10 - Neuroevolution

Question 1

Evolving Weights

Answer 1

Instead of using backprop, use evolution to find weights

Mutation is the main operator to search new solutions

Question 2

Advantages of neuroevolution vs backprop

Answer 2

• Evolving weights can be applied to any network arch
• … can be applied to any transfer function

Question 3

Evolving architectures (in terms of model layers)

Answer 3

Evolution can be used to find the optimal model layers etc

Question 4

3 features of NeuroEvolution of Augmenting Topologies

Answer 4

• Tracking genes with historical markers to make crossover less destructive
• Apply speciation. New topologies are like a new species and protected during evolution
• starts small and add

Question 5

Evolving learning rules

Answer 5

• Problem/arch dependent

evolve the parameter (eg learning rate)
evolve the function
decide the inputs of the rule
decide the function type
tune the parameters of the rule

Question 6

Direct Representation

Answer 6

ABCABCABCABCABCABC

Question 7

Indirect Representation

Answer 7

ABCx6
Closer to biological inspiration
Compressed

Question 8

Disadvantages of direct rep

Answer 8

in ABCABCABCABCABCABC

changing all As would be costly. in indirect it is not

Question 9

Disadvantage of indirect rep

Answer 9

If there is a lot of uniqueness in the pattern it’s difficult

Or if you wabt to change just one A in ABCx6 then it is hard

Question 10

Reuse and Rediscovery

Answer 10

Reuse is reusing structures in different places but there can be variation

Question 11

Grammatical Developmental Encodings

Answer 11

• Uses properties of grammars and computer languages
  subroutines and hierarchies

Question 12

Cell Chemistry dev encodings

Answer 12

• simulate low level chemical and biological properties
• simulate diffusion of chemicals, reaction, production, signalling etc

Question 13

Grammatical: L-systems

Answer 13

Good for fractal like structures.

Using substitutions of A B etc

EXAMPLE ONLY:
Variables : A B
axiom: A
rules (A->AB), (B-> A)

n = 0 : A
n = 1: AB
n = 2: ABA
n = 3: ABAAB

Question 14

CPPN Compositional Pattern Producing Networks

Answer 14

Artificial indirect encoding to abstract how embryos are encoded through DNA

Question 15

HyperNEAT

Answer 15

• Uses CPPN Compositional Pattern Producing Networks
• Uses an indirect encoding approach
• Can evolve network with millions of connections

1. Query each potential connection on substrate
2. Feed each coordinate pair into CPPN
3. Set weight of connection between pair to value of output

n by n neurons means (n * n) ^2 connections

Question 16

HyperNEAT more neurons has what effect on genome

Answer 16

None.

The genome is the CPPN so it does not increase

Question 17

Adaptive HyperNEAT

Answer 17

In addition to inputting coord pairs we also input a pre and post and w,

Pre and post synaptic values (how active the network is)

Question 18

Generative development systems are good when

Answer 18

• problem is likely to benefit from reuse
• likely to be regularities in solution
• we don’t know how the solution will look

Question 19

Generative development systems are NOT good when

Answer 19

• problem/solution has no regularity
• small search spaces
• simple control tasks, optimsation problems