=Paper=
{{Paper
|id=Vol-2600/paper13
|storemode=property
|title=Cognitive Twin: A Cognitive Approach to Personalized Assistants
|pdfUrl=https://ceur-ws.org/Vol-2600/paper13.pdf
|volume=Vol-2600
|authors=Sterling Somers,Alessandro Oltramari,Christian Lebiere
|dblpUrl=https://dblp.org/rec/conf/aaaiss/SomersOL20
}}
==Cognitive Twin: A Cognitive Approach to Personalized Assistants==
https://ceur-ws.org/Vol-2600/paper13.pdf
Cognitive Twin: A Cognitive Approach to Personalized Assistants
Sterling Somers, Alessandro Oltramari and Christian Lebiere
Psychology Department, Carnegie Mellon University, Pittsburgh, PA
Bosch Research and Technology Center, Pittsburgh, PA
{sterling@sterlingsomers.com,Alessandro.Oltramari@us.bosch.com, cl@cmu.edu}
Abstract enables advanced monitoring, predictive analytics,
and ubiquitous computing. The benefits of digital
This paper presents and evaluates an early pro- twins are evident in different domains, from indus-
totype of a cognitive twin: a digital reflection of
the user, intended to make decisions and carry
trial manufacturing (Kritzinger et al. 2018), where
out tasks on the user’s behalf. The intention of anticipating machine failures through digital simu-
the cognitive twin is to model the cognitive pro- lations can dramatically reduce maintenance costs,
cesses underlying the user’s decisions. To that end, to healthcare, where the growing number of inter-
we implemented the twin in a computational cog- connected health-monitoring systems (e.g, wear-
nitive architecture. The model specifically uses ables) is prompting providers to create personal-
an Instance-Based Learning approach to modeling ized digital solutions for their patients. Healthcare
human decision making that leverages the architec- is a perfect example of an area where IoT virtual-
ture’s memory mechanisms. The task of this pro- ization also factors the human in, e.g., by creating
totype model is to help organize a social gather- digital twins of the human body that can be used
ing. Data is generated in a discrete simulation that
we are developing to freely experiment with issues
for real-time remote-monitoring of physiological
ranging from environment structure to data avail- functions (Bruynseels, Santoni de Sio, and van den
ability. We implement two versions of the cogni- Hoven 2018). Beyond modeling the human as an
tive twin and show that it provides an effective per- organism, digital twins can expand to cover social
sonalized assistant with limited data availability. dimensions of human life, which are characterized
by our daily interactions with portable electronics,
digital services, social networks, etc. Accordingly,
Introduction in this paper we introduce the notion of cognitive
In the context of the Internet of Things (IoT), a (digital) twin, to highlight the key role that cogni-
digital twin is the virtual replica of a sensor-based tive mechanisms play in modeling human decision
connected device and of the processes that are as- making in the IoT digital space.
sociated with it. A network of digital twins can
then be conceived as a virtual space that mirrors The aim of the cognitive twin is to make the de-
the physical properties of IoT entities, and that cisions that a user would would make and to deal
with those aspects of life that can be handled over
Copyright c 2020 held by the author(s). In A. Mar- the internet. Much of our social lives are planned
tin, K. Hinkelmann, H.-G. Fill, A. Gerber, D. Lenat, R. over the internet and a cognitive twin would be
Stolle, F. van Harmelen (Eds.), Proceedings of the AAAI
2020 Spring Symposium on Combining Machine Learn-
able to handle tasks such as the planning of so-
ing and Knowledge Engineering in Practice (AAAI- cial events: who should attend, when it should take
MAKE 2020). Stanford University, Palo Alto, Califor- place, what activities should occur, etc. With an in-
nia, USA, March 23-25, 2020. Use permitted under Cre- crease in the use of IoT devices, we can imagine a
ative Commons License Attribution 4.0 International cognitive twin that not only knows when to turn
(CC BY 4.0). on the dishwasher, when to buy the soap, where to
buy the soap, but actually buy the soap, and even act as well (for instance the guest list might deter-
turn on the dishwasher. mine when the party can take place and vice versa)
We propose a hybrid data-driven and resulting in a potential complex process with mul-
knowledge-based approach. Because our con- tiple cycles of replanning.
text is human decision making, we implement If the cognitive twin is going to plan a party for
our digital twin in a computational cognitive a user, it has to know who that user would want
architecture. Our approach is to take advantage to attend. From a practical perspective, the twin
of existing knowledge ontologies, and structured likely has to learn about that user’s friendships,
procedural knowledge, but also learn decisions professional relationships, etc. One could imagine
from data. We believe this approach can bal- a twin that works by asking about attendees but,
ance domain generality while maintaining a low then, such a twin has less utility because it requires
implementation and data cost. significant effort on the part of the user.
The architecture we are working with is a hybrid From a practical perspective, there are a num-
cognitive architecture, ACT-R (Anderson 2007). ber of tasks we would expect the digital twin to
The ACT-R architecture, discussed more thor- learn and a number of tasks that are more prac-
oughly below, consists of both symbolic elements: tically represented as knowledge. Between these
declarative knowledge and procedural knowledge; two extremes of pure data-driven learning, and
as well as sub-symbolic elements that support sta- pure knowledge acquisition and authoring, there
tistical learning at a human-timescale. The tight in- are cases where we would expect a balanced hy-
tegration of these two levels of the architecture en- brid approach to be the most practical.
ables the combination of knowledge representation
Knowledge-based Party planning is a complex
and reasoning with statistical machine learning in
process. There are a number of sub-tasks that are
every aspect of decision making rather than arti-
required and the ordering of the sub-tasks is impor-
ficially segregating them into one or the other. In
tant. The knowledge of how to plan a party, how-
this project we use a cognitive modeling approach,
ever, would be difficult to learn from data because
Instance-Based Learning (IBL), in which past de-
the steps are not necessarily performed in any fixed
cisions are used to inform future decisions by di-
sequence, the patterns of occurrence are abstract,
rectly leveraging the mechanisms of the cognitive
and requires tying together different facets of in-
architecture.
formation that are rarely explicitly exhibited in be-
havior (therefore, would be difficult to collect data
Dinner Party Planning for). Instead those processes are typically commu-
We have chosen as our proof-of-concept scenario nicated at least in part explicitly as instructions,
to implement the planning of a dinner party. We in a process known as Interactive Task Learning
chose this scenario for a number of reasons. (Laird et al. 2017).
Perhaps one of the most obvious reasons for us- Data-Driven There are plenty of aspects to the
ing a party planning scenario is that it is such a party planning task that can be purely data-driven.
strong use-case for a cognitive twin. Party plan- For example, preferences in diets (e.g. vegetarian)
ning requires back and forth negotiation between are exhibited in behavior and, therefore, leaves a
the potential attendees with respect to scheduling trace that can plausibly be learned from. Other
and also requires information about the social dy- regularities like schedules, party sizes, and social
namics of possible attendees (avoid having people preferences, could all be plausibly learned from
that do not like each other at a party), food prefer- available data.
ences or diet restrictions, etc. Some of that infor-
mation is sensitive, such as personal likes and dis- Hybrid We believe that the party-planning sce-
likes. These are the the sort of tedious details that nario can benefit from a balanced, hybrid ap-
not only require time of the host but time of the proach. There are some aspects that can be and
potential guests. Planning a party requires a great should be data-driven (e.g. who your friends are),
deal of structured knowledge and systematic pro- while other aspects can benefit from both struc-
cess in order to make decisions about the various ture and data while others can largely be repre-
facets that need to be planned. Those facets inter- sented as structured knowledge (e.g. the process
of planning itself, diet categories, etc). An integra- chunks depending on how close their similarity is
tive framework that can combine both process as- to the requested pattern. Finally, chunk activation
pects is therefore required. Previous attempts have includes a stochastic component that makes mem-
been made at integrating cognitive architectures ory retrieval a probabilistic process.
and knowledge ontologies (Oltramari and Lebiere For this project we make use of the Instance-
2013). Based Learning methodology (IBL), which lever-
ages declarative memory by making decisions
ACT-R based on retrieval from memory of previous deci-
sions rather than through heuristics or production
ACT-R is a hybrid (symbolic, sub-symbolic) com-
rules (Gonzalez, Lerch, and Lebiere 2003).
putational cognitive architecture. The components
of the architecture represent a unified, neurally in-
spired theory of cognition (Anderson 2007). The Data / Simulation
architecture is comprised of a set of modules, each In order to collect data for a cognitive twin we
of which are constrained by the biological pro- have chosen to implement a discrete simulation.
cesses and limitations of the brain. While the archi- The simulation is currently under development but
tecture does include modules for vision, audition, produces sufficient data to test the cognitive twin.
and motor activity, the core of the system is com- Our aim, currently, with the simulation is a vehi-
prised of a central production system and a declar- cle for generating data in liue of human data. The
ative memory. simulation represents human agents going about
The central production system represents proce- their lives, while personal data is being collected
dural knowledge (skills). Each production ‘fires’ and stored. The agents in the simulation are dis-
in response to particular patterns or ‘chunks’ of in- tinct from cognitive agents. In a sense, the simu-
formation held in buffers that constitute the cog- lated agents represent ‘real’ human beings (in liue
nitive architecture’s working memory. Skill acqui- of human data) and the cognitive agents represent
sition, response competition, and multitasking ca- their cognitive twins. Long-term, we aim to expand
pacity are among the cognitive capacities modeled on the simulation, making the ‘lives’ of the agents
in the production system. more veridical, the properties of the network to re-
Declarative memory is a long-term knowledge flect human social networks, etcetera. Also, as we
storage and retrieval system. Declarative knowl- develop the simulation further, we hope to make
edge is represented as chunks, that are proposi- the data collected about users reflect more closely
tional compositions of other chunks or more basic the quality of data potentially collectable from real
values. Each chunk in declarative memory has an humans. In the following we describe the simula-
activation that modulates its retrieval. The history- tion in its current form, including aspects we are
based activation component is called base-level ac- not currently using in development and evaluation
tivation, which reflects the chunk’s recency (when of the cognitive twin.
it was encoded or recited) and frequency (number The simulation creates a population of n agents
of times it has been recited) to capture regulari- that carry out both scheduled (e.g. work) and un-
ties of human behavior such as the power laws of scheduled events (socializing, eating) over a period
practice and forgetting. Furthermore a chunk’s ac- of days. For this paper we simulate 720 days of
tivation is affected by spreading activation, such simulation. As our simulation is generated stochas-
that related chunks (e.g., context information) can tically, for this analysis we set a random seed, and
boost the activation of chunks in memory, depend- all analysis is performed on a single, generated
ing on their degree of co-occurrence. Chunks can world. Relevant information about the activities in-
be retrieved through a pattern-matching process cluding the time, the event type, and specific de-
in which a sub-pattern is requested and a com- tails about the event are recorded in data files.
plete pattern is (potentially) returned. Importantly, The simulation is controlled by a number of pa-
memory retrieval in ACT-R is not perfect. Chunks rameters. To generate a ‘world’, a user specifies a
with a low activation can fail to be retrieved population of n size. The simulation then generates
(forgetting) while a partial-matching process can families, where the member-size is controlled by a
generalize to similar but not perfectly matching weighted distribution parameter. Although family
dynamics is not captured in the simulation at this of which is a free parameter. Values for the social
early stage of development, we have designed the links are real numbers ranging from -1 to 1. The
simulation to accommodate more complex interac- social links are both incoming and outgoing and,
tions in the future. as a result, can be asymmetric (i.e., one agent may
Since our main proof-of-concept case is the favor another, but may not be favored in return).
party planning scenario, we have developed the We consider a negative connection between per-
simulation such that activities include social activ- son A and person B to represent tension or nega-
ities, to track positive and negative social interac- tive feelings of A towards B. Large negative values
tions; food preparation, to track diet restrictions; represent strong dislike. Positive values represent a
and daily activities, to track an individual’s sched- positive relation and we define values greater than
ule. 0.75 as friendship. When the simulation is running,
the social links are used when setting up social in-
Schedule The simulation ticks by the hour teractions in the simulation. The social links be-
and, currently, events have a 1-hour resolution. tween agents change over time as they interact with
Implementation-wise, there is no known restric- one another (see Figures 4 and 5). Roughly, friends
tions on resolution. Days of the week are soft- become more likely to interact and enemies less
coded into the simulation (by dividing hours of likely, as described below.
simulation time) and are used to design the sched- Social interactions (parties) are created during
ule. Each individual is assigned a static work and run-time at the start of each simulation day. Social
sleep schedule, reflecting real-world schedules (9- interactions, on any given day, are potentially cen-
5 Monday to Friday). Eating can occur nested tered around randomly selected individuals. The
within another event (i.e. eating at work), and the parameter s is set as a proportion of the total
start and end times are chosen from a weighted population and we used a value of 25 to ensure
distribution of pre-selected hours. For example, that we generate sufficient data for testing. This
breakfast can occur between 6 am and 9 am with value was qualitatively assessed to maintain steady
different weights for different hours (free parame- world statistics and, as described below, we evalu-
ter). ate the model with different ranges of history. Once
At generation, agents are assigned scheduled centers (hosts) are selected, attendees are selected
events. These events are mandatory and are called to attend.
‘sports’, ‘clubs’, and ‘hobbies’. The intention is
to create regularly-occurring events that should Attendees to each social interaction are selected
be considered when scheduling a social event. from the population such that their likelihood of
The events are selected from a weighted random being selected is weighted by the outgoing social
list, with the possibility of repeats reduced for link from the ‘center’ to the potential attendees.
each by re-weighting the events. For example, the The outgoing links between the center and the at-
likelihood of sports being selected is reduced if tendees are transformed to produce a probability
sports is selected once. There is also a placeholder distribution, and non-uniform random samples are
‘None’ event. Events are selected, scheduled, and selected. A free parameter which is set to con-
re-weighted; the process repeats until an event fails trol the minimum social interaction connection re-
to be scheduled on randomly selected days or until quired to attend is set to a minimum value of -0.1,
the None event is selected. It is, therefore, possi- so that people with slight dislike for each other
ble (though low probability) that some people will might attend. A small bias is also added to the
have no activities for which they are committed. weights of connections above the friend threshold
(how we separate friends and close friends). The
All times, event types, and event details are
bias is set to 0.3 and the friend threshold is set to
recorded in an individual’s log-file as well in a
0.75. We use the values to help control the patterns
global log-file.
of interactions. These values were set qualitatively
Social Interactions When a ‘world’ is gener- to produce steady world statistics.
ated, social links between every pair of individual Not all centers will result in a social interaction.
agents in the simulation are generated randomly Once attendees are picked, the social interaction
from a truncated normal distribution, the variance then has to be scheduled. Scheduling may fail due
to firm commitments by attendees. The minimum with the gray area showing one standard deviation.
size of social interactions are selected randomly As can be seen from those two figures, the rela-
from a truncated normal distribution between 4 and tionships in the simulation are fairly stable. These
10. The shape of that distribution is also a free can be modified using the free parameters but we
parameter. If an insufficient number of people are have chosen to keep the relationships rather stable,
able to attend, the social interaction is canceled. as we assume this reflects the nature of relation-
Furthermore, social interactions are resolved in ships in the real world.
a queue, which is ordered by the original selection
of centers. Attendees that are in the queue for mul-
tiple events (lots of friends) may become busy, po- Figure 1: Social Connections
tentially becoming unavailable for events further
down the queue.
