Transcript Applied Interactive Story

Artificial Intelligence and
Software that Learns and Evolves
DIG 3563 – Fall 13
Dr. J. Michael Moshell
University of Central Florida
Adapted from A Special Presentation
for Ajou University
Autumn 2013
hplusmagazine.com
1
The Plan of the Lecture
0: What is a problem? What is intelligence?
1. The classical approach: logic and deduction
2. The knowledge-based approach: large databases
3. Cognitive science: models of human reasoning
4. Evolutionary Computing
-2 -
0: What is a Problem?
"Something that is difficult to deal with."
(Dictionary definition)
-3 -
0: What is a Problem?
"Something that is difficult to deal with."
(Dictionary definition)
For a small child, this is a problem:
Anna had $2.00. She spent $0.75 for candy.
How much money does Anna have now?
www.towngreendistrict.co
m
-4 -
0: What is a Problem?
"Something that is difficult to deal with."
(Dictionary definition)
For a small child, this is a problem:
Anna had $2.00. She spent $0.75 for candy.
How much money does Anna have now?
For the President of the United States, this is a problem:
www.nps.gov
Can we change the laws so that everyone has a job,
and the economy grows in a safe, steady fashion?
-5 -
Classifying Problems
Problems
Well-formulated problems
other problems
clear goals
mixed goals
limited action space
clear rules
infinite action space
rules are changing
-6 -
Classifying Problems
Problems
Well-formulated problems
other problems
clear goals
mixed goals
limited action space
clear rules
infinite action space
rules are changing
Easy
problems
Tractible
problems
kardwell.com
Intractible
problems
en.wikipedia.org
artsbeat.blog.nytimes.com
-7 -
Tractible: definition
Easily handled or worked.
examples:
Wood is a tractible material for making furniture.
bizchair.com
OPPOSITE: intractible
Titanium is an intractible material for making furniture.
worldchair.com
-8 -
The Traveling Salesman Problem
A man must visit 50 cities. He must visit each city ONE TIME.
Find the shortest path for his travel.
-9 -
The Traveling Salesman Problem
A man must visit 50 cities. He must visit each city ONE TIME.
Find the shortest path for his travel.
A man must visit n cities. He must visit each city ONE TIME.
Find the shortest path for his travel.
How long to compute?
-10 -
The Traveling Salesman Problem
A man must visit 50 cities. He must visit each city ONE TIME.
Find the shortest path for his travel.
A man must visit n cities. He must visit each city ONE TIME.
Find the shortest path for his travel.
time = k c n
How long to compute?
(for some constants k and c).
As n gets large, time gets VERY BIG VERY FAST
-11 -
The Traveling Salesman Problem
for k=1 microsecond
and c=2,
50 cities takes
313,000 hours or
35 years!
-12 -
Classifying Problems
Problems
Well-formulated problems
other problems
clear goals
mixed goals
limited action space
clear rules
infinite action space
rules are changing
Easy
problems
Tractible
problems
kardwell.com
Intractible
problems
en.wikipedia.org
In 1975:
artsbeat.blog.nytimes.com
-13 -
IBM's Deep Blue
Chess Playing Computer
In 1989, IBM's computer and programming
team defeated Garry Kasparov, world
chess champion.
ibm.com
It did not defeat the exponential time cost of chess.
It simply made k and c small enough, and explored more
futures than the human could.
en.wikipidia.org
-14 -
Classifying Problems
Problems
Well-formulated problems
other problems
clear goals
mixed goals
limited action space
clear rules
infinite action space
rules are changing
Easy
problems
Tractible
problems
kardwell.com
Intractible
problems
en.wikipedia.org
In 1990:
artsbeat.blog.nytimes.com
-15 -
trim-a-tree.co.uk
Decision Trees and
Exponential Time-Cost
Many problems are analyzed by building a decision tree
and seeking a path to a winning node. Here, n=9 (nine options)
en.wikipidia.org
-16 -
Decision Trees and
Exponential Time-Cost
trim-a-tree.co.uk
If each decision leads to a growing tree of other decisions,
the time required to explore all the branches time
=kcn
and that is too long
for anything but
very small n.
