=Paper=
{{Paper
|id=Vol-2558/short1
|storemode=property
|title=Anticipatory Thinking in Cognitive Architectures with Event Cognition Mechanisms
|pdfUrl=https://ceur-ws.org/Vol-2558/short1.pdf
|volume=Vol-2558
|authors=Steven Jones,John Laird
|dblpUrl=https://dblp.org/rec/conf/aaaifs/JonesL19
}}
==Anticipatory Thinking in Cognitive Architectures with Event Cognition Mechanisms==
https://ceur-ws.org/Vol-2558/short1.pdf
Anticipatory Thinking in Cognitive Architectures
with Event Cognition Mechanisms
Steven J. Jones, John Laird
University of Michigan
2260 Hayward Street
Ann Arbor, MI 48109-2121
{scijones,laird}@umich.edu
Abstract Including mechanisms for human-like event cognition
There is no comprehensive theory for how anticipa-
in cognitive architectures can provide such an integration
tory thinking capabilities emerge from cognitive pro- to better understand how human AT functionality emerges
cesses. Event cognition describes some human anticipa- from cognitive processes. Cognitive architectures are theo-
tory thinking capabilities, but is not integrated with gen- ries for the fixed computational mechanisms that underlie
eral theories of cognition. We use Event Segmentation cognition. While many architectures initially made differ-
Theory to motivate a theoretical account for how the ent and conflicting assumptions or described isolated aspects
Soar cognitive architecture and the Common Model of of cognition, over time a consensus has emerged. This con-
Cognition can be extended to support event cognition, sensus is formalized through the Common Model of Cog-
and in turn account for anticipatory thinking processes nition, which is a theoretical specification of the computa-
and reasoning. Current cognitive architectures appear to tional processes underlying cognition (Laird, Lebiere, and
require additional mechanisms to create computational
models implementing this theoretical account.
Rosenbloom 2017). Extending the Common Model to in-
clude event cognition provides a model for how AT is real-
Anticipatory thinking (AT) is an emergent cognitive func- ized in human-like cognition.
tionality. AT has been described as the ability to proactively
guide attention and take preparatory action (Klein, Snow- Event Segmentation Theory
den, and Pin 2011). We propose the development of a cog-
With support from observations of human behavior (Eisen-
nitive theory of human AT functionality based on the com-
berg, Zacks, and Flores 2018), memory (Sargent et al. 2013),
bination of event cognition research and research on cogni-
and brain activity (Baldassano et al. 2017), Event Segmenta-
tive architecture. A general cognitive theory of human AT
tion Theory (EST) has become the dominant theory for event
could predict how human AT changes as a result of specific
cognition. It provides the process model depicted in Figure
training, experience, environments, and/or access to differ-
1. The theory is that humans understand their experience in
ent kinds of knowledge. Additionally, with a theory of how
terms of discrete segments of experience called events (Za-
AT is realized in human cognition, AT can be implemented
cks and Swallow 2007). Similarly to AT as a form of sense
in artificial systems with similar computational structure.
making, the segmentation of experience into events is con-
Event cognition research studies the human ability to per-
sidered part of ongoing comprehension.
ceive, understand, and remember everyday events (Radvan-
sky and Zacks 2014). This research has the potential to pro- The theory proposes that humans use mental models for
vide insight into how AT is realized in human cognition. events. These models are divided into event models de-
Event cognition research hypothesizes that humans simul- scribing specific situations and event schemas describing the
taneously perceive and predict events to guide attention in
real-time and also use the same mental representations both
for guiding action and comprehending the actions of others
(Richmond and Zacks 2017). We propose that these proper-
ties of human event cognition are also core aspects of AT.
While there is a neuro-physiological account for some as-
pects of event cognition (Franklin et al. 2019), event cogni-
tion is not currently integrated with a general theory of cog-
nition. Such an integration would allow an understanding of
how additional cognitive processes enable the decision mak-
ing and response preparation necessary for functional AT.
