Analogical Concept Memory for Architectures Implementing the Common Model of Cognition

2025-04-30 0 0 9.49MB 14 页 10玖币
侵权投诉
Analogical Concept Memory for
Architectures Implementing the Common Model of Cognition
Shiwali Mohan, Matthew Klenk
Abstract
Architectures that implement the Common Model of Cognition - Soar, ACT-R, and Sigma - have a prominent place
in research on cognitive modeling as well as on designing complex intelligent agents. In this paper, we explore how
computational models of analogical processing can be brought into these architectures to enable concept acquisition
from examples obtained interactively. We propose a new analogical concept memory for Soar that augments its current
system of declarative long-term memories. We frame the problem of concept learning as embedded within the larger
context of interactive task learning (ITL) and embodied language processing (ELP). We demonstrate that the analogical
learning methods implemented in the proposed memory can quickly learn a diverse types of novel concepts that are
useful not only in recognition of a concept in the environment but also in action selection. Our approach has been
instantiated in an implemented cognitive system Aileen and evaluated on a simulated robotic domain.
Keywords: cognitive architectures, common model of cognition, intelligent agents, concept representation and
acquisition, interactive learning, analogical reasoning and generalization, interactive task learning
1. Introduction
The recent proposal for the common model of cognition
(CMC; Laird et al. 2017) identifies the central themes in
the past 30 years of research in three cognitive architec-
tures - Soar (Laird, 2012), ACT-R (Anderson, 2009), and
Sigma (Rosenbloom et al., 2016). These architectures have
been prominent not only in cognitive modeling but also in
designing complex intelligent agents. CMC architectures
aim to implement a set of domain-general computational
processes which operate over domain-specific knowledge to
produce effective task behavior. Early research in CMC
architectures studied procedural knowledge - the knowl-
edge of how to perform tasks, often expressed as if-else
rules. It explored the computational underpinnings of a
general purpose decision making process that can apply
hand-engineered procedural knowledge to perform a wide-
range of tasks. Later research studied various ways in
which procedural knowledge can be learned and optimized.
While CMC architectures have been applied widely, Hin-
richs and Forbus (2017) note that reasoning in them fo-
cuses exclusively on problem solving, decision making, and
behavior. Further, they argue that a distinctive and ar-
guably signature feature of human intelligence is being able
to build complex conceptual structures of the world. In
the CMC terminology, the knowledge of concepts is declar-
ative knowledge - the knowledge of what. An example of
declarative knowledge is the final goal state of the tower-of-
hanoi puzzle. In contrast, procedural knowledge in tower-
of-hanoi are the set of rules that guide action selection
in service of achieving the goal state. CMC architectures
agree that conceptual structures are useful for intelligent
behavior. To solve tower-of-hanoi, understanding the goal
state is critical. However, there is limited understanding of
how declarative knowledge about the world is acquired in
CMC architectures. In this paper, we study the questions
of declarative concept representation, acquisition, and us-
age in task performance in a prominent CMC architecture
- Soar. As it is similar to ACT-R and Sigma in the orga-
nization of computation and information, our findings can
be generalized to those architectures as well.
1.1. Declarative long-Term memories in Soar
In the past two decades, algorithmic research in Soar
has augmented the architecture with decalartive long-term
memories (dLTMs). Soar has two - semantic (Derbinsky
et al., 2010) and episodic (Derbinsky and Laird, 2009) -
that serve distinct cognitive functions following the hy-
potheses about organization of memory in humans (Tulv-
ing and Craik, 2005). Semantic memory enables enriching
what is currently observed in the world with what is known
generally about it. For example, if a dog is observed in the
environment, for certain types of tasks it may be useful to
elaborate that it is a type of a mammal. Episodic mem-
ory gives an agent a personal history which can later be
recalled to establish reference to shared experience with a
collaborator, to aid in decision-making by predicting the
outcome of possible courses of action, to aid in reasoning
by creating an internal model of the environment, and by
keeping track of progress on long-term goals. The history
is also useful in deliberate reflection about past events to
improve behavior through other types of learning such as
reinforcement learning or explanation-based learning. Us-
Preprint submitted to Cognitive Systems Research October 24, 2022
arXiv:2210.11731v1 [cs.AI] 21 Oct 2022
ing dLTMs in Soar agents has enable reasoning complexity
that wasn’t possible earlier (Xu and Laird, 2010; Mohan
and Laird, 2014; Kirk and Laird, 2014; Mininger and Laird,
2018).
However, a crucial question remains unanswered - how
is general world knowledge in semantic memory acquired?
We posit that this knowledge is acquired in two distinctive
ways. Kirk and Laird (2014, 2019) explore the view that
semantic knowledge is acquired through interactive in-
struction when natural language describes relevant declar-
ative knowledge. An example concept is the goal of tower-
of-hanoi a small block is on a medium block and a large
block is below the medium block. Here, the trainer pro-
vides the definition of the concept declaratively which is
later operationalized so that it can be applied to recognize
the existence of a tower and in applying actions while solv-
ing tower-of-hanoi. In this paper, we explore an alternative
view that that this knowledge is acquired through exam-
ples demonstrated as a part of instruction. We augment
Soar dLTMs with a new concept memory that aims at ac-
quiring general knowledge about the world by collecting
and analyzing similar experiences, functionally bridging
episodic and semantic memories.
1.2. Algorithms for analogical processing
To design the concept memory, we leverage the com-
putational processes that underlie analogical reasoning
and generalization in the Companions cognitive architec-
ture - the Structure Mapping Engine (SME; Forbus et al.
2017) and the Sequential Analogical Generalization En-
gine (SAGE; McLure et al. 2015). Analogical matching,
retrieval, and generalization is the foundation of the Com-
panions Cognitive architecture. In Why we are so smart?,
Gentner claims that what makes human cognition superior
to other animals is “First, relational concepts are critical
to higher-order cognition, but relational concepts are both
non-obvious in initial learning and elusive in memory re-
trieval. Second, analogy is the mechanism by which rela-
tional knowledge is revealed. Third, language serves both
to invite learning relational concepts and to provide cogni-
tive stability once they are learned” (Gentner, 2003). Gen-
tner’s observations provide a compelling case for exploring
analogical processing as a basis for concept learning. Our
approach builds on the analogical concept learning work
done in Companions (Hinrichs and Forbus, 2017). Previ-
ous analogical learning work includes spatial prepositions
Lockwood (2009), spatial concepts (McLure et al., 2015),
physical reasoning problems (Klenk et al., 2011), and ac-
tivity recognition (Chen et al., 2019). This diversity of
reasoning tasks motivates our use of analogical processing
to develop an architectural concept memory. Adding to
this line of research, our work shows that you can learn
a variety of conceptual knowledge within a single system.
Furthermore, that such a system can be applied to not
only learn how to recognize the concepts but also acting on
them in the environment within an interactive task learn-
ing session.
1.3. Concept formation and its interaction with complex
cognitive phenomenon
Our design exploration of an architectural concept mem-
ory is motivated by the interactive task learning problem
(ITL; Gluck and Laird 2019) in embodied agents. ITL
agents rely on natural interaction modalities such as em-
bodied dialog to learn new tasks. Conceptual knowledge,
language, and task performance are inextricably tied -
language is a medium through which conceptual knowl-
edge about the world is communicated and learned. Task
performance is aided by the conceptual knowledge about
the world. Consequently, embodied language processing
(ELP) for ITL provides a set of functional requirements
that an architectural concept memory must address. Em-
bedding concept learning within the ITL and ELP contexts
is a significant step forward from previous explorations in
concept formation. Prior approaches have studied concept
formation independently of how they will be used in a com-
plex cognitive system, often focusing on the problems of
recognizing the existence of a concept in input data and
organizing concepts into a similarity-based hierarchy. We
study concept formation within the context of higher-order
cognitive phenomenon. We posit that concepts are learned
through interactions with an interactive trainer who struc-
tures a learner’s experience. The input from the trainer
help group concrete experiences together and a generaliza-
tion process distills common elements to form a concept
definition.
1.4. Theoretical Commitments, Claims, and Contribu-
tions
Our work is implemented in Soar and consequently,
brings to bear the theoretical postulates the architecture
implements. More specifically, we build upon the following
theoretical commitments:
1. Diverse representation of knowledge: In the past
decade, the CMC architectures have adopted the view
that architectures for general intelligence implement
diverse methods for knowledge representation and
reasoning. This view has been very productive in not
only studying an increasing variety of problems but
also in integrating advances in AI algorithmic research
in the CMC framework. We contribute to this view
by exploring how algorithms for analogical processing
can be integrated into a CMC architecture.
2. Deliberate access of conceptual knowledge: Follow-
ing CMC architectures, we assume that declarative,
conceptual knowledge is accessed through delibera-
tion over when and how to use that knowledge. The
architectures incorporates well-defined interfaces i.e.
buffers in working memory that contain information
as well as an operation the declarative memory must
execute on the information. Upon reasoning, informa-
tion may be stored, accessed, or projected (described
in further detail in Section 4).
2
3. Impasse-driven processing and learning: Our ap-
proach leverages impasse in Soar, a meta-cognitive
signal that can variably indicate uncertainty or fail-
ure in reasoning. Our approach uses impasses (and
the corresponding state stack) to identify and pursue
opportunities to learn.
4. A benevolent interactive trainer: We assume existence
of an intelligent trainer that adopts a collaborative
goal with the learning system that it learns correct
definitions of concepts. Upon being prompted, the
trainer provides correct information to the learner to
base its concept learning upon.
Based on these theoretical commitments, our paper con-
tributes an integrative account of a complex cognitive phe-
nomenon - interactive concept learning. Specifically, this
paper:
1. defines the concept formation problem within larger
cognitive phenomenon of ELP and ITL;
2. identifies a desiderata for an architectural concept
memory;
3. implements a concept memory for Soar agents using
the models of analogical processing;
4. introduces a novel process - curriculum of guided par-
ticipation - for interactive concept learning;
5. introduces a novel framework for evaluating interac-
tive concept formation.
Our implementation is a functional (and not an architec-
tural) integration of analogical processing in Soar’s declar-
ative long-term memory systems. It characterizes how
an analogical concept memory can be interfaced with the
current mechanisms. Through experiments and system
demonstration, we show that an analogical concept mem-
ory leads to competent behavior in ITL. It supports learn-
ing of diverse types of concepts useful in ITL. Learned
concept representations support recognition during ELP
as well as action based on those concepts during task per-
formance. The concepts are from few examples provided
interactively.
2. Preliminaries - The AILEEN Cognitive System
Aileen is a cognitive system that learns new concepts
through interactive experiences (linguistic and situational)
with a trainer in a simulated world. A system diagram is
shown in Figure 1. Aileen lives in a simulated robotic
world built in Webots1. The world contains a table-top
on which various simple objects can be placed. A sim-
ulated camera above the table captures top-down visual
1https://www.cyberbotics.com/
information. Aileen is engaged in a continuous perceive-
decide-act loop with the world. A trainer can set up a
scene in the simulated world by placing simple objects on
the scene and providing instructions to the agent. Aileen
is designed in Soar which has been integrated with a deep
learning-based vision module and an analogical concept
memory. It is related to Rosie, a cognitive system that
has demonstrated interactive, flexible learning on a vari-
ety of tasks (Mohan et al., 2012, 2014; Mohan and Laird,
2014; Kirk and Laird, 2014; Mininger and Laird, 2018),
and implements a similar organization of knowledge.
Visual Module. The visual module processes the image
taken from the simulated camera. It produces output in
two channels: object detections as bounding boxes whose
centroids are localized on the table-top and two perceptual
symbols or percepts corresponding to the object’s shape
and color each. The module is built using a deep learn-
ing framework - You Only Look Once (YoLo: Redmon
et al. (2016)). YoLo is pre-trained with supervision from
the ground truth in the simulator (12,000 images). It is
detects four shapes (error rate <0.1%) - box (percept
-CVBox), cone (CVCone), ball (CVSphere), and cylinder
(CVCylinder).
For colors, each detected region containing an object
is cropped from the image, and a K-means clustering is
applied all color pixel values within the crop. Next, two
weighted heuristics are applied that selects the cluster that
likely comprises the detected shape among any background
pixels and/or neighboring objects. The first heuristic se-
lects the cluster with the maximum number of pixels. The
second heuristic selects the cluster with the centroid that
is closest to the image center of the cropped region. The
relative weighted importance of each of these heuristics is
then trained using a simple grid search over w1and w2:
Score =w1Rs+w2(1Cs), s D, where w1+w2= 1, Dis
the set clusters, Rsdenotes the ratio between the number
of pixels in each cluster and the the number of pixels in
the image crop, and Csis the Euclidean distance between
the centroid of the cluster and the image center normal-
ized by the cropped image width. The average RGB value
for all pixels included in the cluster with the highest score
is calculated and compared with the preset list of color
values. The color label associated with the color value
that has the smallest Euclidean distance to the average
RGB value is selected. The module can recognize 5 colors
(error rate <0.1%): CVGreen,CVBlue,CVRed,CVYellow,
and CVPurple. Note that the percepts are named so to be
readable for system designers - the agent does not rely on
the percept symbol strings for any reasoning.
Spatial Processing Module. The spatial processing module
uses QSRLib (Gatsoulis et al., 2016) to process the bound-
ing boxes and centroids generated by the visual module to
generate a qualitative description of the spatial configu-
ration of objects. For every pair of objects, the module
extracts qualitative descriptions using two spatial calculi
3
摘要:

AnalogicalConceptMemoryforArchitecturesImplementingtheCommonModelofCognitionShiwaliMohan,MatthewKlenkAbstractArchitecturesthatimplementtheCommonModelofCognition-Soar,ACT-R,andSigma-haveaprominentplaceinresearchoncognitivemodelingaswellasondesigningcomplexintelligentagents.Inthispaper,weexplorehowcom...

展开>> 收起<<
Analogical Concept Memory for Architectures Implementing the Common Model of Cognition.pdf

共14页,预览3页

还剩页未读, 继续阅读

声明:本站为文档C2C交易模式,即用户上传的文档直接被用户下载,本站只是中间服务平台,本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私,请立即通知玖贝云文库,我们立即给予删除!
分类:图书资源 价格:10玖币 属性:14 页 大小:9.49MB 格式:PDF 时间:2025-04-30

开通VIP享超值会员特权

  • 多端同步记录
  • 高速下载文档
  • 免费文档工具
  • 分享文档赚钱
  • 每日登录抽奖
  • 优质衍生服务
/ 14
客服
关注