en.wikipidia.org
-17 -
Heuristic: A plan to
choose options that are 'most
likely to succeed'
trim-a-tree.co.uk
Eliminate those branches that your heuristic function tells you
are not likely to succeed. Then expand the promising ones.
en.wikipidia.org
-18 -
trim-a-tree.co.uk
Heuristic: A plan to
choose options that are 'most
likely to succeed'
A simple heuristic from chess:
Pawn=1 unit
Do not exchange pieces
if you lose more pawn-units
than your opponent loses.
Knight,
Bishop=3 pawns
Rook=5 pawns
Queen=9 pawns
-19 -
Heuristic: A plan to
choose options that are 'most
likely to succeed'
trim-a-tree.co.uk
A simple heuristic from chess:
Pawn=1 unit
Do not exchange pieces
if you lose more pawn-units
than your opponent loses.
Knight,
Bishop=3 pawns
Example:
Rook=5 pawns
Do not exchange your queen
for two knights.
Queen=9 pawns
-20 -
Intelligence = Problem Solving Ability?
zmescience.com
Most people agree that an intelligent agent
must be able to solve some problems (not all problems.)
However,
Many people feel that if you have a well-formed problem,
the hard work has already been done.
The BIG challenge is transforming a real-world
problem into a well-formed symbolic problem.
-21 -
Natural Language: a great place
to find ill-formed problems
zmescience.com
Imagine a computer program that could answer questions:
"Can a cat drive a car?"
Computer and Program
worldoffemale.com
-22 -
Natural Language: a great place
to find ill-formed problems
zmescience.com
Imagine a computer program that could answer questions:
"Can a cat drive a car?"
Computer and Program
"No. A cat has no hands and cannot drive a
car."
-23 -
The Turing Test for Intelligence
Alan Turing was a British mathematician who
thocp.net
played a key role in World War II code-breaking
and helped to develop the digital computer.
He thought about intelligence
and proposed a test.
-24 -
The Turing Test for Intelligence
Is "mystery system" intelligent?
thocp.net
Ask questions via a Teletype machine.
Mystery
System
Is "mystery system" a human or
a machine? If you cannot accurately
decide (and it's a machine) then
the machine is intelligent.
-25 -
The Turing Test for Intelligence
Has any system passed the Turing Test yet?
scoopertino.com
Ask Siri ...
www.apple.com
Most people quickly conclude that Siri does not yet pass
the Turing Test. But it's getting better all the time...
-26 -
1. The Classical (Logical) Approach
to Artificial Intelligence
Basic concepts:
hci.stanford.edu/~wino
grad
1. LOGIC is powerful enough to solve AI problems.
2. KNOWLEDGE must be represented in a formal system.
3. INFERENCE is the key mechanism to answer questions.
All humans will die.
John is a human
therefore, John will die.
-27 -
1. The Classical (Logical) Approach
to Artificial Intelligence
Knowledge representation as a "semantic net"
hci.stanford.edu/~wino
grad
of related concepts
en.wikipedia.org
-28 -
1. The Classical (Logical) Approach
to Artificial Intelligence
Example: Terry Winograd's SHRDLU System
hci.stanford.edu/~wino
grad
A "Toy world" of colored blocks (simulated by computer)
Questions and commands (in English):
1) Translate into formal propositions
2) Try to prove or disprove them
from the known facts.
3) Change system state if possible.
University of Utah
-29 -
1. The Classical (Logical) Approach
to Artificial Intelligence
Example: Terry Winograd's SHRDLU System
hci.stanford.edu/~wino
grad
Person: Pick up a big red block
Computer: OK
Person: Grasp the pyramid
University of Utah
-30 -
1. The Classical (Logical) Approach
to Artificial Intelligence
Example: Terry Winograd's SHRDLU System
hci.stanford.edu/~wino
grad
Person: Pick up a big red block
Computer: OK
Person: Grasp the pyramid
Computer: I don't understand which
pyramid you mean.