Copyright c 2020 for this paper by its authors. Use permitted un-
der Creative Commons License Attribution 4.0 International (CC Figure 1: The structure of the Event Segmentation Theory
BY 4.0) process model. (Reproduced from (Zacks et al. 2007).)
commonalities for a given type or class of event (Radvansky
and Zacks 2011). A mental model is an abstract representa-
tion of a situation used for reasoning. It is composed of in-
dividual elements (such as entities and relations) that can be
rearranged and that are grounded to perceptual representa-
tions. An example of a mental model is representation of an
animal in terms of an arrangement of body parts. An event
model is a mental model for a specific event. Event mod-
els are entities and relations describing a particular span of
space and time, but usually in a single location. Event mod-
els include labels, spatial relations, and relations that con-
vey a temporal ordering. An event schema is a mental model
for a class of event models, where multiple event model in-
stances belong to the same event schema. As an example, a
specific memory for having watched a film is an event model Figure 2: The structure of the Soar Cognitive Architecture.
while an understanding for how a visit to the theater gener-
ally proceeds is an event schema. Event models are created
by specializing event schemas to a set of observations. Both
representations contain causal relations between changes. bility is to explore the integration of EST with the Common
During everyday tasks, event models predict changes to Model of Cognition. Unfortunately, due to its abstract na-
the current situation and guide perceptual processing (Za- ture, the Common Model does not provide the level of de-
cks et al. 2007). For example, predictive-looking describes tail necessary for implementation of running computational
the human behavior of looking to where changes are ex- models. Instead, we use Soar, an architecture consistent with
pected to occur. This ability is diminished near the bound- the Common Model, as a model for how cognitive architec-
aries between events (Eisenberg, Zacks, and Flores 2018). tures (and, more abstractly, the Common Model) can realize
We use the term working event model to refer to event mod- event cognition functionality.
els used to describe the current situation (Radvansky 2012).1 Soar models cognition as a series of deliberate actions that
As shown in Figure 1, when a prediction fails, a prediction perform reasoning steps, retrievals from long-term memo-
error is detected and signals that the working event model ries (episodic or semantic), or motor actions (Laird 2012).
does not match the situation. In this case, humans create a The actions are initiated by knowledge retrieved from pro-
new event model to interpret the situation by retrieving an cedural memory, based on the contents of working memory.
event schema that matches to recent sensory input and cre- Working memory contains a symbolic representation of the
ating new expectations. current situation (derived from perception and internal rea-
The EST process model focuses on descriptions of on- soning), current goals, and intended actions. A cognitive cy-
going perception for a directly-experienced event. Anticipa- cle, which consists of processing input, a deliberate decision,
tory thinking appears to require reasoning that includes ex- and output to the motor system, maps onto approximately 50
pectations for future events beyond short-term expectations ms of human behavior. This low-latency perception and ac-
for the currently-experienced event, which motivates our ac- tion cycle provides reactivity to both changes in perception
count of event cognition using a cognitive architecture. and knowledge retrieved from long-term memory. Complex
behavior arises from a sequence of cognitive cycles. Figure
Event Cognition in Soar 2 shows Soar’s structure.
To theoretically model event cognition phenomena using
Cognitive architectures are computational models for the Soar, we map the different mental representations specified
fixed mechanisms and processes that underlie cognition. by EST to Soar’s memory systems. Both event schemas and
These architectures act as theories for the functionality pro- event models contain relational information and lack percep-
vided by different memory systems and cognitive processes. tual detail. They contain entities and relations for describ-
They also can be used to implement artificial cognitive sys- ing an event, which are directly supported by the memory
tems. However, these architectures do not currently exhibit systems of Soar. We assume that event schemas and mod-
the event cognition functionality found in humans. els have relations depicting changes over time, allowing for
EST specifies representations of events, but does not de- representation of future state using these relations. The as-
scribe how (together with other mental models) they are en- sociation of event schemas and event models to the memory
coded, stored, or retrieved from memory systems, nor the systems of Soar is depicted in Figure 3.
reasoning processes that use them. Cognitive architectures
can extend event cognition theory by including the memory The working event model is grounded to ongoing action
systems and reasoning that EST lacks. An intriguing possi- and perception. It is also used in reasoning about the current
situation. To provide this functionality, it must be composed
1
In EST, event models are hierarchical. Thus, a single work- of working memory structures and also representations of
ing event model describes the current situation, but it can contain perception and action. Figure 3 depicts the working event
nested sub-events that are event models for smaller space and/or model within working memory, but specifically as including
shorter segments of time. the representations for perception and control.
of AT processes. Anticipatory thinking is associated with
Reasoning
Event Schemas
Previous three distinct processes. These processes are “recognition of
Strategies Event Models a situation based on current cues derived from previous ex-
perience, extrapolation of a system state to a different state,
and construction of a mental model of the system based
Retrieved Event Models on variable evidence” (Geden et al. 2019). These processes
have also been referred to as “pattern matching,” “trajectory
Working Event Model tracking,” and “convergence,” respectively (Klein, Snowden,
and Pin 2011). To explain how the proposed model supports
these processes, consider the following scenario:
You observe someone else printing papers. You recog-
nize that they are likely creating exam packets. You in-
fer that they will need to staple these papers together.
You observe that they do not have a stapler. You fetch a
Figure 3: Soar’s memory systems populated with event cog- stapler to help them achieve their goal.
nition knowledge.
The process of recognition uses cues from the present sit-
uation to retrieve knowledge for similar situations from the
In contrast, event models representing prior situations and past. An example of pattern matching is the recognition of
event schemas require long-term storage and need to be someone in the act of creating exam packets by observing
stored within the long-term declarative memory systems in them in the copy room printing papers. In our model of event
Soar. Event models have been hypothesized as the episodes cognition, there are two forms of recognition. When there is
of episodic memory in humans (Ezzyat and Davachi 2011). knowledge for a type of event that generalizes multiple spe-
In Soar, they naturally belong in episodic memory as a re- cific events, this is stored as an event schema in semantic
sult of automatic storage of working event models in work- memory. Recognition can take the form of retrieval of an
ing memory. In Soar, semantic memory provides a means event schema from semantic memory based on the cue that
to knowledge independent of the exact situation in which it someone is printing papers. However, if such knowledge is
was learned, and thus, as shown in Figure 3, is where event not available, there can also be knowledge of a specific simi-
schemas are stored. To be used in reasoning, these stored lar event from the past. Recognition can thus also result from
event schema and model representations must be retrieved retrieval of an event model from episodic memory.
into working memory. The process of extrapolation involves not only predict-
By assigning the mental representations described by EST ing future states, but also guides action in conjunction with
to the memory systems of Soar, we can replicate the EST predictions to realize a desired future state. An example
process model using the mechanisms available in Soar. Per- is catching a ball, but extrapolation also refers to narra-
ception feeds into working memory and cues for retrieval of tive understanding and prediction (Klein, Snowden, and Pin
an event schema from semantic memory. This event schema 2011), not only to the ongoing real-time prediction of per-
can then be grounded to perception and action to form a ception performed by working event models. An exam-
working event model that is compared to perception. Then, ple of such extrapolation is creating the expectation that
the ways in which event models within working memory can someone will need a stapler. They may not currently need
be used for reasoning depends on the reasoning strategies a stapler to proceed, but they are doing a task which in-
within procedural memory. volves later use of a stapler. In human event cognition, event
models are also used to simulate future events consistent
The contribution of this model is that reasoning is not
with episodic future thinking (Richmond and Zacks 2017;
limited to only performing the EST process model loop of
Szpunar, Spreng, and Schacter 2014). Additionally, event
maintaining a working event model to describe the present.
models can be used to understand indirectly experienced
In situations where an agent performs long-term planning
narratives and situations (Radvansky and Zacks 2011). Us-
(representing and reasoning about the future beyond the cur-
ing this as inspiration, in our model extrapolation results
rent event), additional event models that represent expected
from the structure and contents of the retrieved schema.
distant future states are retrieved into working memory to
When the schema for creating exam packets is retrieved,
be used in reasoning. Also, previous event models can be
this knowledge includes causal relations and expected future
retrieved for comparison of a specific previous situation to
state. Because this future state is currently retrieved to work-
the present for case-based reasoning. These additional rea-
ing memory, it is available for reasoning despite this state not
soning capabilities arise from the general cognitive mech-
yet having occurred. This ability to use a representation of
anisms available for using and manipulating stored event
future state in current reasoning provides AT extrapolation.
knowledge.
