Genetic Network Programming (GNP)

G for Genetic

• Inspired by the mechanisms of evolutionary biology (Darwin)
Basic principle of biological evolution: “Beneficial traits resulting from random variation are favored by natural selection.”
• Application of selection, reproduction, and mutation
• Goal: Find optimal solutions within the defined search space

N for Network

• Network structure leads to an implicit memory function
• Ideal for dynamic environments and time series analyses
• The networks are transparent and not a "black box model"

N for Network

P for Programming

• Developed and researched since 2000
• Focus: dynamic environments (Time series)
• Let's run a GNP...

Let's run a GNP...

Let's run a GNP...

Variable-size genetic network programming

Challenges

Challenges

• Deleting to many nodes

Challenges

• Deleting to many nodes
• Deleting "important" nodes

Challenges

• Deleting to many nodes
• Deleting "important" nodes
• Adding to many nodes (bloat problem)

Challenges

• Deleting to many nodes
• Deleting "important" nodes
• Adding to many nodes (bloat problem)
• Overfitting

Challenges

• Deleting to many nodes
• Deleting "important" nodes
• Adding to many nodes (bloat problem)
• Overfitting
• Faulty crossover

The novel Mutation Operator

• Novel operator is fitness neutral.

The novel Mutation Operator

• Novel operator is fitness neutral.
• Only delete unused nodes.

The novel Mutation Operator

• Novel operator is fitness neutral.
• Only delete unused nodes.
• Only add nodes if all existing nodes are used.

Simulation Results

Network Example

Node Numbers

Future Research

• Mutation of start node function
• Adding multiple nodes
• Shifting of interval boundaries
• Multicriteria optimization
• Extension with Reinforcement Learning
• Crossover adaption : Exchanging indices

Appendix

Faulty Crossover

Mutation

Algo