Title: CAP6938 Neuroevolution and Developmental Encoding Non-Neural NEAT and Closing Remarks
1CAP6938Neuroevolution and Developmental Encoding
Non-Neural NEAT andClosing Remarks
- Dr. Kenneth Stanley
- October 30, 2006
2Outline
- Complexification is a general concept
- Protecting innovation is a general concept
- Therefore, they can apply to anything without a
defined dimensionality - Example Cellular Automata
3Complexification is a General Concept
- Solving a smaller version of a problem and
expanding the solution - Making a rough estimate and refining it
- Building a structure piece by piece
- Elaboration of a pre-existing concept
4Complexification Does Not Mean Optimizing Random
DimensionsFrom a Set
- Example 10-dimensional search space
- Now hold d2 through d10 constant and search d1
- Once you get a good value for d1, start searching
both d1 and d2 together, and so on - This is not complexification
- It is a naïve search assuming independent
variables - Subject to simple deception
- Usually wont work
5Then What Does it Mean?
- Complexification means increasing information
about the solution - (Optimizing d1 does not increase general
information about the solution) - Initial dimensions are a complete solution on
their own (nothing is held at zero) - Complexification means finding the dimensionality
of the solution is part of the problem - A neural network can have any number of weight
dimensions and solve the same problem - Most dimensions outside the current structure
have no meaning on their own
6Example
1
3
2
- 3 dimensions
- Is dimension 2372 held at zero?
- What exactly is dimension 2372?
- It depends on how the other 2371 dimensions turn
out to relate to each other - It is undefined it doesnt exist
- Not like d10 in prior example, which always
exists - Complexification is searching infinite undefined
dimensions, or rather, it is not performing
search in the usual sense. It is increasing
information.
7Example 2
y
x
- Problem Find an expression of this function
- Complexification says start with a very low
dimensional approximation as accurate as possible
in its space - Red line 2-dimensional estimate ymxb
- Now we could add new terms and refine the
estimate - Analogous to bending the line like a rubber band
for each new dimension added - New estimate does not necessarily need exactly
the same term mxb
8Protecting Innovation is a General Concept
- New ideas need time to mature
- Children need time to grow up
- Ph.D. students need room to make mistakes
- Bigger often means slower, but not stupider
- Einstein was not the teachers pet
- The long run is what matters
- If we kicked him out early, wed all lose
9Speciation Protects Innovation
- An idea is represented as a niche
- The niche is a local, protected competition
- One niche does not directly compete with another
- Only the absolute worst are purged after
sufficient opportunity is spent
10General Concepts Means They Dont Have to Apply
to a Neural Network
- Complexification and protection of innovation go
hand in hand - In order to elaborate, one must protect potential
elaborations - In order to grow one must have room
11Novel Phenotype Cellular Automata
- Set of pixels that change over time according to
neighborhood rules - The Game of Life is a familiar example of 2D
cellular automata
From http//www.bitstorm.org/gameoflife/
122D Cellular Automata
- Pixels are in a line instead of a plane
- Change over time can be represented as a vertical
graph
time
From Melanie Mitchell, James P. Crutchfield, and
Rajarshi Das, "Evolving Cellular Automata with
Genetic Algorithms A Review of Recent Work", In
Proceedings of the First International Conference
on Evolutionary Computation and Its Applications
(EvCA'96), Russian Academy of Sciences (1996).
13Neighborhood Rules
- Next state for pixel determined by pixels in its
neighborhood within some radius
2(2r1) bits per rule table
From Melanie Mitchell, James P. Crutchfield, and
Rajarshi Das, "Evolving Cellular Automata with
Genetic Algorithms A Review of Recent Work", In
Proceedings of the First International Conference
on Evolutionary Computation and Its Applications
(EvCA'96), Russian Academy of Sciences (1996).
14Can It Do Anything Useful?
- Maybe it can compute functions
- Popular task Fill the line with whichever color
is in the majority (Density Classification) - Successful attempts (r3 128 bits/genome)
15Assessing Performance
- Measure correct over unbiased distribution of
many initial conditions - Best performance is 86 on 149 pixels with r3
(Juillé and Pollack 1998) using coevolution of
rules and initial conditions - Could we do a lot better than 86?
- Maybe with complexification
16Complexifying Cellular Automata
17Complexifying Cellular Automata
- How? Expand the neighborhood
- Neighborhood doesnt need to be symmetric or even
contiguous - Is this really complexification?
- Yes Unexpressed dimensions are undefined without
knowing all the dimensions - The initial rules give us only a little
information, but good information - The dimensions of the search space are the bits
of the rule, not the neighborhood positions - The rule includes the neighborhood positions,
i.e. there is structure. Position is a
historical marker in this case.
00 0
01 0
10 1
11 0
0 0
1 1
18Even More Abstract Complexification Solution
- A very wide neighborhood could be input into a
neural network that computes a function of those
inputs and outputs the next value for the bit in
the middle - The network that computes the function can
complexify
Evolved Topology
19Conclusion
- Complexification and protection of innovation may
allow more complex and therefore powerful
neighborhood functions to evolve (maybe beat 86
by using with coevolution?) - Complexification and protection of innovation may
allow far more complex solutions to anything
20NE DE What Have We Learned
- Search is not just optimization
- Expanding complexity over generations is a
powerful idea - Protecting innovation is as well
- Neural networks can be grown with this method
- The mapping between genotype and phenotype is
equally important - Reuse of genes is powerful
- The neural model can be enhanced in several ways
- NEAT can evolve any kind of structure, including
DE/indirect encodings, and CPPNs
21What Is Its Significance?
- These are the forces of nature
- We are unlocking natures box by understanding
the underlying algorithms - Much of the beauty and complexity of nature, and
civilization itself, resulted from these simple
processes - Biological evolution was an unguided process
- What will we create if we take its reigns and
guide it?
22Where is the Field?
- Two parallel streams
- How to evolve with complexification
- How to represent with reuse (DE)
- More progress so far on 1 than 2
- Indirect encoding is on the brink
- The two streams are merging (e.g. CPPNs)
- A complexifying system with an efficient encoding
(mapping) is the next generation system
23Next TopicsTechnical topics in implementing
complexifying evolutionary systems and presenting
results.
- Practical implementation issues
- Questions and group discussion/problem solving
- How to present research results
Reference Section 3 of Real-time Neuroevolution
in the NERO Video Game by Stanley, Bryant, and
Miikkulainen (2005). (Has most up-to-date NEAT
description)