Cognition’s Interface to the World
and to Metacognition in MIDCA
MICHAEL T. COX
D E PA R T M E N T O F C O M P U T E R S C I E N C E & E N G I N E E R I N G W R I G H T S TAT E U N I V E R S I T Y, D AY TO N , O H
Virtual International Symposium on Cognitive
Architecture 2021 ♠ 8 May 2021
Levels of Cognitive Computation
Ground Level
◦
Action
◦
Perception
Object Level
◦
Problem Solving
◦
Comprehension
Meta-Level
◦
Meta-level Control
◦
Introspective Monitoring
Cox, M. T., & Raja, A. (2011). Metareasoning: An introduction.The
Metacognitive
Integrated
Dual-Cycle
Architecture
(MIDCA)
Cox, M. T., Alavi, Z., Dannenhauer, D., Eyorokon, V., Munoz-Avila,
Object Level
Meta-Level
Metacognition
Cognition
Memory Mission & Goals( ) World Model (MΨ) Episodic Memory Semantic Memory Problem Solving Comprehensiongoal change goal input goal insertion Intend Plan Evaluate Interpret Goals subgoal Goal Actions MΨ MΨ MΨ State Hypotheses goal change Intend Controller Plan Evaluate Monitor Interpret Meta Goals goal insertion subgoal goal input Goal Hypotheses Algorithms Trace Meta-Level
Control Introspective Monitoring
Memory Reasoning Trace ( ) Strategies (△ ) Episodic Memory Metaknowledge Self Model ( ) Mental Domain = Ω Goal Management
Executive Metacognition
Memory
Mission & Goals( ) World Model (MΨ) Episodic Memory Semantic Memory ( ) & Ontology Plans( ) & States( ) Problem Solving Comprehensiongoal change goal input
Introduction
Cognition’s Interfaces
Cognition’s Interface to the Environment
Cognition’s Interface to Metacognition
Conclusion
Cognition’s Interfaces
Dannenhauer, Z. A., & Cox, M. T. (2018). Rationale-based perceptual monitors. AI Communications, 31(2), 197-212.
Cognition’s
Interfaces in
MIDCA
MIDCA Cognitive cycle
MIDCA Metacognitive cycle
Memory T rac e C ont rol
Standard Simulator
BuffersFeedback Audio Image
Cognition’s Interface to
the Environment
Domain Structure Definition
Object Detection
◦
Type hierarchy
◦
Object attributes and values
◦
Relationships between objects
Action Execution
◦
Operators mapped to low-level
communication channels
Configuration File parsed for new domain
◦
Generates object detection stubs
◦
Creates handlers for MIDCA
Act and Perceive phases
MIDCA / ROS API
on(block1,block2)
( { block1 color:“red”; z: “11”; y: “2”; x: “2”}, { block2 color:“green”; z: “6”; y: “2”; x: “2”} ) Command Nodesunstack(block1,block2)
loc_cmd(11,2,2) grab_cmd() raise_cmd() Object Detection NodesMIDCA’s Knowledge
of states and actions is
in terms of
◦
Logical predicates
◦
PDDL operators
Mapped to and from
◦
ROS nodes through
◦
ROS topics
Topics:
Perception
Sensor Handler
◦
Detects objects in environment
◦
Places their IDs and attributes in
perceptual buffer
Perceive Phase
◦
Infers predicate representations of objects
and their relationships
Action
Execution Handler
◦
Packages command arguments into strings
◦
Sends message to command nodes through
associated topics to control effectors
Act Phase
◦
Watches for feedback from ROS
Problem: Too much unnecessary
Plan Monitoring Demo
Neo
◦
The Baxter humanoid robot
Plan and Goal Monitors
◦
Manage perception
Cognition’s Interface to
Metacognition
The
Metacognitive
Cycle of
MIDCA
goal changeIntend
Controller
Plan
Evaluate
Monitor
Interpret
Meta Goals
goal insertion subgoal goal inputGoal
Hypotheses
Algorithms
Trace
Meta-LevelControl Introspective Monitoring
Cognitive Phases as mental actions
Memory as mental states
Memory
Snapshot 1
Intend
Plan
Introspective Monitoring
Memory
Snapshot 2
Snapshot 3
Memory
(& Speak)
Act
Metacognitive Expectations
A Mental State is a vector of variables ⃑" = "
1
, … , "'
(
)
= (+
,
, -
., /, 0Y, 1, χ, 3
,
)
◦
Current goal set
◦
Goal agenda
◦
Current plan
◦
World state
Metacognitive Expectation as Boolean function over segment of mental trace
56
)
(
7
)
, 3
7
)
, (
789
)
→ {⊤, ⊥}
◦
Discrepancies
◦
Explanation
Metacognitive Expectation for ‘Explanation’ Mental Action
("
#
$
, &'()*+*,-.+, "
#/0
$
)
Memory
Snapshot 1
Problems
Detect
Explanation
Metacognitive Expectation for Interpret
Memory
Snapshot 2
Snapshot 3
Memory
Formulation
Goal
Prior mental
state
Mental
action
Post mental
state
Meta-Level Control
Goal Change
◦
Goal transformation
◦
Goal priorities
Execution Control
◦
Over cognitive phases
Learning
◦
Action models
◦
Other?
Executive Metacognition
World =Ψ
Memory
Mission & Goals( ) World Model (MΨ) Episodic Memory Semantic Memory ( ) & Ontology Plans( ) & States( ) Problem Solving Comprehension goal change goal inputMetacognition Example
Plant Protection Domain with faulty spray op
◦
Goals: all invasive plants; all native plants alive
◦
Spray herbicide in one cell but kills native plant
in adjacent cell
Goal Expectation Failure detected
by Interpret at object level
◦
Goal to preserve native plant but now dead
◦
MIDCA cannot explain the failure
Metacognitive Expectation Failure detected
by meta-level Interpret at meta-level
Experimental Results in Two Domains
Average performance as a function of problem complexity
in the Plant Protection domain.
Open-source Code available at https://github.com/COLAB2/midca
Integrating Behavior, Cognition and Metacognition is hard for any architecture
◦
MIDCA has explored metacognition from the beginning but only lightly
◦
MIDCA has ignored grounded cognition for a long time
◦
But both have started to take off as of late
The Future is interesting
Backup Slides
Origins of MIDCA: Don Norman (1986)
Norman, D. 1986. Cognitive Engineering. In D. Norman and S. Draper eds., User-centered system design: New perspectives on
human-computer interaction. Hillsdale, NJ: Lawrence Erlbaum.
Norman’s HCI model
MIDCA’s Cognition model
Cox, M. T.; Oates, T.; and Perlis, D. 2011. Toward an integrated metacognitive architecture. In P. Langley ed., Advances in