Crate eant2 [−] [src]
An implementation of the EANT2 algorithm used for training neural networks. The algorithm requires very little information about the problem, and creates minimal, high fitness networks.
While this algorithm is extremely general and can learn almost any task, it can take a very long time to run. Because of this, it is sometimes necessary to adjust the options to make it run faster. It is important to design a good task, otherwise the algorithm can take days to run. Here are some tips for creating a good one:
Lower the amount of inputs and outputs to each neural network. Design the fitness function to either do some of the work, or just simplify the problem. Another option is to train a separate neural network to do processing of raw inputs, then feed its simplified output to the EANT2 network.
Run the algorithm on good hardware to take advantage of the multithreaded support. The algorithm is a one time thing; create a neural network with it and distribute and run it on less powerful computers. On a related note, it is okay if it takes some time to run because it will only have to run once.
Adjust individual options for the algorithm through
EANT2Options
. The documentation provides a lot of information on how each option effects the algorithm.
Note that it is important to get everything right the first time, otherwise the results might be below expectation, which can mean running the algorithm again.
Examples
Complete this section when the project is finished
Reexports
pub use self::fitness::NNFitnessFunction; |
pub use self::options::EANT2Options; |
pub use self::eant::eant_loop; |
Modules
eant | |
fitness |
The fitness function. |
options |
Option types for the EANT2 algorithm. |
Structs
Network |
Enums
CMAESEndConditions |
An enum representing a condition under which to terminate the CMA-ES algorithm. |
TransferFunction |
Represents which transfer function to use for evaluating neural networks. |
Functions
foo |