-
Notifications
You must be signed in to change notification settings - Fork 93
Portal: Arguments
Haoxi Zhan edited this page Dec 10, 2013
·
12 revisions
Arguments are defined in clojush/pushgp/pushgp.clj.
:use-single-thread false ;; When true, Clojush will only use a single thread
:random-seed (random/generate-mersennetwister-seed) ;; The seed for the random number generator
:initial-population nil ;; (MAY BE BROKEN) Can point to a file where an initial population is stored and can be read
:save-initial-population false ;; When true, saves the initial population:error-function (fn [p] '(0)) ;; Function that takes a program and returns a list of errors
:error-threshold 0 ;; Pushgp will stop and return the best program if its total error is <= the error-threshold
:atom-generators (concat @registered-instructions ;; The instructions that pushgp will use in random code
(list
(fn [] (lrand-int 100))
(fn [] (lrand))))
:population-size 1000
:max-generations 1001
:max-points 50 ;; Maximum size of push programs, as counted by points in the program
:max-points-in-initial-program 50 ;; Maximum size of initial programs in generation 0
:evalpush-limit 150 ;; The number of Push instructions that can be evaluated before stopping evaluation
:evalpush-time-limit 0 ;; The time in nanoseconds that a program can evaluate before stopping, 0 means no time limit
:reuse-errors true ;; When true, children produced through direct reproduction will not be re-evaluated but will have the error vector of their parent:reproduction-probability 0.1 ;; Direct reproduction, which makes a direct copy of the parent
:mutation-probability 0.4 ;; Subtree mutation, similar to that found in tree-GP
:crossover-probability 0.5 ;; Subtree crossover, similar to that found in tree-GP
:simplification-probability 0.0 ;; Auto-simplification of the program
:ultra-probability 0.0 ;; Uniform Linear Transformation with Repair and Alternation -- a uniform crossover and mutation operator
:gaussian-mutation-probability 0.0 ;; Gaussian mutation affects only the float literals in a program by adding Gaussian noise
:boolean-gsxover-probability 0.0 ;; Geometric Semantic Crossover -- creates random code that determines which of the two parents to use for each test cases
:deletion-mutation-probability 0.0 ;; A mutation operator that deletes random instructions
:parentheses-addition-mutation-probability 0.0 ;; Operator that randomly inserts a parentheses pair somewhere in the program
:tagging-mutation-probability 0.0 ;; Operator that chooses a piece of code, moves it to beginning of program and tags it, and puts a tagged call in its place
:tag-branch-mutation-probability 0.0 ;; Operator that inserts a tag-branch into the program. tag-branches compare two items on a stack and then jump to one of two tags depending on the comparison:mutation-max-points 20 ;; The maximum number of points that new code will introduce during mutation
:reproduction-simplifications 1 ;; The number of simplification steps that will happen during simplification reproduction
:ultra-alternation-rate 0.1 ;; When using ULTRA, how often ULTRA alternates between the parents
:ultra-alignment-deviation 1 ;; When using ULTRA, the standard deviation of how far alternation may jump between indices when switching between parents
:ultra-mutation-rate 0.1 ;; The probability of each token being mutated during ULTRA
:use-ultra-no-paren-mutation false ; If true, ULTRA will use no-paren mutation, which means that parentheses won't be added or deleted during mutation.
:gaussian-mutation-per-number-mutation-probability 0.5 ;; The probability that any given float literal will be affected by a pass of gaussian-mutate
:gaussian-mutation-standard-deviation 0.1 ;; The standard deviation of a gaussian-mutated float
:boolean-gsxover-new-code-max-points 20 ;; The maximum size of the random code fragment used in boolean-gsxover
:tag-branch-mutation-type-instruction-pairs [] ;; A list of types and comparators that can be used by tag-branch-insertion-mutation:node-selection-method :unbiased ;; The node selection method can be :unbiased, :leaf-probability, or :size-tournament
:node-selection-leaf-probability 0.1 ;; If the node-selection-method is :leaf-probability, this is the percent of selections that will happen in leaves of the tree
:node-selection-tournament-size 2 ;; If node-selection-method is :size-tournament, this is the size of the node selection tournaments:tournament-size 7 ;; If using tournament selection, the size of the tournaments
:trivial-geography-radius 0 ;; If non-zero, this is used as the radius from which to select individuals for tournament or lexicase selection
:decimation-ratio 1 ;; If >= 1, does nothing. Otherwise, is the percent of the population size that is retained before breeding. If 0 < decimation-ratio < 1, decimation tournaments will be used to reduce the population to size (* population-size decimation-ratio) before breeding.
:decimation-tournament-size 2 ;; Size of the decimation tournaments
:use-historically-assessed-hardness false ;; When true, total error for tournament selection will depend on historically-assessed hardness
:use-rmse false ;; When true, total error for tournament selection will depend on the root mean square error of the error vector
:use-lexicase-selection false ;; If true, uses Lexicase Parent Selection (see Spector paper in GECCO-UP 2012 workshop proceedings)
:use-elitegroup-lexicase-selection false ;; If true, uses elitegroup lexicase selection, an experimental change to lexicase that thus far is often worse:tournament-size 7 ;; If using tournament selection, the size of the tournaments
:trivial-geography-radius 0 ;; If non-zero, this is used as the radius from which to select individuals for tournament or lexicase selection
:decimation-ratio 1 ;; If >= 1, does nothing. Otherwise, is the percent of the population size that is retained before breeding. If 0 < decimation-ratio < 1, decimation tournaments will be used to reduce the population to size (* population-size decimation-ratio) before breeding.
:decimation-tournament-size 2 ;; Size of the decimation tournaments
:use-historically-assessed-hardness false ;; When true, total error for tournament selection will depend on historically-assessed hardness
:use-rmse false ;; When true, total error for tournament selection will depend on the root mean square error of the error vector
:use-lexicase-selection false ;; If true, uses Lexicase Parent Selection (see Spector paper in GECCO-UP 2012 workshop proceedings)
:use-elitegroup-lexicase-selection false ;; If true, uses elitegroup lexicase selection, an experimental change to lexicase that thus far is often worse:report-simplifications 100 ;; The number of simplification steps that will happen during report simplifications
:final-report-simplifications 1000 ;; The number of simplification steps that will happen during final report simplifications
:problem-specific-report default-problem-specific-report ;; A function can be called to provide a problem-specific report, which happens after the normal generational report is printed
:print-errors true ;; When true, prints the error vector of the best individual
:print-history false ;; When true, prints the history of the best individual's ancestors' total errors
:print-timings false ; If true, report prints how long different parts of evolution have taken during the current run.
:print-cosmos-data false ; If true, report prints COSMOS data each generation.
:print-ancestors-of-solution false ; If true, final report prints the ancestors of the solution
:maintain-ancestors false ; If true, save all ancestors in each individual (costly):print-csv-logs false ;; Prints a CSV log of the population each generation
:print-json-logs false ;; Prints a JSON log of the population each generation
:csv-log-filename "log.csv" ;; The file to print CSV log to
:json-log-filename "log.json" ;; The file to print JSON log to
:log-fitnesses-for-all-cases false ;; If true, the CSV and JSON logs will include the fitnesses of each individual on every test case
:json-log-program-strings false ;; If true, JSON logs will include program strings for each individual:parent-reversion-probability 0.0 ;; The probability of a child being reverted to its parent if the parent has better fitness or equal fitness and is smaller
:use-bushy-code false ;; When true, random code and code changed by ULTRA mutation, will be "bushy", as in close to a binary tree