RPD-Based Cognitive Modeling : A Framework for the Comprehensive Computational Representation of Human Decision Processes
Chris Forsythe
Sandia National Laboratories
Albuquerque, NM
jcforsy@sandia.gov RPD-Based Cognitive Modeling
Acknowledgements : Acknowledgements Sandia National Laboratories Team
Patrick Xavier, Eric Parker, Ed Thomas, Dave Schoenwald, Sabina Jordan, Caren Wenner, Nathan Brannon, Elaine Raybourn, Howard Hirano, …..
University of New Mexico
Tim Goldsmith
Northern Illinois University
Joe Magliano, Mary Anne Britt, Christoph Guess
Program Objectives : Program Objectives Objective: collection of projects focused on development of a framework for the comprehensive modeling of factors that shape human decision processes in naturalistic settings
Technical Challenge: realistically model decision making processes in a manner that is comprehensive, defensible, extensible and doable. Problem Space: (1) Synthetic humans for training and analytic tools; (2) Intelligent machines; and (3) Core technology for augmented cognition solutions
Critical Requirements for Realism of Synthetic Entities : Critical Requirements for Realism of Synthetic Entities Representation of emotional processes including the interaction between emotions, arousal and cognitive processes
Representations of knowledge that provide a broad range of relevant, and also, irrelevant behavioral responses
Mechanisms to address variations in knowledge and emotional associations attributable to cultural differences
Mechanisms that enable non-linear patterns of behavior and reasonably realistic reactions to non-linear behavioral responses
Theoretical Synthesis : Theoretical Synthesis The simulation incorporates a theoretical synthesis of Klein’s RPD model of naturalistic decision making (Klein, 1993), oscillating systems theory of semantic and episodic memory processes (Klimesch, 1996) and LeDoux’s (1998) model of the relationship between cognitive and emotional processes (Forsythe, 2001).
Modeling Recognition Primed Decision Making : Modeling Recognition Primed Decision Making Semantic Knowledge Situational Knowledge
- Situation A
- Situation B
- Situation C
-
- Pattern Recognition Interpretation of Situation Cues and Knowledge of Ongoing Events Cues in the environment activate concepts in semantic net Patterns of activation are recognized that correspond to known situations Knowledge of appropriate actions (e.g., scripts) are implicit to situation recognition Situations and corresponding knowledge
Slide7 : Cognitive Model: Initial Pattern Recognition Approach Environmental Cues and
Knowledge of Ongoing Events Concept A Concept B Concept C Concept D Concept E Situation 1 Situation 2 Situation 3 Begin Snapshot End Snapshot Semantic Network
Perceptual events and knowledge lead to activation of concepts in a semantic network with activation spreading to related concepts Pattern Recognition
The pattern of activation in the semantic network is matched with patterns associated with known situations Situation Recognition
Sequential snapshots increment or decrement evidence for alternative situations in a race model
Physiology-Based Engine: Foundation in Oscillating Systems Theory : Physiology-Based Engine: Foundation in Oscillating Systems Theory Knowledge Network Situation Recognition Oscillators represent each node in the network. In operation, the activity of these oscillators may vary in frequency and amplitude. A single oscillator underlies the pattern recognition process. Operating at a lower frequency (10-13 Hz for semantic processes versus 4-7 Hz for recognition) it provides a relatively broad window for recognition of patterns in the activation of the semantic network.
Properties of Semantic Processes: Reverse Engineering and First Steps Toward V&V : Properties of Semantic Processes: Reverse Engineering and First Steps Toward V&V Example of Specification:
The rate of information processing corresponds to the dominant frequency in the 10-13 Hz Bandwidth
(Klimesch 1996) Frequency of Oscillations for Neural Units (Hz) Decision Time (sec) Simulated Event-Related Potential Actual Event-Related Potential
Emotions: Fear : Emotions: Fear Emotion leads to increased activation of neural assemblies corresponding to the stimulus event or situation associated with emotional reaction, and active inhibition otherwise (LeDoux 1998) Activation Observed in the Absence of Fear-Inducing Stimulus No association Association No association Association Activation Observed with Fear-Inducing Stimulus Presented at Approximately 2 Seconds.
