step - Ben Goertzel

advertisement
Novamente:
An Integrative Approach
to Embodied
Artificial General
Intelligence
Ben Goertzel
Novamente LLC
Overview
Artificial General Intelligence
Historical Approaches to AI
The Importance of Embodiment
Stages of Cognitive Development
The AGI-SIM Simulation World
The Novamente AI Architecture
Learning to Play Fetch and Tag
Object Permanence and Word-Object
Association
Steps Toward Human-Level AI
Artificial General
Intelligence (AGI)
“The ability to achieve complex goals in complex
environments using limited computational resources”
• Autonomy
• Practical understanding of self and others
• Understanding “what the problem is” as opposed to
just solving problems posed explicitly by
programmers
Existing AI Paradigms
Paradigm
Strengths
Weaknesses
GOFAI
Representation of abstract
knowledge
Reasoning (short proofs)
Pattern recognition
Learning
Autonomy
Neural nets
Pattern recognition
Learning
Associative memory
Perception/action/cognition
integration
Representation of abstract
knowledge
Abstract reasoning
Learning
Autonomy
Evolutionary
Programming
Pattern recognition
Learning of complex procedures
Representation of abstract
knowledge
Abstract reasoning
Autonomy
Probabilistic
Reasoning
Representation of abstract
uncertain knowledge
Reasoning (short proofs)
Hypothesis formation
Autonomy
Pattern recognition
Subsumption
Robotics
Autonomy
Learning
Perception-action integration
Cognition
Representation of abstract
knowledge
The Power of
Embodiment
Embodiment (real or virtual) provides a would-be
AGI with
Symbol grounding

Most crucially: grounding of subtle words like prepositions
An effective medium for learning complex cognitive skills



attention allocation
procedure-learning
inference control
A sense of self

Critical for cognition as well as mental health
Empathy with humans
Stages of Cognitive
Development
(a path toward embodied AGI, inspired
by the classic work of Jean Piaget)
Infantile
Concrete Operational
•
Instinct
•
Conservation Laws
•
Imitation
•
Theory of Mind
•
Word-Object Association
•
Complex object classifications
•
Object Permanence
•
Advanced syntax
Pre-Operational
Formal
•
Simple syntax
•
Abstract deductive reasoning
•
Systematic word-object
associations
•
Scientific testing of hypotheses
•
Single-parameter object
classifications
AGI-SIM:
An Open-Source Simulation
Environment for AGI
AI systems can sense and act in real-time via
embodiment in a 3D virtual world
Uses CrystalSpace (open-source game engine)
for visualization
Provides AI systems with multisensory inputs
 visual inputs at varying levels of granularity:
pixels, polygons or objects
 hearing, touch, proprioception, …
Integration with natural language interface for
fluid, situated communication
Suitable for teaching/learning based on a
developmental-psychology-based methodology
Compatible with Novamente but usable by any
AI system via a simple sockets-based protocol
AGI-SIM Status
Version 1 complete and usable

Available on SourceForge
Work in progress:



More complete set of wireframe models
Multi-room simulated world
Realistic physics


Via integration of Open Dynamics Engine
Integration of NLP chat
Post-Embodied AI
AI systems may viably synthesize
knowledge gained via various means
virtually embodied experience

AGI-SIM
physically embodied experience

robotics
explicit encoding of knowledge in natural language
ingestion of databases

quantitative and relational
The Novamente Project
Long-term goal:


creating "artificial general intelligence" approaching and then exceeding the
human level
to be approached via a series of incremental phases
Novamente AI Engine: an integrative AI architecture



synthesizes ideas from cognitive and neural science with computer science
algorithms such as evolutionary programming and probabilistic inference
uses a unique "weighted labeled hypergraph" knowledge representation
Efficient, scalable C++/Linux implementation
Currently parts of the Novamente codebase are being used for
commercial projects


natural language processing
biological data analysis
The Novamente Project
Moving Toward General Intelligence
Novamente is being used to control an embodied virtual agent in a 3D
simulation world called AGI-SIM.
A loosely-Piaget-inspired series of cognitive developmental stages is
being used to guide the process of teaching Novamente to carry out a
series of progressively more complex tasks in the AGI-SIM
environment.
The same approach being used within AGI-SIM may be used in future
to embody Novamente in physical robots, including humanoid robots or
automated vehicles.
Key Conceptual Aspects
of Novamente
Knowledge representation is transparent wherever possible
Procedural and declarative knowledge are easily
interconvertible
Evolutionary learning is used to supply creativity for both
procedural and declarative knowledge
Probabilistic logical inference is used for basic reinforcement
learning as well as abstract thinking
Architecture supports interaction of multiple specialized agents
acting on a common knowledge store, along with processes of
truly general scope
Design covers all aspects of human and machine cognition
Key Novamente
Design Aspects
Aspect
Function
Node
Link
Denotes a percept, set, list or action
MindAgent
A persistently active cognitive, perceptual
or active process
Core
The “Mind OS” that maintains the store of
nodes/links and executes MindAgents
and Tasks
Map
A semantically significant pattern of
activation of MindAgents, Nodes and
Links
Unit
A collection of nodes, links and MindAgents
grouped together to carry out some particular
function (like perception, language processing,
or abstract reasoning)
Denotes a logical, associative or action
relationship
Novamente
Node and Links
Node types:
 WordNode, NumberNode, CharacterNode,…
 ConceptNode, ProcedureNode, PredicateNode,…
Link types:
 SimilarityLink, InheritanceLink, ImplicationLink,
EquivalenceLink,…
 ProcedureApplicationLink, PredicateEvaluationLink,…
Nodes and links weighted with:
 Probabilistic truth values
 Attention values similar to neural network activations
Architecture of a
Novamente Lobe
MindAgents are based on
•Probabilistic Term Logic
•BOA-based Evolutionary Procedure
Learning
•Frequent Itemset Mining
•Stochastic Local Search
Example MindAgents:
•First-order probabilistic inference
•Schema Learning with BOAP
•Probabilistic Attention Allocation
•Procedure and Predicate Evaluation
•Sentence Parsing
•Word Sense Disambiguation
•Sentence Production
•…
Distributed
Novamente Architecture
Novamente
Architecture:
High-Level View
Learning to Play Fetch
ifelse holding
(ifelse facingteacher
step
rotate)
(ifelse nearball
pickup
(ifelse facingball
step
rotate))
Example program for
learning to fetch a ball
and bring it to the
teacher, learned using
Novamente’s BOA
procedure learning
algorithm
Learned via giving
Novamente partial
reinforcement for
bringing the ball near the
teacher
Learning to Play Fetch
Interpretation of Example Program
If holding the ball and facing the teacher move
forward (to give the ball to the teacher)
Otherwise, if holding the ball, rotate ( in order to face
the teacher)
Otherwise, if near the ball, pick it up
Otherwise, if facing the ball, move forward (to get the
ball)
Otherwise, rotate (in order to face the ball)
Learning to Play Tag
Step size significantly smaller than
robot diameter
(ifelse it
opponentangle
- opponentangle 3.1416)
step
Step size equal to robot diameter
(ifelse it
opponentangle
- opponentangle 1.5708)
step
Agents controlled by
BOA learn rules for
playing “tag” via
reinforcement learning
in a tag tournament
Optimal rules learned
depend on ratio
between agents’ speed
and step size
An example of
“cooperative learning”
among a community of
agents
Learning to Play Tag
Not-IT’s
Great Escape
Next Steps
Fall 2005 Milestones:
Oct. -- Complete integration of Novamente
reasoning/learning with AGI-SIM
Nov./Dec. -- Complete first two “infantile
stage” tasks:


Object permanence
Word-object association
Object Permanence
Word-Object Association
Long-Term
Novamente Project
Goals
Phase One: Definition of design and implementation of initial versions of key components

Goal: a teachable AI system capable of embodiment in a simulation world
Phase Two: Refinement of design and implementation in the course of teaching the AI
system to control an agent in a simulation world, according to a loosely Piagetan learning
plan


Goal: an “artificial child” with qualitatively complex though not humanlike English conversation ability,
and the approximate problem-solving ability of an average ten-year old human child within the context
of its simulation world
Optionally: Initiate parallel development in “real robotics” via partnership with a robotics research
team
Phase Three: Instruction of “artificial child” in relevant topics, including ethics and science
Phase Four: Instruction of AI system in AI design and general computer science

Goal: an ethical AI capable of modifying its own implementation with a goal of self-improvement
Credits
Novamente:
Cassio Pennachin
Moshe Looks
Ari Heljakka
Andre Senna
Izabela Goertzel
Welter Silva
Michael Ross
Hugo Pinto
Rodrigo Barra
AGI-SIM:
Ari Heljakka
Welter Silva
Download