Construction is the ability to reason about and create men-
tal models for a situation. We model this as reasoning about
Cognitive Modelling of Anticipatory Thinking the conditions and connections between events. When rec-
The cognitive mechanisms available for manipulating event ognizing that someone is involved in a task which will re-
representations in Soar’s memory systems enable modelling quire a stapler in the future, we have the ability to integrate
an event model for delivery of that stapler with an event Backcasting is reasoning that finds ways or paths to a par-
model for someone else’s future use of the stapler, allowing ticular future state. Means-ends analysis performs similar
us to coherently model both our delivery and the satisfac- reasoning in Soar. In order to determine a path to a future
tion of their task. This ability to evaluate how our planned state, reasoning proceeds backwards from the future state,
actions will impact external events in the future is an exam- attempting to create a plan of actions that can achieve the
ple of conditional AT which uses causal relations between future state, while recursively attempting to achieve the pre-
event models. Soar supports construction through relations conditions of those actions until a path is found with precon-
in working memory that link different event models. The ditions that are satisfied in the current state.
model associated with preparing exams and the model for Retrospective branching also involves determination of
fetching a stapler can be combined to form a composite men- the preconditions for achieving a given state. It can be im-
tal model within working memory. plemented with means-ends analysis, but using the present
These processes do not directly map to individual mech- state as the initial cue for retrieval instead of a future state.
anisms in the architecture, but are supported by existing Traditionally, these forms of reasoning leverage action-
mechanisms and representations. In combination, these pro- model knowledge stored in procedural memory. Action
cesses enable different types of AT reasoning. models feature preconditions and causally-related effects.
However, AT includes reasoning for distant or indirectly-
Types of Reasoning for Anticipatory Thinking experienced states while action models describe local expe-
In EST, event models are updated following misprediction. rience. Using event schemas and event models generalizes
However, events often proceed as expected with little addi- the aforementioned forms of reasoning to perform AT2 .
tional reasoning required to guide action. During these pe- With each of these methods, the same underlying archi-
riods, proactive reasoning can be performed to prepare for tecture is used and reactivity to the current situation is main-
future events without jeopardizing reactivity in the present. tained by incremental processing. As in human AT, if an ac-
This is one case in which it is possible to perform AT. tion must be taken in the moment, this reasoning may be
Alternatively, an agent may have a goal, but does not have interrupted or forgotten. Additionally, as in human AT, this
sufficient event schema knowledge for how to realize its reasoning can fail if there is simply insufficient knowledge
goal. (The knowledge may not exist as an event schema in available in memory. The main distinction between existing
semantic memory or it may be difficult to cue for retrieval.) reasoning methods and the provision of AT functionality is
In this case, an agent needs additional knowledge to proceed. the use of event models as the knowledge for simulating the
In either of these cases, an agent can perform additional environment. Thus, the main challenge in implementing AT
types of reasoning beyond the default EST behavior of re- in this model is learning and encoding event schema knowl-
trieving a single event schema to update the current working edge that includes causal relations and preconditions.
event model. Soar provides mechanisms to account for these
types of additional reasoning. Future Work and Implementation
In Soar, agents can detect when their knowledge for the
current situation is insufficient to select additional actions. So far, we have only considered a theoretical specification.