During a social interaction, all agents in atten-
dance receive an interaction score with the other
guests. The interaction scores are selected from a
truncated normal distribution between -1 and 1,
and the shape of the distribution is a free param-
eter. The total score that an interaction receives is
a linear combination of: the random score (-1 to Figure 2: Friendship With Party Centre
1), the link from person A to person B, and the
link from person B to person A. For this paper,
those factors are all weighed the same as we have
yet to analyze the impact of these values on the
social dynamics across time. Values above 1 or
below -1 are truncated. As a result, it is possible
for the quality of relationships to change: friend- Figure 3 represents the average connection be-
ship links between agents may go from positive to tween all guests in attendance at the social inter-
negative after interactions and, likewise, negative actions. Note that the simulation does not try to
links can become positive after interaction. Nega- maximize all the social links of a social interaction,
tive links becoming positive are most likely to oc- as attendees are chosen stochastically, weighted by
cur in social interactions not centered by the agent the outgoing social link from the party center.
with outgoing negative links. Either the attendees
show up at a mutual friends party or there is an im-
balance in their outgoing and incoming links. Al- Figure 3: Average Connection Between Party
though crude, this is meant to capture the influence Guests
of past interactions. The time of interaction and the
scores between attendees is recorded in the log-file
for the individual as well as a global log-file.
To get a sense of what the socialization looks
like, we have plotted some basic statistics of the
interactions over 720 days of simulation.
Figure 1 shows the percent of people who have a To have a sense of the social networks over time,
zero or greater (i.e., positive) outgoing social link Figures 4 and 5 represent the steadiness of people
between them (blue), with the shaded area (gray) that have positive connections with a target over
showing one standard deviation. In orange is il- time. It shows the percent of change of people in
lustrated the percent of people who are above the the positive social network per day (blue), as well
‘close-friend’ threshold, with the orange shaded as the change since time zero. Figure 5 shows the
area showing one standard deviation. Figure 2 same stats (blue vs orange) but where the members
shows the percent of people at a party who are of the social network are above the friend threshold
close friends with the party center (or host) (blue) (0.75).
integrate with simulations, and ultimately to de-
Figure 4: Positive Social Network Over Time velop and distribute on portable platforms with
limited footprint and computing power, as well
as real-time performance and network-size scaling
requirements. The key equations leveraged were
those for declarative memory, as it reflects trade-
offs in recency vs frequency in generalizing pat-
terns of user activity, as well as generalization in
continuous spaces such as high-dimensional vec-
tors used in distributional semantics.
Figure 5: Friend Social Circles Over Time In the current model, batch activity data is
loaded into memory. We will discuss later the type
of protocol that could be used to request informa-
tion from peer cognitive twins, and their impact
on the algorithm used by the main cognitive twin
planning the activity.
Finding the most compatible set of guests re-
lies on ratings of past joint social activity. Those
Diet and Food are represented as chunks in memory consisting of
Individual agents in the simulation are either om- the user who provided the rating, the agent with
nivores, vegetarians, or vegans, and consume food whom that user interacted, and the rating of the in-
according to their diets. The ontology of diets is teraction. Those ratings are not necessarily sym-
modeled from ConceptNet (Liu and Singh 2004) metrical, so each pairwise interaction might result
and the foods in the simulation are, therefore, in two separate chunks with distinct values. Those
structurally consistent with that knowledge. Build- memory chunks have associated a base-level ac-
ing the structure into the simulation is important tivation that reflect the recency (and frequency
for what we are trying to achieve, as we expect that if the same rating between the same agents has
a learning system should recover that structure. We been provided multiple times) of the rating. Each
propose that a system like the cognitive twin can chunk activation can therefore be interpreted as
take advantage of existing knowledge graphs like the relative importance of that rating among com-
ConceptNet to take advantage of known structure peting ones, and factored accordingly by memory
but should also be able to learn that structure, or retrieval processes. Specifically, the blending re-
any combination thereof. In this work, we are not trieval process (Lebiere 1999) is used to produce
not learning diet restrictions but intend to do so for a consensus estimate of social rating between two
future work. agents by retrieving and aggregating individual rat-
ings weighted according to their activation. These
social compatibility ratings are used in a greedy al-
Cognitive Twin Model gorithm that starts with the central user organizing
At a high-level, the cognitive twin is intended as the gathering. The rating of all potential guests are
an automated decision maker. It uses your data to evaluated against the current set of invitees (start-
make the kinds of decisions you would make. In ing with the host) and the guest with the highest
the party-planning scenario, your cognitive twin is rating is selected. The process is repeated until the
intended to do some of the leg work of party plan- guest list is full. It is possible that this process does
ning for you. The cognitive twins, described below, not yield an optimal outcome because of the se-
use the personal data from the simulation agents to quential selection process could lead to a local op-
plan their next party for them (given an existing timum. An optional second phase (not yet imple-
history of social interactions). mented) would consist of computing the social rat-
The cognitive twin was developed in ACT-UP ing of each guest with respect to all other guests,
(Reitter and Lebiere 2010), a toolkit implemen- then swapping out the guest with the lowest rating
tation of ACT-R developed to make it easier to if another guest with a higher rating with the re-
maining guests could be found. This process would according to two measures, displayed in Figure 6.
be repeated until no such guest would be found. The first measure is the average percentage of each
The time scheduling process operates in a simi- party that is made of close friends to the host. The
lar manner. All considered time slots are generated centered version of the cognitive twin (69.75%)
by a process that combines user input with knowl- outperforms the distributed version of the cogni-
edge constraints (e.g., social events are not sched- tive twin (56.75%) because it is more narrowly fo-
uled in the middle of the night, or during regular cused on optimizing compatibility with the host
work hours). Past and scheduled activities of each social connections rather than mutual guest con-
potential guest are represented in memory, each as nections. They both outperform the simulation al-
a chunk including identity of the agent, day, time gorithm (49.25%) because of the built-in stochas-
and nature of activity. For each potential time slot, ticity of that algorithm.
the availability of each guest is generated through The second measure is the average connection
a blending process that generalizes from past and strength between each pair of guests (including
future activities to infer regularities such as typi- the host). By that measure, the distributed version
cal activities for a particular day/time slot without of the cognitive twin (0.680) outperforms the cen-
needing to represent them explicitly, such as by ac- tered version of the twin (0.464) because it factors
cessing a calendar. The time slot with the highest the mutual guests connections instead of strictly
expected availability is selected. An optional pro- those from the host. Both outperform the simula-
cess (not yet implemented) would trigger another tion algorithm (0.358) because of its combination
guest selection process if enough guests cannot at- of narrow focus on the host and stochasticity.
tend at a given time slot.
Results Figure 6: Ratio Of Friends and Average Score at
Social Event
The model was evaluated against a simulation of
100 different agents run for two years of simulated
time. At the end of that time, the model consists of
100 distinct cognitive twins, each loaded with the
history of its specific agent. That results in about
16,000 chunks in the memory for each twin, about
150 per week, corresponding to a mix of social
events and interaction values.
The model is run to generate a party for each
of the 100 agents. Two versions of the model were
run. Recall, this simulation is a set random seed
of possible worlds. The distributed version, as de-
scribed above, factors in the social interactions The data requirements of the model were also
with all current scheduled guests when adding a evaluated. For that purpose, the model was run in
new guest. The centered version, however, only the same settings as above, but with a more limited
selects guests based upon their social interactions history of social interactions. The length of history
with the host, as is the case for the simulation. was varied from 1 week to the full 2 years. Re-
The parameters of the model are set to provide sults for both measures describe above are plotted
fairly accurate performance, including a mismatch in Figure 7.