(because there are two of them.)
University of Utah
-31 -
1. The Classical (Logical) Approach
to Artificial Intelligence
Example: Terry Winograd's SHRDLU System
Watch the SHRDLU movie (3 minutes 20 seconds of it)
University of Utah
-32 -
1. The Classical (Logical) Approach
to Artificial Intelligence
Excitement! SHRDLU worked for Blocks World.
hci.stanford.edu/~wino
grad
followed by
Disappointment: Most domains are MUCH harder.
-33 -
2. The Knowledge-Based Approach:
Doug Lenat's talk at Google: Brittle Software
(Lenat video: first 14 minutes)
-34 -
2. The Knowledge-Based Approach:
Key concept: Today we have brittle (easily broken) software
Danger: Power is in the hands of "smart idiots".
Examples of Cyc's successes:
Request: Find a picture of someone smiling
 Cyc found picture of a man helping his daughter
take her first step
Request: Find something that could harm an airplane
 Cyc located a video about an SA-7 missile
-35 -
2. The Knowledge-Based Approach:
LARGE databases of facts.
If SHRDLU's world was too small,
let's build a big world of knowledge.
Cyc Project – started in 1984 by Douglas Lenat
Estimated effort (1986): 250,000 rules
and 350 man-years of effott.
Up until now: >1 million rules, and no end in sight.
-36 -
2. The Knowledge-Based Approach:
LARGE databases of facts.
If SHRDLU's world was too small,
let's build a big world of knowledge.
Cyc Project – started in 1984 by Douglas Lenat
cYcorp distributes the OpenCyc 4.0 database (for free), with
~ 239,000 terms
~ 2,093,000 "triples" (rules) that attempt to represent
human common sense.
-37 -
2. The Knowledge-Based Approach:
cYcorp also has a private database with many
cycorp.org
more assertions and rules, in the CycL language.
Example:
(#$isa #$BillClinton #$UnitedStatesPresident)
-38 -
Cyc: An example of the complexity
cycorp.org
University of Utah
-39 -
Cyc: Method for Growing the Database
* Attempt to automatically read encyclopedia articles.
(enCYClopedia!)
* Analyze successes & failures
* Apply human "knowledge engineering" to improve rules
-40 -
Cyc example: Terrorism Database
* Analyze literature on terrorism
* Predict future events.
Success:  predicted anthrax mailings, 6 months before 9/11
Miss:  Predicted 1000 dolphins from Al-Qaeda to attack
Hoover Dam
www.usbr.gov
-41 -
Cyc: Status and Hope for the Future
Cyc will eventually become smart enough to teach itself.
The results thus far:
* Government sponsors basic research and terrorism database
* Some commercial applications are being tried.
-42 -
Cyc: Status and Hope for the Future
Cyc will eventually become smart enough to teach itself.
The results thus far:
* Government sponsors basic research and terrorism database
* Some commercial applications are being tried.
* Many people in the Artificial Intelligence community doubt
that Cyc will play a key role in successful AI
Why? It's too logical. Humans are inconsistent, emotional,
intuitive – they act on their FEELINGS --
-43 -
3. Cognitive Science –
How Humans Think
wikimedia commons
-44 -
Philosophy:
Example: Mind-Body Problem
Is the mind part of the body? Or separate?
Metaphors:
"The brain is a telephone switchboard"
"The brain is a computer" 
Mind is software (can be changed)
Brain is hardware (can be broken)
 New ideas on good and evil
-45 -
Philosophy:
Example: Deductive Logic
If A, and AB, then B
A: A Hyundai is a car
B: Cars are made by humans
so: Hyundais are made by humans
-46 -
Philosophy:
Inductive logic:
If the events in class C are probable,
and A is in class C, then
A is probable.