These situations are architecturally-recognized as impasses. Soar, and potentially other cognitive architectures, imple-
Note that this is distinct from misprediction. An agent could ment the forms of reasoning described above. However, cog-
have a good model of the environment, but not have the nitive architectures do not generally contain the necessary
knowledge for how to act or how the currently available mechanisms to implement event cognition. A full implemen-
actions will impact goal achievement. To resolve these im- tation of event cognition includes, but is not limited to: au-
passes, additional knowledge is brought into working mem- tomatic learning of event schemas, event model mispredic-
ory to guide action. These moments during which it is un- tion or surprise detection, memory for the past in terms of
clear which actions to perform (either in the present or in event models, and mechanisms for retrieving event models
preparation for the future) provide opportunities for AT. and event schemas based on their contents (Franklin et al.
Geden et al. describe three types of anticipatory thinking 2019). A full account of event cognition also describes how
that depend on the aforementioned AT processes: prospec- event models and schemas are used for reasoning and not
tive branching, backcasting, and retrospective branching. just the constraints placed on memory systems.
Using our model of event cognition, these types of antici- Other cognitive architectures besides Soar have included
patory thinking emerge from general cognitive processes in mechanisms that partially support event cognition. These ar-
Soar (and potentially in other cognitive architectures) that chitectures include Sigma, ACT-R/e, and Icarus.
support search-based planning and means-ends analysis. Sigma has mechanisms for detecting surprise (Rosen-
Prospective branching refers to imagining potential future bloom, Gratch, and Ustun 2015) and misprediction (Rosen-
states, given the current state. Search-based planning is an blooma, Demskia, and Ustuna 2017). Each can be used as
analogous form of reasoning in which an agent imagines a measure for detecting when to use a new event model to
potential futures by simulating actions using action mod- characterize the current situation. Sigma also includes some
els. When an agent has a goal, but the agent does not have episodic memory functionality (Rosenbloom 2014).
knowledge for which actions will accomplish this goal, an
agent performs search-based planning to simulate how avail- 2
This is similar to the approach taken by Cardona-Rivera et al.
able actions would change the situation. that used a planning-based knowledge representation for narratives.
ACT-R/e has been used to model some aspects of Geden, M.; Smith, A.; Campbell, J.; Spain, R.; Amos-Binks,
event cognition explicitly (Khemlani, Harrison, and Trafton A.; Mott, B.; Feng, J.; and Lester, J. 2019. Construction and val-
2015). The ACT-R/e implementation of event boundary en- idation of an anticipatory thinking assessment. PsyArXiv 9reby.
coding supports aspects of segmentation-based retrieval. Khemlani, S. S.; Harrison, A. M.; and Trafton, J. G. 2015.
Icarus supports event cognition with a dedicated episodic Episodes, events, and models. Frontiers in human neuroscience
memory store (Ménager and Choi 2016) and a measure of 9:590.
expectation violation explicitly presented as providing event Klein, G.; Snowden, D.; and Pin, C. L. 2011. Anticipatory
segmentation (Ménager et al. 2018). Icarus has been eval- thinking. Informed by knowledge: Expert performance in com-
uated for its ability to model human memory for events plex situations 235–245.
(Ménager, Choi, and Robins 2019). Laird, J. E.; Lebiere, C.; and Rosenbloom, P. S. 2017. A
These architectures motivate extending the specification standard model of the mind: Toward a common computational
of the Common Model to provide a formal account for framework across artificial intelligence, cognitive science, neu-
event cognition and anticipatory reasoning. Limitations to roscience, and robotics. AI Magazine 38(4).
the Common Model include insufficient specification of how Laird, J. E. 2012. The Soar cognitive architecture. MIT press.
query mechanisms retrieve event models and event schemas,
Ménager, D., and Choi, D. 2016. A robust implementation of
no event schema learning, little evaluation of error detection
episodic memory for a cognitive architecture. In Proceedings
or misprediction mechanisms for event models, no delin- of the 38th Annual Meeting of the Cognitive Science Society.
eation between episodic and semantic memory, and no di-
rect specification for what generally constitutes event cog- Ménager, D.; Choi, D.; Roberts, M.; and Aha, D. W. 2018.
nition functionality. A specification of event cognition func- Learning planning operators from episodic traces. In 2018
AAAI Spring Symposium Series.
tionality in general (including reasoning, memory, and learn-
ing) could motivate further implementation and evaluation Ménager, D. H.; Choi, D.; and Robins, S. K. 2019. A hybrid
among architectures. theory of event memory. Advances in Cognitive Systems.