penalty parameter (scaling the partial matching For both model versions and both measures, per-
process) of 5.0 and an activation noise parame- formance actually increases with decreasing his-
ter (scaling the stochastic activation factor) of 0.1 tory length. This results because changing social
as well as the usual time-based decay parameter connections means that historical data becomes in-
(weighting recency against frequency) of 0.5. creasingly inaccurate over time. Cognitive mech-
This evaluation focuses on the guest selection anisms such as blending are usually able to gen-
process rather than the time scheduling. The mod- eralize quite well with limited data. The optimum
els are evaluated against the simulation baseline history length is between 2 and 4 weeks depending
upon the measure and model version. privacy of the guests. Developing an infrastructure
that can optimize those trade-offs while safeguard-
ing user data is an essential goal of the cognitive
Figure 7: Ratio of Friends and Average Connection twin approach.
by Weeks of Data
References
Anderson, J. R. 2007. How Can The Human Mind
Occur In The Physical Universe? New York, NY:
Oxford University Press.
Bruynseels, K.; Santoni de Sio, F.; and van den
Hoven, J. 2018. Digital twins in health care:
ethical implications of an emerging engineering
paradigm. Frontiers in genetics 9:31.
Corbett, A. T., and Anderson, J. R. 1995. Knowl-
edge tracing: Modeling the acquisition of procedu-
ral knowledge. User Modelling and User-Adapted
Discussion Interaction 4(4):253–278.
Parameters of the model were selected to be com-
Gonzalez, C.; Lerch, J. F.; and Lebiere, C. 2003.
patible with standard values used to match human
Instance-based learning in dynamic decision mak-
performance in ACT-R models (see models and
ing. Cognitive Science 27(4):591–635.
publications on the ACT-R web site at http://act-
r.psy.cmu.edu) and have not been optimized for Kritzinger, W.; Karner, M.; Traar, G.; Henjes, J.;
this specific simulation environment. For instance, and Sihn, W. 2018. Digital twin in manufacturing:
a larger time decay parameter would allow com- A categorical literature review and classification.
bining the accuracy of a longer history without the IFAC-PapersOnLine 51(11):1016–1022.
penalty of overly relying on obsolete information. Laird, J. E.; Gluck, K.; Anderson, J.; Forbus, K. D.;
Another method for evaluating the model is Jenkins, O. C.; Lebiere, C.; Salvucci, D.; Scheutz,
model tracing (Corbett and Anderson 1995). That M.; Thomaz, A.; Trafton, G.; Wray, R. E.; Mohan,
approach, developed to parameterize general mod- S.; and Kirk, J. R. 2017. Interactive task learning.
els to match individual behavior traces, would con- IEEE Intelligent Systems 32(4):6–21.
sist of running the cognitive twin alongside the Lebiere, C. 1999. The dynamics of cognition:
simulation, making predictions at each step that An act-r model of cognitive arithmetic. Kognition-
could be compared to simulation events, as well swissenschaft 8:5–19.
as gradually learning an increasingly long history
Liu, H., and Singh, P. 2004. Conceptnet—a prac-
of its users.
tical commonsense reasoning tool-kit. BT technol-
Improved generalization over limited data could
ogy journal 22(4):211–226.
also be obtained by relying on similarities between
agents rather than treating agents as distinct sym- Oltramari, A., and Lebiere, C. 2013. Knowledge
bolic entities. For instance, similarities between in Action: Integrating Cognitive Architectures and
agents could be set to reflect the degree of overlap Ontologies. Berlin, Heidelberg: Springer Berlin
in their activities or social connections. Heidelberg. 135–154.
The structure of the various versions of the cog- Reitter, D., and Lebiere, C. 2010. Accountable
nitive twin also has implications for communica- modeling in act-up, a scalable, rapid-prototyping
tion protocols between cognitive twins. For in- act-r implementation. In Proceedings of the 10th
stance, the centered version of the model does not International Conference on Cognitive Modeling,
rely on any social information from other cogni- ICCM 2010, 199–204.
tive twins, while the distributed version does re-
quire information from invited guests regarding fu-
ture potential guests. This leads to a trade-off be-
tween overall quality of the social gathering and