Genetic Algorithm Generalization
Finding the best neural network structure for testing and training
Abstract
We're using a genetic algorithm to train neural networks to deduce what type of neural networks can generalize better. It is current theory that the simplest model that fits a problem is the best model, but we are predicting that a bit more complexity than the simplest model will increase its generalization. Our environment is a simple map traversal that mimics race tracks and mazes; The individuals are given input corresponding to speed, direction, and distance to walls; the output includes doing nothing, speeding up, slowing down, turning left, and turning right. We will compare the results of a neural network training vs testing on a map.
Problem Definition
Figure out how the structure of neural networks relate to the performance during training vs testing
Determine a way to predict test performance
Existing Solutions
Comparing various NN in a specific environment
Generalization is measured using the bias and variance of performance between environments
Limitations
Can be slow and complex
Not guaranteed to find the best solution
Difficult to Optimize
GAs tend to overfit NNs
Expected Impact
Determine correlations between model structure and generalization ability
Invent a way to estimate the ability of models without training