
tree of ‘plot points’ (pairs of input gestures and agent responses)
of arbitrary length and branching factor, and 3)
the full Viewpoints AI interactive art installation.
The RQ2 and RQ3 were evaluated using empirical quantitative
and qualitative methods in a pilot study (sample
size of 10). For the empirical evaluation, three different
experimental conditions were used. Condition 1 had only
mirroring of interactor movement as our baseline for comparison.
Condition 2 had mirroring of interactor movement
along with random movement responses, selected from a
library of prerecorded movements, when the participant
was performing rhythmic repeated movements. Finally,
condition 3 had the full response capabilities of the agent
available to respond whenever the interactor was making
rhythmic repeated movements. The order of the experimental
conditions was also randomized each time. In each
case, participants interacted with the experiment for 3
minutes, filled out two surveys administered online, and
then debriefed with a semi-structured interview. RQ2 was
evaluated using a set of validated survey instruments
measuring system usability, flow, and enjoyment of the
installation (Brooke 1996; Jackson et al. 2008; Vorderer et
al. 2004). RQ3 was evaluated using a set of validated survey
instruments measuring the creativity support index
(CSI) and effectance of the installation (Cherry and Latulipe
2014; Klimmt et al. 2007). The individual scales (excluding
the CSI) were administered online as part of the
IRIS Measurement Toolkit (Vermeulen et al. 2010). The
CSI had responses on a 7 point Likert scale while the IRIS
Measurement Toolkit used a 5-point Likert scale.
Results
Formal Analysis
The interaction-based authoring approach was designed
primarily to address the knowledge-authoring bottleneck.
Therefore the results of the formal analysis directly estimate
how much of an improvement is achieved using this
approach for acquiring knowledge within a co-creative
agent in the movement improvisation domain. For the three
experimental conditions described earlier (as with most
existing literature in the field) only the variability was used
as a factor for quality of the user experience. Therefore
authorial leverage was calculated as the ratio of the number
of unique experiences (variability) to the number of authorial
inputs involved in creating the system. In addition, a
few assumptions were made during the calculation. 1) In
order to compare the Viewpoints AI installation variants to
existing interactive narrative literature, the notion of plot
points was loosened to represent sequences of human –
agent movements or gestures. 2) The authorial inputs were
considered to be the sum of the number of gestures that
were authored prior to the start of the calculation in addition
to any manually authored transition rules between
them or between interactor gestures and agent responses.
For the first condition evaluating the purely mirroring
agent, it was assumed that both interactor and agent
movement responses were occurring simultaneously. Thus
a plot point would represent the interactor gesture and the
same agent gesture performed simultaneously. Therefore
the same sequence of N interactor gestures input to the
system would always return the same sequence of N gestures
back as responses. The authorial leverage is thus
nearly infinite since there is almost no prior manual authoring
of instantial content (authorial inputs near zero).
For the second condition with the pre-authored branching
tree of input gesture and agent response pairs, a tree of
average branching factor b and depth d would have at most
a total of (b
(bd
)−1) ⁄ (b−1) nodes or loose analogs to
plot points. Also, such a tree would have at most (bd
) linear
paths through it from root to leaf node representing unique
experiences. Therefore, the authorial leverage is roughly
(bd
)
(b−1) ⁄ (b
(bd
)−1). This function has an asymptotic
upper bound of 1 given any b or d.
Proceedings of the Sixth International Conference on Computational Creativity June 2015 240
Finally, the third condition with the full installation active
has the capability to select responses dynamically
based on the reasoning processes mentioned earlier. For
the first condition, the only possible response was to mirror
the interactor’s input gesture simultaneously. In the full
installation that capability is present (though mimicking
not mirroring) in addition to various other responses possible.
Therefore the number of unique responses possible for
a specific input gesture can only be higher.
For the same N input gestures as in the first condition,
the number of possible agent responses would be RN,
where R is the total number of unique responses to any one
input gesture given a set of response strategies (rather than
just mirroring). In the worst case this is 1 and RN reduces
to N possible unique agent responses. In the best case, this
becomes ΣRi
N where Ri represents the total number of
unique responses to one input gesture from the i
th response
strategy. Each of the response strategies is analyzed below.
For the “no response” response strategy, there is only
ever one agent response. For the “repeat input gesture”
response strategy, Ri
N is N since each input gesture returns
the same gesture as the agent output. For the “transform
input gesture” response strategy, Ri
N becomes 2(V+F)N,
where V and F are the number of Viewpoints dimensions
and functional transformations that the agent can use to
transform the input gesture into an agent response. In this
case 2(V+F) represents the cardinality of the power set containing
(V+F) elements. For the “emotion algebra” response
strategy, the number of emotionally appropriate
gestures available to respond with is dependent on the past
history of gestures learned by the agent. For a history of H
gestures with h appropriate gestures, that amounts to hN
possible responses. In the worst case, this reduces to repeating
the input gesture and Ri
N reduces to N. This is
justified by equating the emotional mirroring taking place
with emotional contagion (Hatfield et al. 1994). In the best
case, the entire history of gestures has the appropriate emotional
content and Ri
N becomes HN. For the “novel response
from episodic memory” response strategy, the exact
magnitude of Ri
N is difficult to estimate for the best case
since it is completely dependent on the past ordering of
learned gestures and received input gestures. However, the
lower bound for Ri
N is N since in the worst case, if no
gesture is found that is similar to the input gesture, the input
gesture is repeated as the agent output. Finally, for the
“learned interaction patterns” response strategy, given a set
of b learned responses on average for the right hand side of
each of N input gestures, the Ri
N would be bN.
Therefore RN or ΣRi
N for all the i response strategies in
the Viewpoints AI installation becomes 1 + N + 2(V+F)N +
N + N + bN or (1 / N + 3 + 2(V+F) + b)×N in the worst case.
This becomes (1 + N + 2(V+F)N + HN + ≥N + bN) or (1 / N
+ ≥2 + 2(V+F) + H + b)×N in the best case. Thus, the number
of unique experiences possible with the full installation
is much higher than in the first condition. The amount of
authorial input is equally minimal in the full Viewpoints
AI system. Therefore, since the authorial leverage for the
first condition is very large, the authorial leverage for the
full Viewpoints AI system is even larger. In addition, if
complexity were a factor in our calculation of authorial
leverage, it is visibly clear that the full installation has significantly
higher complexity in its decision-making and in
the user experience offered than the mirroring version of
the system.
Pilot Empirical Study
The aggregated results for both the IRIS Measurement
Toolkit and Creativity Support Index are presented in Figure
2. The system usability, flow, and enjoyment scales
were used to evaluate the system in terms of its ability to
produce high quality experiences for the user. The effectance
and creativity support index scales were used to
evaluate the ability of the system to co-create alongside the
participant with equal creative agency. The results show
that each of the experimental conditions did well, though
no statistically significant results could be obtained between
the different conditions potentially due to the small
sample size (sample size of 10). However, regardless of the
apparent lack of difference between the conditions, the
Figure 2: Results of Pilot Empirical Study. For descriptions of specific questions refer provided citations.
Proceedings of the Sixth International Conference on Computational Creativity June 2015 241
survey ratings for the third condition show clearly that usage
of the interaction-based authoring approach instantiated
within the Viewpoints AI installation can indeed both
create high quality subjective experiences for participants
interacting with the installation as well as support collaboration
with equal creative agency.
The semi-structured interviews were used to guide future
development and contained questions regarding feedback
about the experience, goals that users had while interacting
with the system, what they liked / disliked about the
installation, etc. The feedback was overwhelmingly positive,
with particular emphasis on the aesthetics of the
VAI’s visual representation, freedom of creative expression
felt by participants, amount of fun had by users, and
sheer “cool factor” of the installation. Some of the negative
feedback suggested that more was required to show that
the agent was actually doing something other than mimicking
the user. In addition, potentially indicating a miscommunication
of the design goals for the installation, it became
clear that some users felt like they should have been
able to control the agent’s actions to a greater degree. The
goals of the users varied depending on how many times
they had interacted with it and how inhibited they were.
The goals generally went from exploring the boundaries of
the system, to trying to get the agent to do certain reactions
/ responses, to trying to do novel interactions with the system
that hadn’t been tried before.
Discussion
The results given above help answer the three questions
used to evaluate the interaction-based authoring approach
instantiated within the Viewpoints AI installation. Using
the interaction-based authoring approach led to a significantly
higher authorial leverage (the ratio of variability of
the experience, or more generally the quality of the system,
to the amount of authorial input) than any pre-authored or
pure mirroring version of the installation. The pilot study
showed that the interaction-based authoring approach also
led to high quality experiences, as judged by the system
usability, flow, and enjoyment metrics administered. In
addition, the study revealed that the interaction-based authoring
approach was able to support collaboration with
equal creative agency using the effectance and creativity
support index metrics. However, it did not show a significant
difference in ratings between the three experimental
conditions for any of the survey metrics.
The lack of significant different between ratings for the
different experimental conditions could be because of a
number of reasons. Firstly, the study was conducted using
a very small sample size. However, given that the ratings
were so similar for all three, it is also possible that users
had difficulty distinguishing between the different conditions
in terms of the metrics used. Secondly, in terms of the
evaluation, users were blind to the nature of the experimental
condition as well as blind to the processes occurring
within the virtual agent. According to Colton (2008),
the process and the product are equally important to influence
evaluation of creativity within the system. Therefore
Turing-test style approaches to evaluation are found lacking.
This seems particularly true when the creative domain
is improvisation where participants evaluate the improvisational
experience / process.
The results (especially from the semi-structured interviews)
suggest that in order to improve the differential
ratings of the full Viewpoints AI installation to the other
conditions, the system’s actions and outputs should be
more noticeably different to highlight the system’s original
efforts better. This points to the requirement for a more full
featured list of implemented transforms (both Viewpoints
and functional transforms as well as gestural combination).
In addition, video analysis showed that novice users had a
hard time triggering the system’s rhythmic repeated
movement gesture segmentation mechanic. Thus current
efforts focus on replacing the existing gesture segmentation
algorithm with a more naturalistic automated gesture
segmentation algorithm from Kahol et al. (2004). Finally,
the experimental design is being refined to make the framing
more explicit and will be scaled up.
Conclusion
In conclusion, this paper introduced a hybrid approach to
knowledge authoring for co-creative systems called interaction-based
authoring. The approach incorporates ideas
from case-based learning and imitative learning, while emphasizing
incorporation into open-ended co-creative application
domains. This paper then presented an instantiation
of the interaction-based authoring approach within the
Viewpoints AI installation. The installation was then evaluated
in terms of the extent to which it mitigated the
knowledge-authoring bottleneck, produced high quality
subjective experiences, and supported equal creative agency.
Finally, the results of the evaluation were discussed in
terms of guiding the future iterative development and evaluation
of the installation.
<references_biblio/>
References
Behrends, A., Müller, S., & Dziobek, I. 2012. Moving in
and out of synchrony: A concept for a new intervention
fostering empathy through interactional movement and
dance. Arts in Psychotherapy, 39(2), 107–116.
Boden, M. A. 2003. The creative mind: myths and
mechanisms. Psychology Press.
Bogart, A., & Landau, T. 2005. The viewpoints book: a
practical guide to viewpoints and composition. New York:
Theatre Communications Group.
Brooke, J. 2013. SUS : A Retrospective. Journal of
Usability Studies, 8(2), 29–40.
Chen, S., Nelson, M. j, Sullivan, A., & Mateas, M. 2009.
Evaluating the Authorial Leverage of Drama Management.
In Proceedings of the AAAI 2009 Spring Symposium on
Interactive Narrative Technologies II (pp. 20–23).
Cherry, E., & Latulipe, C. 2014. Quantifying the Creativity
Support of Digital Tools through the Creativity Support
Proceedings of the Sixth International Conference on Computational Creativity June 2015 242
Index. ACM Transactions on Computer-Human
Interaction, 21, 1–25.
Colton, S. 2008. Creativity Versus the Perception of
Creativity in Computational Systems. In Proceedings of
AAAI Spring Symposium on Creative Systems (pp. 14–20).
Colton, S., & Wiggins, G. A. 2012. Computational
creativity: The final frontier? In L. De Raedt, C. Bessiere,
& D. Dubois (Eds.), 20th European Conference on
Artificial Intelligence: (Vol. 242, pp. 21–26).
Csikszentmihalyi, M. 1997. Flow and the Psychology of
Discovery and Invention. New York: HarperPerennial.
Csinger, A., Booth, K. S., & Poole, D. 1994. AI meets
authoring: User models for untelligent multimedia.
Artificial Intelligence Review, 8, 447–468.
Davis, N., Popova, Y., Sysoev, I., Hsiao, C.-P., Zhang, D.,
& Magerko, B. 2014. Building Artistic Computer
Colleagues with an Enactive Model of Creativity. In
Proceedings of the Fifth International Conference on
Computational Creativity (ICCC 2014). Ljubljana,
Slovenia.
Hall, M., National, H., Frank, E., Holmes, G., Pfahringer,
B., Reutemann, P., & Witten, I. H. 2009. The WEKA Data
Mining Software : An Update. SIGKDD Explorations,
11(1), 10–18.
Hastings, E. J., Guha, R. K., & Stanley, K. O. 2009.
Automatic content generation in the galactic arms race
video game. IEEE Transactions on Computational
Intelligence and AI in Games, 1(4), 245–263.
Hatfield, E., Cacioppo, J., & Rapson, R. 1994. Emotional
contagion. Current Directions in Psychological Science
(Vol. 2, pp. 96–99). Cambridge University Press.
Hecker, C., Raabe, B., Enslow, R. W., DeWeese, J.,
Maynard, J., & van Prooijen, K. 2008. Real-time motion
retargeting to highly varied user-created morphologies.
ACM Transactions on Graphics (TOG), 27(3).
Hoffman, G., & Weinberg, G. 2010. Gesture-based humanrobot
jazz improvisation. In Proceedings - IEEE
International Conference on Robotics and Automation (pp.
582–587).
Hsu, W. 2008. Two approaches for interaction
management in timbre-aware improvisation systems. In
Proceedings of the International Computer Music
Conference (ICMC). Belfast
Jackson, S. a, Martin, A. J., & Eklund, R. C. 2008. Long
and short measures of flow: the construct validity of the
FSS-2, DFS-2, and new brief counterparts. Journal of
Sport & Exercise Psychology, 30, 561–587.
Jacob, M., Coisne, G., Gupta, A., Sysoev, I., Verma, G. G.,
& Magerko, B. 2013b. Viewpoints AI. In Proceedings of
the Ninth Annual AAAI Conference on Artificial
Intelligence and Interactive Digital Entertainment
(AIIDE). Boston, MA.
Jacob, M., Zook, A., & Magerko, B. 2013a. Viewpoints
AI: Procedurally Representing and Reasoning about
Gestures. In Proceedings of the Digital Games Research
Association (DiGRA). Atlanta, GA.
Kahol, K., Tripathi, P., & Panchanathan, S. 2004.
Automated gesture segmentation from dance sequences. In
Proceedings - Sixth IEEE International Conference on
Automatic Face and Gesture Recognition (pp. 883–888).
Klimmt, C., Hartmann, T., & Frey, A. 2007. Effectance
and control as determinants of video game enjoyment.
Cyberpsychology & Behavior : The Impact of the Internet,
Multimedia and Virtual Reality on Behavior and Society,
10(6), 845–847.
Laird, J. 2012. The Soar cognitive architecture. MIT Press.
Latulipe, C., Charlotte, U. N. C., Huskey, S., Beasley, R.,
& Nifong, N. 2011. SoundPainter. In Proceedings of the
8th ACM conference on Creativity and cognition (pp. 439–
440).
MacDonald, L., Brosz, J., Nacenta, M. a., & Carpendale, S.
2015. Designing the Unexpected: Endlessly Fascinating
Interaction for Interactive Installations. In Proceedings of
the Ninth International Conference on Tangible,
Embedded, and Embodied Interaction - TEI ’14 (pp. 41–
48).
Magerko, B., Permar, J., Jacob, M., Comerford, M., &
Smith, J. 2014. An Overview of Computational Cocreative
Pretend Play with a Human. In Proceedings of
First Workshop on Playful Virtual Characters at the
Fourteenth Annual Conference on Intelligent Virtual
Agents. Boston, MA.
O’Neill, B., Piplica, A., Fuller, D., & Magerko, B. 2011. A
knowledge-based framework for the collaborative
improvisation of scene introductions. In Proceedings of the
4th International Conference on Interactive Digital
Storytelling (pp. 85–96). Vancouver, Canada.
Reidsma, D., van Welbergen, H., Poppe, R., Bos, P., &
Nijholt, A. 2006. Towards Bi-directional Dancing
Interaction. In Entertainment Computing - ICEC 2006 (pp.
1–12).
Thom, B. 2000. BoB: an interactive improvisational music
companion. In Proceedings of the fourth international
conference on Autonomous agents (pp. 309-316). ACM.
Tomasello, M. 2000. The Cultural Origins of Human
Cognition. Harvard University Press.
Vermeulen, I. E., Roth, C., Vorderer, P., & Klimmt, C.
2010. Measuring user responses to interactive stories:
Towards a standardized assessment tool. In Interactive
Storytelling (pp. 38–43).
Vorderer, P., Klimmt, C., & Ritterfeld, U. 2004. Enjoyment:
At the Heart of Media Entertainment. Communication
Theory, 14, 388–408.
Proceedings of the Sixth International Conference on Computational Creativity June 2015 243