# This was a plan for a "case study" file. The actual contents are
# just unimplemented ideas as of yet!! Actual grammar to be used is
# JSON, as in our earlier Hadoop experiments. It is just so easy to
# parse using boost. So, these remain only as an idea roadmap:

# Sections are marked up as in INI-files. Fields are not, though!
[Case]

# Field names are to end with ':' and be separated from value by whitespace.
# One line for each field. One value for each name.
# Any white space within the value is interpreted as a single ' '.

# The 'tag:' field will identify the case study for LaTeX documentation etc.
tag:   sample1

# The 'brief:' field should contain a brief description of the case.
brief: A first example

# The 'type:' field determines the learning task. As of now, only
# 'classification' is supported.
case_type: classification

[Dataset]
file: iris.train
# generate: 2dvshape  -- maybe also like this?

# Parameters for population size
[Population]
pop_size: 100

# Constraints (maximum layer sizes) on MLP architecture.
[MLP_constraints]
# Input and output sizes should be set according to data dimensions ("*"): 
max_size: *-3-4-*
# Except that we probably want to do feature extraction somehow:
#max_size: fixed-3-4-*
#max_size: featuresel-3-4-*

# How to initialize the MLPs of the first population.
[MLP_init]
init_size: *-3-4-*
init_weights: uniform(-1,1)

# Objective functions, aggregation, and algorithm(s) used in
# local improvement
[MLP_improvement]
improvement_obj: mse
improvement_alg: backprop/.8 prunesmall/.2
improvement_steps: 10
improvement_regc: adaptive

# only ideas as of now: 'classwise_random_mse'

# Objective functions used in global Pareto search
[MLP_evolution]
evolution_obj: classwise_misprediction number_neurons
evolution_alg: weight_jog/adaptive weight_del/.2