Transcript Lecture 14

CAP6938
Neuroevolution and
Artificial Embryogeny
Evolving Adaptive Neural
Networks
Dr. Kenneth Stanley
March 1, 2006
Remember This Thing?
What’s missing from current neural models?
An ANN Link is a Synapse
(from Dr. George Johnson at http://www.txtwriter.com/Backgrounders/Drugaddiction/drugs1.html )
What Happens at Synapses?
• Weighted signal transmission
• But also:
– Strengthening
– Weakening
– Sensitization
– Habituation
– Hebbian learning
– None of these weight changes during a
lifetime are happening in static models!
Why Should Weights Change?
•
•
•
•
The world changes
Evolution cannot predict all future possibilities
Evolution can succeed with less accuracy
The Baldwin Effect
– Learning smooths the fitness landscape
– Traits that initially require learning eventually become
instinct if the environment is consistent
• If the mind is static, you can’t learn!
How Should Weights Change?
• Remember Hebbian Learning? (lecture 3)
– Weight update based on correlation:
– Incremental version: wi   xi y
wi (new)  wi (old )  xi y
• How can this be made to evolve?
– Which weights should be adaptive?
• Which rule should they follow if there is more than one?
– Which weights should be fixed?
– To what degree should they adapt (evolve alpha)
• Evolve alpha parameter on each link
Floreano’s Weight Update
Equations
• Plain Hebb Rule:
• Postsynaptic rule:
– Weakens synapse if postsynaptic node fires
alone
• Presynaptic rule:
• Covariance rule:
Strengthens when
correlated,
weakens when not
Floreano’s Genetic Encoding
Experiment: Light-switching
Fully Recurrent Network
• Task: Go to black area to turn on light,
then go to area under light
• Requires a policy change in mid-task:
Reconfigure weights for new policy
Blynel, J. and Floreano, D. (2002) Levels of Dynamics and Adaptive Behavior in Evolutionary Neural Controllers. In B. Hallam, D. Floreano, J. Hallam, G. Hayes, and J.-A.
Meyer, editors. From Animals to Animats 7: Proceedings of the Seventh International Conference on Simulation on Adaptive Behavior, MIT Press.
Results
• Adaptive synapse networks evolved
straighter and faster trajectories
• Rapid and appropriate weight
modifications occur at the moment if
change
However, It’s Not That Simple
• A recurrent network with fixed synapses
can change its policy too
• The activation levels cycling through the
network are a kind of memory that can
affect its functioning
• Do we need synaptic adaptation at all?
• Experiment in paper:
Kenneth O. Stanley, Bobby D. Bryant, and Risto Miikkulainen (2003). Evolving Adaptive Neural
Networks with and without Adaptive Synapses, Proceedings of the 2003 IEEE Congress on
Evolutionary Computation (CEC-2003).
Experimental Domain: Dangerous
Food Foraging
•
•
•
•
Food may be poisonous or may not
No way to tell at birth
Only way to tell is to try one
Then policy should depend on “pain” or not
Condensed Floreano Rules
• Two adaptation rules: One for excitatory
connections, the other for inhibitory:
• First term is Hebbian, second term is a
decay term
NEAT Trick: Use “Traits” to Prevent
Dimensionality Multiplication
•
•
•
•
One set of rules/traits
Each connection gene points to one of the rules
Rules evolve in parallel with network
Weights evolve as usual
Robot NNs
Surprising Result
• Fixed-weight recurrent networks could
evolve a solution more efficiently!
• Adaptive networks found solutions, but
more slowly and less reliably
Explanation
• Fixed networks evolved a “trick”: Strong inhibitory
recurrent connection on left turn output cause it to stay
on until it experiences pain. Then it turns off and robot
spins (from right turn output) until it doesn’t see food
anymore, and it runs to the wall
• In adaptive network, 22% of connections diverge after
pain, causing network to spin in place: a holistic change
Discussion
• Adaptive neurons are not for everything, not
even all adaptive tasks
• In non-adaptive tasks, they only add
unnecessary dimensions to the search space
• In adaptive tasks, they may be best for tasks
requiring holistic solutions
• What are those?
• Don’t underestimate the power of recurrence
Next Topic:
Leaky Integrator Neurons,
CTRNNs, and Pattern Generators
• Real neurons encode information as spikes and spike trains with
differing rates
• Dendrite may integrate spike train at different rates
• Rate differences can create central pattern generators without a
clock!
Levels of dynamics and adaptive behavior in evolutionary neural controllers by Blynel, J., and Floreano,
D. (2002)
Evolution of Central Pattern Generators for Bipedal Walking in a Real-Time Physics Environment by
Torsten Reil and Phil Husbands (2002)
Optional: Evolution and analysis of model CPGs for walking I. Dynamical modules by Chiel, H.J., Beer,
R.D. and Gallagher, J.C. (1999)
Homework due 3/8/06 (see next slide)
Homework Due 3/8/06
Genetic operators all working:
•Mating two genomes: mate_multipoint, mate_multipoint_avg, others
•Compatibility measuring: return distance of two genomes from each other
based on coefficients in compatibility equation and historical markings
•Structural mutations: mutate_add_link, mutate_add_node, others
•Weight/parameter mutations: mutate_link_weights, mutating other parameters
•Special mutations: mutate_link_enable_toggle (toggle enable flag), etc.
•Special restrictions: control probability of certain types of mutations such as
adding a recurrent connection vs. a feedforward connection
Turn in summary, code, and examples demonstrating that all
functions work. Must include checks that phenotypes from
genotypes that are new or altered are created properly and work.
Project Milestones (25% of grade)
•
•
•
•
•
•
2/6: Initial proposal and project description
2/15: Domain and phenotype code and examples
2/27: Genes and Genotype to Phenotype mapping
3/8: Genetic operators all working
3/27: Population level and main loop working
4/10: Final project and presentation due (75% of grade)