This document summarizes Valentina Rho's thesis on extending a hybrid knowledge representation system called Dual-PECCS into the cognitive architecture ACT-R. Dual-PECCS represents concepts using both classical and typical information based on dual-process theory. The objectives were (a) extending Dual-PECCS and (b) integrating it into ACT-R. Dual-PECCS was translated into ACT-R chunks and a new action allows accessing its subsystems. Experiments using riddles showed the integrated system had similar accuracy to humans in conceptual categorization and representation proxyfication. Future work includes improving generalization in the typical system and integrating Dual-PECCS into other architectures.
Extending a Hybrid Knowledge System into the ACT-R Cognitive Architecture
1. Extending and integrating a
hybrid knowledge representation
system into the cognitive
architecture ACT-R
Valentina Rho
Università degli Studi di Torino
supervisor: Daniele P. Radicioni
co-supervisor: Antonio Lieto
15th International Conference of the Italian Association for Artificial Intelligence, 1 December 2016
2. Thesis objectives
(a) To extend a hybrid knowledge representation
system based both on classical and typical
information (S1S2)
(b)To integrate this system within the well-known
cognitive architecture ACT-R
3. What is a concept and how
to represent it?
• A concept is an abstract mix of information about a
set of items that share common characteristics.
• Different theories try to find a way to represent
concepts: for example classical, prototypes or
exemplar-based theories.
6. Dual process theory
• In our mind there are two types of reasoning
processes:
• S1: fast, instinctive and emotional
• S2: slower, more deliberative, and more logical
Kahneman (2011)
8. The proxytypes theory
Prinz (2002)
…
birds
…
black
penguins
penguins
proxyfication
short-term
memory
long-term memory
9. Heterogeneous proxytypes
Lieto (2014)
Short-term memory
Long-term memory
exemplars
MAMMAL
concept
prototypes
…
whale
exemplar proxyfication
similarity-
based check
classical
representation
Stimulus
concept v
concept y
concept x
concept z
10. Dual-PECCS algorithm
“The big fish that eats plankton”
Typical
System - S1
(Conceptual
Spaces)
Classical
System - S2
(OpenCyc)
Information
Extractor
Internal
representation
(dimension: big,
family: fish,
feeding: plankton)
whale 1.0
whale-shark 0.8
shark 0.7
…
whale is not a fish
whale-shark is ok
our first unconscious answer would be whale
our conscious and reasoned answer
would be whale shark
11. Cognitive architectures
• The objective of a cognitive architecture is to define
a comprehensive theory about the structure and
the underlying mechanisms of the human mind.
• Some examples: ACT-R, Clarion, SOAR
12. ACT-R Architecture
Anderson et al. (2004)
External Environment
Vision module Aural module
Motor module
Visual buffer
Visual-
location buffer
Manual buffer
Goal buffer
Retrieval
buffer Imaginal
buffer
Goal module
Declarative module
Imaginal module
Procedural module
(match; select; fire) Aural buffer
Aural-location
buffer
Speech module
Vocal buffer
Working memory
(buffers)
Production list
[…]
Chunks list
[…]
Long-term memory
13. Integration in ACT-R
What we’ve done?
• Translated the Dual-PECCS typical KB into chunks,
considering bodies-of-knowledge chunks and
conceptual chunks
• Extended the ACT-R DM with a dedicated action to
allow access to the Dual-PECCS subsystems
• Implemented the main reasoning algorithm within
the ACT-R production rules system.
14. Additional points of
extension
• We extended the attentional markers of ACT-R to
emulate the “change of mind” process when the S2
system doesn’t confirm the fast typical answer.
• We preliminarily studied how to extend the
activation formulas of ACT-R (based on recency
and frequency of retrieval) in order to follow the
intuition that the activation of a concept should be
function of the activation values of its
representations.
15. Experiments
• We used 90 textual riddles in two types of
experiments:
1. with manual information extraction
2. with automatic information extraction
i.e.“The big fish that eats plankton”
The results produced by the system have been
compared to the responses provided in a
psycological experiment by 10 human volunteers.
16. Results
• CC-Acc is the conceptual categorization accuracy:
when the system returns the correct concept.
• P-Acc is the proxyfication accuracy: when Dual-
PECCS not only returns the correct concept but
also proxyfies the correct representation of it.
17. P-Acc analyses
• The system fails mostly when we are expecting a
Prototype but and Exemplar is proxyfied. This
means we need to improve the generalization
process within the S1 system.
18. What’s next / now?
• Integration of Dual-PECCS in other cognitive
architectures (SOAR, Clarion)
• Automatic population of the typical knowledge
base
• Improving generalization within typical system (S1)