90% of humans are right-handed.
Jack is a human.
so Jack is probably right-handed.
-47 -
Philosophy and Intelligence
If a thing is intelligent,
we expect it to use deductive logic
and inductive logic.
-48 -
Psychology:
Definition: Study of mental functions and behaviors
Example: Memory
wikipedia.org/mimory
-49 -
Psychology:
Definition: Study of mental functions and behaviors
Some types of long-term memory:
Procedural (how to do something)
-50 everyculture.com
Psychology:
Definition: Study of mental functions and behaviors
Some types of long-term memory:
Procedural (how to do something)
Topographic (where am I, where am I going)
mycharlois.com
-51 -
Psychology:
Definition: Study of mental functions and behaviors
Some types of long-term memory:
Procedural (how to do something)
Topographic (where am I, where am I going)
Episodic (what happened)
-52 fleetowners.com
Psychology:
en.wikipedia.org
Definition: Study of mental functions and behaviors
Some types of long-term memory:
Procedural (how to do something)
Topographic (where am I, where am I going)
Episodic (what happened)
Semantic (facts, definitions, abstract knowledge)
-53 -
Psychology:
Definition: Study of mental functions and behaviors
Some types of long-term memory:
Procedural (how to do something)
Topographic (where am I, where am I going)
Episodic (what happened)
Semantic (facts, definitions, abstract knowledge)
Visual (I've seen this before)
bic.org
-54 -
Psychology:
Definition: Study of mental functions and behaviors
Some types of long-term memory:
Procedural (how to do something)
Topographic (where am I, where am I going)
Episodic (what happened)
Semantic (facts, definitions, abstract knowledge)
Visual (I've seen this before)
gofamilyperks.com
Emotional (things I loved or hated)
-55 -
Psychology:
If a thing is intelligent,
we expect it to need (and have)
most of the types of memory
that humans have.
Why? (back to Philosophy!)
-56 -
Psychology:
If a thing is intelligent,
we expect it to need (and have)
most of the types of memory
that humans have.
Why? Inductive logic.
"Most of the intelligent creatures we have
seen, had these kinds of memory".
-57 -
Linguistics: Scientific
Study of Language
Key insight: Analogies carry meaning.
Definition: An analogy is a comparison of two
systems. If you understand system A, it can help
you to understand system B.
Analogy
Simile
Metaphor
-58 -
Linguistics: Scientific
Study of Language
Key insight: Analogies carry meaning.
Definition: An analogy is a comparison of two
systems. If you understand system A, it can help
you to understand system B.
Analogy:
Simile
Metaphor
"The motor of a car is like
a horse pulling a wagon."
-59 -
Linguistics: Scientific
Study of Language
Key insight: Analogies carry meaning.
Definition: An analogy is a comparison of two
systems. If you understand system A, it can help
you to understand system B.
Analogy:
Simile
Metaphor
"His mother was a tiger!"
-60 -
Linguistics: Scientific
Study of Language
Key insight: Analogies carry meaning.
Science is based on analogies.
Example: Bohr's "Solar system" model of the atom.
wikipediaorg
ou.org
-61 -
Linguistics:
If a thing is intelligent,
we expect that it will understand and use
a natural language (like English or
Korean).
and we expect that it will make and use
analogies to extend and communicate
its knowledge.
-62 -
AI and Cognitive Science:
Marvin Minsky makes
an analogy
Minsky's theory of mind:
The mind is like a complex software system.
The pieces of this software system will interact
in ways that are different from traditional software.
They will interact like a "society".
-63 -
AI and Cognitive Science:
Marvin Minsky makes
an analogy
Minsky's theory of mind:
•
•
•
A mind is a large collection of small agents
They compete for control of the
‘front office’ (consciousness)
The ‘all or none’ theory. You can’t half walk and half sit.
Many of these agents are working at any time
-64 -
Interior Grounding, Reflection and Self-Consciousness
* A woman named Joan is crossing the street.
A car sounds its horn.
-65 -
Interior Grounding, reflection and Self-Consciousness
* A story about a woman crossing the street.
•Reaction: Joan reacted quickly to that sound.
•Identification: She recognized it as being a sound.
•Characterization: She classified it as the sound of a car.
•Attention: She noticed certain things rather than others.
•Indecision: She wondered whether to cross or retreat.
-66 -
Interior Grounding, reflection and Self-Consciousness
* A story about a woman crossing the street.
•Reaction: Joan reacted quickly to that sound.
•Identification: She recognized it as being a sound.
•Characterization: She classified it as the sound of a car.
•Attention: She noticed certain things rather than others.
•Indecision: She wondered whether to cross or retreat.
•Imagining: She envisioned some possible future conditions.
•Selection: She selected a way to choose among options.
•Decision: She chose one of several alternative actions.
•Planning: She constructed a multi-step action-plan.
•Reconsideration: Later she reconsidered this choice.
-67 -
Marvin Minsky:
Interior Grounding, Reflection and Self-Consciousness
* These processes can be classified something like this
.. which is similar to
Freud’s model:
-68 -
Marvin Minsky:
Interior Grounding, reflection and Self-Consciousness
-69 -
Minsky doesn’t like the ‘bottom-up’ idea
that sensations (alone) could lead to higher thought.
He believes
in a rich set
of built-in
capabilities.
The details
of which
language, what
culture, what
house and street
are learned by each individual.
-70 -
Minsky's Influence
"Societies of Mind" has not yet led to a working AI system
.. but Minsky's early work led to the study of Neural Nets
(our next topic)
-71 -
5. Neural Nets, Perception and Learning
-72 -
4. Artificial Evolution
Nature "learns" by creating new species.
bio100.nicerweb.net
Can we model that process, to solve problems?
-73 -
Evolutionary Computing:
* Reviewing Genetics
• Sexual reproduction has a big payoff. What is it?
( In other words: why are males worth having?)
Observation: bacteria and viruses without SR
have evolved several mechanism for swapping DNA.
It’s almost as if the fundamental underlying
metaphor for life is a flea market.
www.ryctx.org
-74 -
Evolutionary Computing:
* Genetics Reviewed
KNOWN BEFORE DNA was discovered:
• The genome is a (very) long sequence of Genes
• Each gene controls the production of one kind of protein
• Proteins are catalysts for chemical reactions
as well as the ‘structural steel’ of living organisms.
A GENE represents a finite alphabet of choices.
The various versions of a gene are called alleles.
If there are 10 ways to make collagen, there would be
10 alleles for the collagen gene.
-75 -
Genotype and Phenotype
• Genotype: your collection of genes
Phenotype: your ‘rendering’ – your
actual body, as built.
• Genes, encoded in DNA, are organized into chromosomes
• Individual humans have 23 pairs of chromosomes
• When reproducing,
each parent randomly
contributes one of the
two chromosomes to
the child.
-76 -
Genotype and Phenotype
Mom
...
1
2
3
23
Dad
XX
XX
XX
X X
-77 -
Genotype and Phenotype
Mom
...
1
2
3
23
Dad
XX
XX
XX
X X
-78 -
Genotype and Phenotype
Mom
Dad
1 XX
2 XX
3 XX
...
23 X X
-79 -
Genotype and Phenotype
Mom
1 XX
2 XX
3 XX
...
23 X X
Dad
A given pair of parents can
produce 223 ~= 8 million
different genetic combinations.
.. it’s a GIRL!
-80 -
Why does this system pay such big dividends?
• The gene pool is a toolkit of variations.
• Consider melanin. Assume variations from black to brown
in various versions of the melanin gene.
• Your tribe moves from Africa to Europe.
• Your random genome remix produces kids of various shades.
The ones with lighter skin get more vitamin D and thrive.
They have more kids. The light-skin gene increases in the
gene pool. Feedback loop.
-81 -
Why does this system pay such big dividends?
• The gene pool is a toolkit of variations.
• Consider melanin. Assume variations from black to brown
in various versions of the melanin gene.
• Your tribe moves from Africa to Europe.
• Your random genome remix produces kids of various shades.
The ones with lighter skin get more vitamin D and thrive.
They have more kids. The light-skin gene increases in the
gene pool. Feedback loop.
NOTE: You didn’t have to INVENT the variation (mutation).
You had it stored away in your toolkit (genome).
Mutation (creation of new alleles or genes) is MUCH slower
than selection among existing alleles.
-82 You need BOTH mechanisms.
Mutation – the big Disaster/Opportunity
• Mutations are rare and usually fatal
• A copying error occurs in a chromosome
- some DNA is duplicated
- some DNA is deleted
- one codon (and its amino acid) replaces another
• Some mutations are beneficial but most are fatal or neutral (now)
• A slightly different kind of hemoglobin might not kill you
but might turn out to be BETTER, against some parasite
that attacks your great great great .... grandchildren
-83 -
Diversity yields robustness
• The environment produces an infinite suite of challenges.
• A rich gene pool provides instant options to try.
• A narrow gene pool is a ticket to extinction (florida panthers.)
• Hybrid vigor is a concept that every farmer knows.
Cross Hereford and Angus cows; calves grow faster.
-84 -
Diversity yields robustness
• The environment produces an infinite suite of challenges.
• A rich gene pool provides instant options to try.
• A narrow gene pool is a ticket to extinction (florida panthers.)
• Hybrid vigor is a concept that every farmer knows.
Cross Hereford and Angus cows; calves grow faster.
It’s like NAFTA or the European Union.
Win-win really is possible! Your kids will survive better
if your partner’s tool-set complements rather
than replicates your own.
-85 -
Unnatural selection
• To build a “learning system” we need three things:
-a genotype (a coded representation)
-a phenotype (a rendering into a ‘real world’ of competition)
-a fitness function (something to measure and kill the losers)
-86 -
Unnatural selection
• To build a “learning system” we need three things:
-a genotype (a coded representation)
-a phenotype (a rendering into a ‘real world’ of competition)
-a fitness function (something to measure and kill the losers)
Is a self-replicating robot without a genome impossible?
That is not proven. But all examples thus far are trivial.
(Crystal growth)
www.dpchallenge.com
-87 -
A Genotype for Mini-Robots
• Karl Sims decided to use a graph-theory genome
• It is applied twice: once for body, once for nervous system
• A random pool of 300 genomes is built
-they are pre-selected by removing:
- creatures with more than N body parts
- creatures whose body parts interpenetrate (share space)
- Rules of the universe are established; e. g. gravity, a floor
- Goal (fitness function) is set: e. g. radius crawled in 1 minute.
-Run simulation. Keep best 1/5 of population (60 individuals)
Re-mix genes to replace the 240 who died.
Run the simulation again.
-88 -
A Genotype for Mini-Robots
• Some of the goals:
- radial distance traveled
- linear distance traveled
- distance swum (or flown) through fluid medium
- speed of approach toward a moving target point
- competition to capture a shared object
-89 -
A Genotype for Mini-Robots
• Some of the goals:
- radial distance traveled
- linear distance traveled
- distance swum (or flown) through fluid medium
- speed of approach toward a moving target point
- competition to capture a shared object
Competitive events: how do you pair them up?
- n x n takes n2 time, and is too slow (each sim is slow!)
- pairwise often means playing against an idiot.
- n vs. best-of-last-round seemed to work well.
-90 -
A Genotype for Mini-Robots
• Some of the goals:
- radial distance traveled
- linear distance traveled
- distance swum (or flown) through fluid medium
- speed of approach toward a moving target point
- competition to capture a shared object
Competitive events: how do you pair them up?
- n x n takes n2 time, and is too slow (each sim is slow!)
- pairwise often means playing against an idiot.
