Integrating Semantic Memory
in Soar
Yongjia Wang
John E. Laird
1
Outline
• Background & Motivation
• Implementations and Experiments
2
What is Semantic Memory
• Definition
– ‘Your memory for meanings and general (impersonal) facts.’
[from WordNet]
• Episodic memory and Semantic memory distinction
– Episodic memory
• Tied to a specific learning episode or experience
• What you remember
– Semantic memory
• General knowledge not tied to a learning experience
• What you know
3
Memory Systems
Memory
Long Term Memory
Declarative
Semantic
Memory
Episodic
Memory
Short Term Memory
Procedural
Perceptual
Representation
System
Procedural
Memory
Working
Memory
4
Related Fields and Motivation
Focus
Feature
Limitations
Architectures
Cognitive Psychology
(ACT-R)
To model human Long-term
Haven’t been used
behavior
declarative memory to build functional
and learning
agents
AI Agent Architectures
(Soar)
To build
Efficient domain
intelligent agents knowledge
engineering
No long-term
semantic memory,
limited learning
Knowledge
Representation Systems
(CYC)
To represent
common sense
semantic
knowledge
Representational
model, not learning
model
Declarative
knowledge
representation
To build
Efficient domain
Our Approach
(Soar + semantic memory) intelligent agents knowledge
engineering and
more learning
capabilities
Constrained by Soar
5
Research Goals
• To improve general functionality of Soar by semantic
memory
– Explore new cognitive capabilities
– Characterize computational functionalities
• To understand semantic memory in the context of a
general cognitive architecture
– How to use semantic memory in specific tasks?
– How semantic memory interacts with other mechanisms in Soar?
– What are the computational implications of semantic memory
and episodic memory distinction?
6
Distinction Between Semantic Memory and
Episodic Memory in Soar
Semantic Memory
Storage & retrieval Single level objects in
working memory
unit
Episodic Memory
Entire working memory
snapshot (episode)
(declarative chunks)
Temporal
information
No architectural temporal
information
Architectural temporal
information (ex: next
episode)
Main purposes
Store general knowledge
Category learning
Store specific events
Case-based reasoning
7
Outline
• Background & Motivation
• Implementations and Experiments
Task: Cognitive Arithmetic
8
Overview of Experiment
• Purpose:
– Integrate a declarative semantic memory component
– Demonstrate related functional advantage of declarative
representation
• Implementation:
– Semantic memory with declarative representation
– Deliberate and automatic semantic learning
• Task: Cognitive arithmetic
– Easy to understand
– Universally performed
– Multiple types of learning
0 9 5 2
+0 0 6 3
1 0 1 5
9
Working Memory Representation of an
Arithmetic Problem
0 9 5 2
state
S1
+ 0 0 6 3
---------------------
arithmetic-problem
1 0 1 5
A1
operation
addition
one-column
NIL
next-column
digit1
9
next-column
next-column
C3
C2
digit2
0
digit1
5
C1
digit2
6
digit1
2
digit2
3
10
Problem Space
process-column
next-column
If last column and carry=0
If last column and carry=1
new-column
stop
get-digit1
get-digit2
compute-result
write-result
If carry = 1
compute-result
write-digit1
11
Solving One Column
An unknown sub-problem
Chunking learns a rule
Compare
General procedural knowledge
Semantic learning
12
Soar
Storage:
Long-Term Memories
Declarative structure consist of triplet elements
Semantic
Procedural
Reinforcement Chunking
Learning Retrieval:
Episodic
Semantic
Learning
Encoding:
Deliberate retrieval
Episodic
Learning
Option1: Deliberately encode
structure specified by task knowledge
Cue is created under ‘cue-link’
Hash based exact match
Option2: Automatically encode all
structures
Short-Term Memory
Use:
Decision
Procedure
Task specific strategies
Action
Perception
Body
13
Underlying Storage
Structure
Declarative
Chunks
Identifier – attribute – value
hash
sum
A1
A1
digit1
digit2
A2
…H
Cue
…H
sum
digit1
digit2
…H
A1
digit2
sum
5
2
digit2
sum
digit1
5
…H
2
…H
3
…H
6
…H
3
…H
3
…H
digit1
3
?
2
Array of elements
A2
(A1, sum, 5)
(A1, digit1, 2)
(A1, digit2, 3)
(A2, sum, 6)
(A2, adigit1, 3) (A2, digit2, 3)
…A
digit2
sum
digit1
6
3
3
A1
…H
A2
…H
5
6
Attribute – value - identifier
hash
A1
…H
sum
…H
2
digit1
A2
3
…H
A1
…H
3
digit2
…H
A2
…H
…H
14
3
Transfer Learning Effect from Semantic
Learning
^digit1 5
^digit2 6
^operation addition
^sum 1
^sum 1
^digit1 5
^digit2 6
^carry-borrow 1
^operation subtraction
^carry-borrow 1
Retrieval
Semantic
learning
sub-state
S1
Counting: 5,6,7,8,9,(1)0, (1)1
carry-borrow sum
1
1
Counting once and learn 1
declarative chunk
digit2
digit1
5
6
15
Procedural Representation Cannot be
Transferred
^sum 1
^digit1 5
^digit2 6
^operation addition
^digit1 5
^sum 1
^digit2 6
^carry-borrow 1
^operation subtraction1
^carry-borrow 1
chunking
chunking
sub-state
sub-state
Counting: 5,6,7,8,9,(1)0, (1)1
Counting: 1,0,(-1)9,(-1)8,(-1)7,(-1)6, (-1)5
Counting twice and
learn 2 rules
16
Comparison of Different learning
Configurations
234 + 345 = 579
579 – 345 = 234
579 – 345 = 234
72
33
9
17
Decision Cycles Breakdown
Situations
All
computations
With
arithmetic facts
After chunking
operators in top-state
(initialization, process-column, next-column)
9
9
9
get-digits (from top-state)
3×3=9
9
0
write-result (to top-state)
3×1=3
3
0
retrieve
3×4=12
12
0
counting
39
0
0
Total
72
33
9
Decision Cycles
Operators
18
Summary
• Nuggets
– Implemented a semantic memory with declarative representation
– Demonstrated the functional advantage of declarative
representation over procedural representation
– Demonstrated transfer learning effect by semantic learning
– Demonstrated the functional Interaction between semantic
learning and chunking
• Coals
– Cognitive arithmetic is an internal mental task
– The task is completely deterministic
19
Thank You
20