Lecture 09 - Neuroevolution Flashcards
Neuroevolution
What is neuroevolution?
Artificial evolution of neural networks using genetic
algorithms (GA)
Neuroevolution
Where has neuroevolution shown promise?
Reinforcement learning tasks, such as robot control and
playing a game.
Neuroevolution
What is NEAT short for?
Neuroevolution of Augmenting Topologies
Neuroevolution
What is TWEANN short for?
Topology and Weight Evolving Artificial Neural Network
Neuroevolution
What’s the relationship between NEAT and TWEANN?
TWEANN is NEAT’s predecessor.
Neuroevolution
What is Direct encoding?
Genome explicitly defines the network
Neuroevolution
(@Rephrase) What is indirect encoding?
Genome has developing rules to grow the network
Neuroevolution
What is the advantage of direct encoding?
Genome is smaller
Neuroevolution
What is the disadvantage of direct encoding?
Reduced topology variety
Neuroevolution
What is NEAT’s solution to the Competing Conventions Problem?
Historical Marking
Neuroevolution
What is Historical Marking in NEAT?
Tracking of “innovations” by enumerating them with a unique ID called the “innovation number”.
Neuroevolution
What is the innovation number?
A rising ID number for each new “innovation” in NEAT.
Neuroevolution
What is an innovation in NEAT?
Any new gene in a genotype, i.e. addition of neurons (nodes) and connections/weights (vertices).
Neuroevolution
How does historical marking work during crossover?
- Matching genes are inherited randomly
Neuroevolution
What are disjoint genes in NEAT?
those that do not match in the middle
Neuroevolution
What are genes that do not match in the middle called?
Disjoint genes
Neuroevolution
What are excess genes in NEAT?
those that do not match in the end
Neuroevolution
What are genes that do not match in the end called?
Excess genes
Neuroevolution
Which parent are disjoint/excess genes inherited from?
The more fit parent
Neuroevolution
How does NEAT protect innovation?
Speciation
Neuroevolution
Why does NEAT use Speciation?
To protect innovation
Neuroevolution
What is speciation?
Separation orf solutions into “species”, where only individuals within the same species are allowed to reproduce with each other.
Neuroevolution
What’s the requirement for a species?
Having similar chromosomes
Neuroevolution
Describe the parts of the compatibility distance.
- cs: importance coefficients
- E: #excess genes
- D: #disjoint genes
- N: #genes in the larger genome
- W: Average weight differences of matching genes.
Neuroevolution
Describe the speciation process. (4)
- Using the distance measure “d” and a threshold “d_t”.
- A random genome from the previous generation represents a species.
- For every genome:
- If the genome is compatible with a species, add it to that species.
- Otherwise, create a new species from that genome.
Neuroevolution
What is adjusted fitness?
The individual’s non-adjusted fitness divided by #individuals in the population.
Neuroevolution
Why is adjusted fitness used?
The adjusted fitness allows the balance of members among all species in a generation
Neuroevolution
How does speciation protect innovation?
The number of offspring of a species is proportional to the sum of adjusted fitnesses of its members
Neuroevolution
How does extinction work?
If the maximum fitness of a species did not improve in 15 generations, the networks in the stagnant species are not allowed to reproduce
Neuroevolution
How does elitism work?
the champion of each species with more than five networks are copied into the next generation unchanged
Neuroevolution
What are the types of mutations in NEAT? (3)
- Add connection
- Add node
- Change weight connection
Neuroevolution
What is HyperNEAT?
A “Hypercube-based NEAT”. An indirect encoding where a CPPN defines the connectivity.
Neuroevolution
What is CPPN short for?
Compositional pattern-producing network
Neuroevolution
What does a Compositional pattern-producing network (CPPN) do?
It produces a geometric pattern.
Neuroevolution
How does HyperNEAT use the CPPN?
HyperNEAT uses CPPN to define the weights of the ANN that executes the desired task
Neuroevolution
How is the CPPN trained?
CPPN is trained using NEAT
Neuroevolution
What are the activation functions in aa CPPN?
Pretty much any activation function, like sigmoid, cosine, linear, absolute value etc.