- n vs. best-of-last-round seemed to work well.
One-species versus two-species (breeding populations)
-91 -
NOW watch the movie at
http://www.youtube.com/watch?v=JBgG_VSP7f8
www.dpchallenge.com
-92 -
A Genotype for Mini-Robots
• So ... how was this done?
NODE and LINK
(The names are just
to help us think.)
Example 1:
Example 1:
From a segment,
link to two other segments.
Repeat any number
of times, recursively.
-93 -
A Genotype for Mini-Robots
• So ... how was this done?
NODE and LINK
(The names are just
to help us think.)
Example 2:
From a body segment,
link to one other body seg.
and two leg segments.
From a leg segment
link once to another leg segment.
Example 2:
-94 -
A Genotype for Mini-Robots
• So ... how was this done?
NODE and LINK
(The names are just
to help us think.)
Example 3:
From a body,
link to a head & four limbs.
From a limb, link to
another limb.
Example 3:
www.dpchallenge.com
-95 -
Brains and bodies
• Each sensor is contained in a specific body part
• Sensors measure joint angles, forces, properties of the world
• The brain is a network of neurons – (but not like real ones)
• Neurons’ functions include: sum, product, sum-threshold,
greater than, .... sin, cos, log, integrate, differentiate,
... smooth, memory, oscillate-wave, oscillate-sawtooth
-96 -
Neurons in Segments
• P0, P1 are photosensors
• C0 and Q0 are contact sensors
• E0 and E1 are effectors (joint angle drivers)
The connections are evolved, not reasoned out.
(There is a graph genome for the neurons, too.)
-97 -
Neurons in Segments
• P0, P1 are photosensors
• C0 and Q0 are contact sensors
• E0 and E1 are effectors (joint angle drivers)
The connections are evolved, not reasoned out.
(There is a graph genome for the neurons, too.)
-98 -
Neurons in Segments
• A single shared neuron
group is also provided.
(“where” is it? Unspecified)
This capability allows for
coordinated control.
-99 -
Neurons in Segments
• A single shared neuron
group is also provided.
(“where” is it? Unspecified)
This capability allows for
coordinated control.
The saw and wav oscillators
are key elements.
-100 -
What Changes in each Generation?
•
NOTE: The system mixes sexual reproduction with mutation
in an un-biological way: mutation occurs in every generation.
MUTATION
1. Internal parameters (weights, oscillation frequencies) are
randomly altered. Small alterations more likely than big ones.
2. A new random node is added to graph. (May not connect; will be
discarded if not.)
3. New random connections are added, existing ones are removed.
4. Unconnected elements are garbage-collected.
Outside (morphology) graphs are altered, then inside (neuro) ones.
-101 -
What Changes in each Generation?
MATING the GRAPHS
a. Crossover operation.
A subset of parent 2 is inserted to replace a subset of parent 1
-102 -
What Changes in each Generation?
MATING the GRAPHS
a. Crossover operation.
A subset of parent 2 is inserted to replace a subset of parent 1
-103 -
What Changes in each Generation?
MATING the GRAPHS
b. Grafting operation.
Two parents are joined together (each loses one node)
-104 -
Results
- Interbreeding populations often converge to uniformity, but
- Successive runs often produce totally different results.
- Swimming produced:
- paddles
- tail wagglers
- specialized scullers
- lots of flippers
- water snakes
-105 -
Results
- Interbreeding populations often converge to uniformity, but
- Successive runs often produce totally different results.
- Walking produced:
- corner-walkers
- rocking blocks
- inchworms
- legs
- hoppers
Light-following worked in
walking and swimming
environments.
-106 -
What happened next?
- not much (at least, nothing so spectacular as Sims’ creatures.)
Why?
- The leap from simple goal-seeking motor activity (“tropisms”) to
interesting perception and cognition is verrrrrry looooong.
- Folks like Brooks and Minsky’s successors are trying
to bridge the gap.
- Fundamental insights are still needed.
-107 -