Additional support for event cognition in cognitive archi- Radvansky, G. A., and Zacks, J. M. 2011. Event perception.
tectures will allow for computational models of human an- Wiley Interdisciplinary Reviews: Cognitive Science 2(6):608–
ticipatory thinking performance. This modelling depends on 620.
integrating event representations with existing agent reason- Radvansky, G. A., and Zacks, J. M. 2014. Event cognition.
ing for achieving goals. Pursuing this specification and im- Oxford University Press.
plementation will provide further constraint into which ar- Radvansky, G. A. 2012. Across the event horizon. Current
chitectural mechanisms and agent knowledge are useful – Directions in Psychological Science 21(4):269–272.
both for modelling humans and for implementing AT func- Richmond, L. L., and Zacks, J. M. 2017. Constructing ex-
tionality in artificial systems. perience: Event models from perception to action. Trends in
cognitive sciences 21(12):962–980.
Acknowledgments Rosenbloom, P. S.; Gratch, J.; and Ustun, V. 2015. Towards
Special thanks to Jeffrey Zacks for comments on EST. The emotion in sigma: from appraisal to attention. In Interna-
work described here was supported by the Office of Naval tional Conference on Artificial General Intelligence, 142–151.
Research under Grant N00014-18-1-2010. The views and Springer.
conclusions contained in this document are those of the au- Rosenbloom, P. 2014. Deconstructing episodic memory and
thors and should not be interpreted as representing the offi- learning in sigma. In Proceedings of the 36th Annual Meeting
cial policies, either expressly or implied, of ONR or the U.S. of the Cognitive Science Society.
Government. Rosenblooma, P. S.; Demskia, A.; and Ustuna, V. 2017. Toward
a neural-symbolic sigma: Introducing neural network learning.
References In Proceedings of the 15th Annual Meeting of the International
Baldassano, C.; Chen, J.; Zadbood, A.; Pillow, J. W.; Has- Conference on Cognitive Modeling.
son, U.; and Norman, K. A. 2017. Discovering event struc- Sargent, J. Q.; Zacks, J. M.; Hambrick, D. Z.; Zacks, R. T.;
ture in continuous narrative perception and memory. Neuron Kurby, C. A.; Bailey, H. R.; Eisenberg, M. L.; and Beck, T. M.
95(3):709–721. 2013. Event segmentation ability uniquely predicts event mem-
Cardona-Rivera, R. E.; Price, T.; Winer, D.; and Young, R. M. ory. Cognition 129(2):241–255.
2016. Question answering in the context of stories generated Szpunar, K. K.; Spreng, R. N.; and Schacter, D. L. 2014. A
by computers. Advances in Cognitive Systems 4:227–245. taxonomy of prospection: Introducing an organizational frame-
Eisenberg, M. L.; Zacks, J. M.; and Flores, S. 2018. Dynamic work for future-oriented cognition. volume 111, 18414–18421.
prediction during perception of everyday events. Cognitive re- National Acad Sciences.
search: principles and implications 3(1):53. Zacks, J. M., and Swallow, K. M. 2007. Event segmentation.
Ezzyat, Y., and Davachi, L. 2011. What constitutes an episode Current Directions in Psychological Science 16(2):80–84.
in episodic memory? Psychological Science 22(2):243–252. Zacks, J. M.; Speer, N. K.; Swallow, K. M.; Braver, T. S.; and
Franklin, N.; Norman, K. A.; Ranganath, C.; Zacks, J. M.; and Reynolds, J. R. 2007. Event perception: a mind-brain perspec-
Gershman, S. J. 2019. Structured event memory: a neuro- tive. Psychological bulletin 133(2):273.
symbolic model of event cognition. BioRxiv 541607.