Introduction

Computational Intelligence uses Biologically inspired “algorithms” or techniques with focus on model output, instead of model workings:

We saw 4 different perspectives:

These and other perspective can be combined!

But why would we mix strategies?

• Alleviate a shortcoming of one CI using another
  • Convergence, e.g.,
    • a method does not converge quickly or
    • converges to a local optimum
  • State representation, e.g.,
    • states are continuous, or,
    • states share features
  • Fitness representation, e.g.,
    • fitness might be too locally defined, or,
    • fitness can not be locally defined.

These techniques can be combined in many different ways. In this lecture, we will see a few examples.

Artificial Neural Networks + Genetic Algorithms

There are several problems where backpropagation is not good enough to handle the complexity. In these cases, we can for example use a Genetic Algorithm:

Mutation: for each gene, add Gaussian noise to each weight.

We typically use Mutation, not crossover. This is because swapping weights is not as efficient as mutating the weights:

Cross-over is usually inefficient for neural network weights.

Evolutionary Neural Networks tries to find the optimal Network Structure: