A problem set I wrote for APPPHYS 237: Quantitative evolutionary dynamics
Some background a student might need: basic understanding of neural nets, probability/counting, and content from the course such as branching process and barcode lineages. Problem 1 doesn't require any bio background and can be tackled by anyone with knowledge of basic mathematics.
feedback much appreciated!
Evolution is nature's best optimization scheme. Can evolution be used to optimize artificial systems?
Neuroevolution uses evolutionary algorithms to generate neural networks to solve complex tasks.
NEAT proposed a method of evolving neural network topologies along with weights, resulting in faster convergence and greater success rate on the double pole balance RL task. The following key insights:
- homologous recombination for crossover of different topologies
- protecting structural innovation using speciation, and
- "complexifying", or evolving from a simple, homogeneous initial population.
In each problem, we focus on a specific innovation from NEAT. We apply evolutionary dynamics to provide insights on neural network evolution.