Issues Now Occupying Our Minds : Issues Now Occupying Our Minds Population of Models with Knowledge
Knowledge Elicitation
Knowledge Representation
Pattern Recognition
Episodic Memory
Perceptual Processes
Scalability, Interoperability
Population Models
Details, Characterization and Validation
Population of Models with Knowledge: Elicitation for Semantic Knowledge : Population of Models with Knowledge: Elicitation for Semantic Knowledge Approach developed and validated by Schvaneveldt, Goldsmith and colleagues based on Pathfinder Networks (Schvaneveldt, 1990)
Identify critical concepts
Collect relatedness data for concept pairs
Develop Pathfinder Network identifying associative links between concepts and associative strength of these links
Implement in model where concepts represented as nodes with associative links and strengths of associative links provide basis for spreading activation
Representation in Model: Semantic Module : Representation in Model: Semantic Module Activation routed to individual nodes Each concept represented as separate node, with oscillatory properties Timer operates as pacemaker sending periodic pulses to network
Larger Semantic Spaces: Latent Semantic Content Analysis : Larger Semantic Spaces: Latent Semantic Content Analysis Goal to develop customized knowledge structures based on domain, culture, knowledge of individual, etc. Books, Magazines, Newspapers, Movie & Television Scripts, etc. Text Parser Concept
Concept
Concept
Concept Sentences/Paragraphs 0 1 1 1 0
1 1 0 1 0
0 0 0 0 1
0 0 1 0 0 1 2 3 4 5 . . . A B C D Singular Value Decomposition Points in Multi-Dimensional Space Semantic Network Representation
Population of Models with Knowledge: Elicitation for Situational Knowledge : Population of Models with Knowledge: Elicitation for Situational Knowledge Utilized approach wherein task analysis emphasizes actions and associated cues, not unlike CDM (Klein, et.al., 1989)
Identify actions and develop associated flowcharts
Identify cues associated with actions providing basis for discrimination of situations
??? Relatedness cues-to-situations, so far done analytically
Implement in model where pattern recognition based on cues and endpoints of flowchart serve as separate actions
??? Pattern recognition, initially template match, subsequently used information accumulation and neural net approaches
Representation in Model: Situation Recognition Module : Representation in Model: Situation Recognition Module Timer operates as pacemaker sending periodic pulses to network Assign evidence based on semantic activation –
??? Knowledge Elicitation Situation activation if continuation of current situation –
??? Assume bias for current situation Assess evidence if not continuation of current situation Assess cues contrary to expectations –
??? Knowledge Elicitation Introduce top-down activation for semantic concepts –
??? Knowledge Elicitation Placeholder for Level 2 & 3 RPD
Episodic Memory : Episodic Memory The objective is a computational model capable of the following:
Store representations of experience
Retrieve representations of experience
Apply experience in meaningful ways to recognize the solution to current problems
Apply experience as a basis for recalling knowledge relevant to current circumstances
Communicate on the basis of shared experience
Equally applicable to synthetic human and intelligent machine applications
Life Experience Generator: Machine Cognitive Model : Life Experience Generator: Machine Cognitive Model Source of Smoke Alarm Coming from Pasages Detect Smoke Detect Smoke Smoke Thickens Smoke Thickens Smoke Lessens Smoke Lessens Smoke Lessens Robot 1 Robot 2 Employ Umbra simulation environment to generate a range of representative experiences for robotic entities
Semantic Activation Across Time : Semantic Activation Across Time For a duration of time, there is a sequential progression in the patterns of concept activation Smoke Passage Hallway Smoke Passage Hallway Smoke Passage Hallway Progression of Time t1 t2 t3 t4
Recognition of “Schema” based on Patterns of Semantic Activation : Recognition of “Schema” based on Patterns of Semantic Activation
Episodic Representation Based on Recurrent Sequences of Schema (i.e., Thremes) : Episodic Representation Based on Recurrent Sequences of Schema (i.e., Thremes)
Statistical Derivation Schema & Themes : Statistical Derivation Schema & Themes Use high-dimensional data vector of robot status (each semantic node represented as vector)
Construct representative training set of data
Use unsupervised learning of training set to develop clusters (each cluster representing general robot state)
Use supervised learning (classification trees) to interpret clusters
Analyze cluster space: by robot, by run, and temporally
Transitions (in time) from cluster to cluster define schema
Consistently occurring sequences of schema define themes
Neural Nets for Recognition of Schema and Themes : Neural Nets for Recognition of Schema and Themes Semantic Knowledge Situational Knowledge
- Situation A
- Situation B
- Situation C
-
- Cues and Knowledge of Ongoing Events Schema Recognition Theme Recognition Interpretation Situation Episodic Memory Record
Training Neural Nets for Recognition of Schema and Themes : Training Neural Nets for Recognition of Schema and Themes Inputs Output Episode Recognition Input layer
Hidden layer(s)
Output layer Neural Net
Model i1 i2 in . . . w1 w2 wn Response Function
Perceptual Processes : Perceptual Processes Exploring concept wherein a federation of perceptual agents provide input to perceptual synthesis leading to semantic activation - ??? Concept Exploration Sensor
Sensor
Data
Perceptual Agent Perceptual Agent Perceptual Agent
Perceptual Synthesis Federation Perceptual Agents Perceptual Processes
Scalability, Interoperability : C++
Semantic Association/Activation Network expected to be large in real applications (60-80 vs 1,000-10,000 nodes).
Real-time (or faster) interactivity desired.
Umbra framework for highly modular simulations and systems.
Extends data-flow model to better support interacting agents and interaction phenomena.
Integration with Tcl/Tk for user interactivity.
Easy to interface to devices, C/C++ libraries.
MPI support for parallelism within Umbra, e.g system of interacting Human Emulators.
HLA support for simulation in the large and federation. Scalability, Interoperability
Population Models : Population Models Exploring concepts in which populations of distinctly different cognitive models interact –
??? Generative Cognitive Models, Parallelization Population
Details, Characterization and Validation : Details, Characterization and Validation Numerous assumptions underlie model, numerous parameters require characterization, and a methodology is required for validation
- ??? differential contribution of cues
- ??? additive and inhibitory influences of cues
- ??? redundancy of cues
- ??? differential saliency of cues
- ??? facilitory priming of cues
- ??? inhibitory priming of cues
- ??? differential sequencing of cues
- ??? top-down facilitation by preceding situation(s)
- ??? top-down inhibition by preceding situation(s)
- ??? situation release and situation nesting
Development of Experimental Protocol for Validation : Development of Experimental Protocol for Validation Study 1: College students will generate the specific instances of general situations with ratings of familiarity and frequency
Study 2: College students will generate cues present when situation is occurring
Events
Knowledge states (expectations, beliefs, and goals)
Emotions
Actions
Conclusion : Conclusion Developed test case models to prove in-principle approach
Characterization and validation effort commencing with NIU collaborators
Emerging capability for simulations based on synthetic humans, although intelligent machine may be more immediate application (e.g., intelligent building based on embedded experts)
Exploring ideas associated with real-time knowledge elicitation
Slide31 : Additional Slides
Outline : Outline Simulation Setup
Summary of Analysis Strategy
Data Vector
Cluster Analysis
Distribution of Robot States
Simulation Setup : Simulation Setup 20 simulations using RMSEL
