SNPs_and_clones

Goal

Implement a Genetic Algorithm to study SNP haplotypes.

Grand Master Plan

• Better population diversity metrics --> (RGB colors with softmax-ed values of first three PCA axis?)
• Implement (again) min_mut in evolve.R?

Scripts

meta_launcher.sh

Status: Functional Script to launch numerous generations, acts as a wrapper for

To launch, write the following command: ./meta_launcher.sh $1 $2 $3 $4 $5

Where: config_file=$1 # Path to config file, absolute or relative to launching dir. max_iter=$2 # Maximum number of generations per GA script launch. total_iter=$3 # Total number of generations to run. save=$4 # 0,1 or 2, determines what outputs are saved afterwards verbose=$5 # 0, 1 or 2, determines how much information is printed in console

gen_algo.R

Status: Functional (de facto main script)

load_inputs.R

Status: Missing weights

To Do:

• Implement weights
• Specify input files characteristics

quality_control.R

Status: To be done

To Do:

• Implement covariates QC

generate_G0.R

Status: Functional

This function is used to generate a starting generation, and if n_nov > 0 to insert random genomes at each generation before crossing-overs.

calc_score.R

Status: Functional with rpart tree

GLM mock function for testing purposes. Rpart tree accuracy score functional.

evolve.R

Status: Functional

To Do:

• Implement unique elites to ensure diversity of elites.

Mutation mechanisms cull feature inflation.

Indeed, mutations occurr symmetrically (similar counts of feature gain / loss) until a threshold is met. This unsures run time is constant and predictable between Gen 0 and Gen n_iter.

Also, results are ranked in order of minimal score first, features count second. This ensures that in the case where a plateau is reached by the optimal solution, unused features (i.e. not in the decision tree) are slowly purged from the elites by random loss mutations.

Crossing-over mechanisms induce the introduction of diversity in genomes.

Indeed, a rare event where a genome is shuffled fully was implemented to lower the probability of the whole generation getting stuck in the neighbourhood of a local optimum.

Right now, this event happens with a probability of 0.01 times the crossing-over rate (cr).

Known issues

Memory fills up with unattainable R objects:

• Implemented quickfix: Dump and source genomes, outputs & parameters (meta_launcher.sh)