Artificial Intelligence Systems

Objective

Implement a Multilayer Neural Network using supervized learning to approximate/simulate world terrains that look real.

Open Octave
Load multilayer.m file.
Run study_terrain(terrain_path, epochs, n_samples) where:
- terrain_path is the relative path to a terrain data file.
- epochs is the number of epochs to run.
- n_samples is the amounts of data points to use from the data file.

Alternatively, you may run study_terrain_custom which offers richer configuration arguments:

study_terrain_custom(terrain_file, epochs, n_samples, learn_function, eta, momentum) where:
- terrain_file, epochs, n_samples are the same as above.
- learn_function must be either "incremental" or "batch" and is used to set the learning function.
- eta is the learning rate.
- momentum is the momentum factor.

multilayer.m contains the functions necessary to make the neural network.
gen_arq.m generates the eleven architectures we experimented with.
study_terrain uses the already pre-set best parameters to learn a new terrain.
test_results.data contains the results of running the 11th architecture with different epochs and sample sizes.
initial_W.data contains the initial weights matrix utilized in the tests.
/terrains contains all the terrains provided.
/images contains the different graphs we gathered, such as different error rates for different architectures and many other things