8 robots per simulation
Fire/smoke location varies across runs
Initial robot status (e.g., location) is consistent across simulations
Fire/smoke location influences robot behavior
Robot Search Paths - Simulation#5 : Robot Search Paths - Simulation#5
Robot Search Paths - Simulation#18 : Robot Search Paths - Simulation#18
Analysis Strategy - Summary : Analysis Strategy - Summary Construct high-dimensional data vector(continuous and binary dimensions) that describes robot status
Construct representative set of high dimensional data
reduced temporal resolution (1/20 sampling rate)
800 randomly chosen observations (10 per robot per 10 simulations)
Develop clusters using unsupervised learning
cluster represents general robot state
Interpret clusters using classification trees (supervised learning)
Analysis Summary (continued) : Analysis Summary (continued) Analyze state space
Over all runs/robots
By robot
By run
Temporally: By robot within run
High-Dimensional Data Vector : High-Dimensional Data Vector 1. Time (t)
2. Xt
3. Yt
4. abs(DXt)
5. abs(Dyt)
6. Smoket
7. DSmoket = Smoket-Smoket-1
8. Smoket - max(robot smoke)
9. Smoket - max(system smoke)
10. IS_Beacon
11. IS_Last
12. IS_Rover
13. RF_Hear_Beacon
14. RF_PING
15. STOP
Cluster Analysis : Cluster Analysis Use representative data set
Unsupervised learning
Candidate cluster analysis methods
Partitioning methods: Kmeans, medoids
Agglomerative methods
Divisive methods
Example uses Kmeans method (5 clusters)
Ref: Finding Groups in Data: An Introduction to Cluster Analysis, L. Kaufman, P. J. Rousseeuw, Wiley, 1990.
Gross Interpretation of Clusters (Robot States) Via Classification Tree (KMEANS) : Gross Interpretation of Clusters (Robot States) Via Classification Tree (KMEANS) 5 3 1 2 4 Moving Stop Small DX Large DX Away from Robot’s
Max Smoke Near Robot’s Max Smoke No RF_Ping RF_Ping
Robot State Distribution Over all 20 runs and all 8 robots : Robot State Distribution Over all 20 runs and all 8 robots
Robot State Distribution By Robot ID : Robot State Distribution By Robot ID
Robot State Distribution By Run (1:10) : Robot State Distribution By Run (1:10)
Robot State Distribution By Run (11:20) : Robot State Distribution By Run (11:20)
Gross Interpretation of Clusters (Robot States) Via Classification Tree (Divisive “DIANA”) : Gross Interpretation of Clusters (Robot States) Via Classification Tree (Divisive “DIANA”) RF_Hear_Beacon OFF RF_Hear_Beacon ON Lo Smoke Hi Smoke Small DX Large DX Small DY Large DY Small DX Large DX 1 2 6 5 3 4
Robot State Distribution By Robot ID : Robot State Distribution By Robot ID
Robot State Distribution By Run (1:10) : Robot State Distribution By Run (1:10)
Robot State Distribution By Run (11:20) : Robot State Distribution By Run (11:20)
Timeline for Sandia National Laboratories Cognitive Modeling and Simulation : Timeline for Sandia National Laboratories Cognitive Modeling and Simulation Next Generation Security Simulation
Concept for cognitive model of human naturalistic decision making FY99 FY00 FY01 Organic Model
Systems model with human as source of organic properties Comprehensive Representation
Merged cognitive and organic models to create prototype framework for comprehensive model Extensible Knowledge-
Based Agents
Modeling domain knowledge and practical application of cognitive models Human-Like Episodic Memory
Model knowledge derived through collective life experiences Intermediate Layer
Models for perceptual and action generation processes Cognitive-Driven Augmented Analyst
Integrates cognitive architecture with data exploitation for next-level capability History of Human Reliability and Vulnerability Analysis for High Consequence Systems
General Cognitive Architecture for Agents : General Cognitive Architecture for Agents Mismatch Sensor
Sensor
Data
Perceptual Agent Perceptual Agent Perceptual Agent Situation/Contextual Knowledge
Situation
Situation Comparator Selective Attention Action Generation Drive Mechanism Emotional Processes Human Interface:
Visual, Auditory, Haptic, Kinesthetic, Etc.
Perceptual Synthesis Federation Perceptual Agents Perceptual Processes