Transcript slides - AGI conferences

Cognitive Model Comparisons:
The Road to Artificial General Intelligence?
Christian Lebiere ([email protected])
Cleotilde Gonzalez ([email protected])
Carnegie Mellon University
Walter Warwick ([email protected])
Alion Science & Technology
Challenges in AI & Cognitive Science
• Both fields have similar history of challenge problems
despite compatible ends but different means
• Artificial Intelligence: maximize task performance
– Started with ambitious but poorly defined test (Turing Test)
– Evolved narrow, precise, overspecialized challenges (Chess)
– Recently attempted broader tests (Robocup, Grand Challenge)
• Cognitive Science: fit human capabilities (design guide)
– Started with ambitious, ill-defined capacities list (Newell Test)
– Organized a series of complex task comparisons (AMBR, HEM)
– Is taking on broader but integrated challenges (DSF?)
3/7/09
Artificial General Intelligence Conference
2
Cognitive Challenge Pitfalls
• Challenge is fundamentally about the task, not cognition
– Too much task analysis and KE, too little cognitive theory
• Task is too narrow; too much data available
– Reduces to data fitting – favors parameterization over
principle
• Task is too specialized (typical cognitive psychology)
– Single cognitive aspect – misses generality, integration
• Lack of common simulation environment
– Each framework/theory only tackles what they do well
• Lack of comparable human data
3/7/09
Artificial General Intelligence
– Emphasizes functionality
– losesConference
cognitive constraints
3
Desirable Challenge Attributes
• Lightweight
– Limit integration overhead and task analysis/knowledge eng.
• Fast
– Rapid model development and collection of monte carlo runs
• Open-ended and dynamic
– Less parameterization, generalization to emergent behavior
• Simple and tractable
– Direct relation from cognitive mechanisms to behavioral data
• Integrated
– Toward integrated agent capturing architectural interactions
3/7/09
Artificial General Intelligence Conference
4
DSF Challenge Comparison
• Dynamic Stocks and Flows – Instance of Dynamic Decision Making
– Control a dynamic system given unexpected environmental fluctuations
– Simple version of real-world situations (financial, ecological, technical,
game)
• Integrated tasks
– Anticipate events
– Control system
• Cognitive functions
– Sequence learning - PC
– Action selection - BG
• Implementation
– VB on Windows
– Text socket protocol
3/7/09
Artificial General Intelligence Conference
5
Generalization Scenarios
• Humans learn to control system over time for simple functions
– Highly variable but quantifiable performance over learning process
– Complexity of task scalable along a number of cognitive dimensions
• Environmental i/o
– Complex sequences
– Stochastic noise
– Multiple variables
• System dynamics
– Feedback delay
– Non-linear effects
– Real-time control
• Multi-agent system
– Other controllers
– Payoff manipulations
3/7/09
Artificial General Intelligence Conference
6
DSF Comparison Schedule
• Official announcement expected March 15
• Task environment with socket connection for model,
data and documentation available on web site
• Symposium April 1st at BRIMS conference (Sundance)
• Model submission by May 15
• Best entries invited to symposium at European
cognitive modeling conference (travel supported)
• Email [email protected] to be added to
distribution list for official announcements/updates
3/7/09
Artificial General Intelligence Conference
7