Package ec

Class Evolve

java.lang.Object
ec.Evolve
public class Evolve extends Object
Evolve is the main entry class for an evolutionary computation run.

An EC run is done with one of two argument formats:

java ec.Evolve -file parameter_file [-p parameter=value]*

This starts a new evolutionary run, using the parameter file parameter_file. The user can provide optional overriding parameters on the command-line with the -p option.

java ec.Evolve -checkpoint checkpoint_file

This starts up an evolutionary run from a previous checkpoint file.

The basic Evolve class has a main() loop with a simple job iteration facility. If you'd like to run the evolutionary system four times, each with a different random seed, you might do:

java ec.Evolve -file parameter_file -p jobs=4

Here, Evolve will run the first time with the random seed equal to whatever's specified in your file, then job#2 will be run with the seed + 1, job#3 with the seed + 2, and job#4 with the seed + 3. If you have multiple seeds, ECJ will try to make sure they're all different even across jobs by adding the job number * numberOfSeeds to each of them. This means that if you're doing multiple jobs with multiple seeds, you should probably set seed.0 to x, seed.1 to x+1, seed.2 to x+2, etc. for best results. It also works if seed.0 is x, seed.1 is y (a number much bigger than x), seed.2 is z (a number much bigger than y) etc. If you set seed.0=time etc. for multiple jobs, the values of each seed will be set to the current time that the job starts plus the job number * numberOfSeeds. As current time always goes up, this shouldn't be an issue. However it's theoretically possible that if you checkpoint and restart on another system with a clock set back in time, you could get the same seed in a later job.

main() has been designed to be modified. The comments for the Evolve.java file contain a lot discussion of how ECJ's main() bootstraps the EvolutionState object and runs it, plus a much simpler example of main() and explanations for how main() works.

Parameters

jobs
int >= 1 (default)		 (The number of jobs to iterate. The current job number (0...jobs-1) will be added to each seed UNLESS the seed is loaded from the system time. The job number also gets added as a prefix (if the number of jobs is more than 1)).
nostore
bool = true or false (default)		 (should the ec.util.Output facility not store announcements in memory?)
flush
bool = true or false (default)		 (should I flush all output as soon as it's printed (useful for debugging when an exception occurs))
evalthreads
int >= 1		 (the number of threads to spawn for evaluation)
breedthreads
int >= 1		 (the number of threads to spawn for breeding)
seed.n
int != 0, or string = time		 (the seed for random number generator #n. n should range from 0 to Max(evalthreads,breedthreads)-1. If value is time, then the seed is based on the system clock plus n.)
state
classname, inherits and != ec.EvolutionState		 (the EvolutionState object class)
print-accessed-params
bool = true or false (default)		 (at the end of a run, do we print out a list of all the parameters requested during the run?)
print-used-params
bool = true or false (default)		 (at the end of a run, do we print out a list of all the parameters actually used during the run?)
print-unaccessed-params
bool = true or false (default)		 (at the end of a run, do we print out a list of all the parameters NOT requested during the run?)
print-unused-params
bool = true or false (default)		 (at the end of a run, do we print out a list of all the parameters NOT actually used during the run?)
print-all-params
bool = true or false (default)		 (at the end of a run, do we print out a list of all the parameters stored in the parameter database?)

  • Field Details

  • Constructor Details

    • Evolve

      public Evolve()

  • Method Details

    • checkForHelp

      public static void checkForHelp(String[] args, String classname, boolean checkpoint)
      Optionally prints the help message.

    • possiblyRestoreFromCheckpoint

      public static EvolutionState possiblyRestoreFromCheckpoint(String[] args)
      Restores an EvolutionState from checkpoint if "-checkpoint FILENAME" is in the command-line arguments.

    • loadParameterDatabase

      public static ParameterDatabase loadParameterDatabase(String[] args)
      Loads a ParameterDatabase from checkpoint if "-params" is in the command-line arguments.

    • determineThreads

      public static int determineThreads(Output output, ParameterDatabase parameters, Parameter threadParameter)
      Loads the number of threads.

    • primeGenerator

      public static MersenneTwisterFast primeGenerator(MersenneTwisterFast generator)
      Primes the generator. Mersenne Twister seeds its first 624 numbers using a basic linear congruential generator; thereafter it uses the MersenneTwister algorithm to build new seeds. Those first 624 numbers are generally just fine, but to be extra safe, you can prime the generator by calling nextInt() on it some (N>1) * 624 times. This method does exactly that, presently with N=2.

    • determineSeed

      public static int determineSeed(Output output, ParameterDatabase parameters, Parameter seedParameter, long currentTime, int offset, boolean auto)
      Loads a random generator seed. First, the seed is loaded from the seedParameter. If the parameter is V_SEED_TIME, the seed is set to the currentTime value. Then the seed is incremented by the offset. This method is broken out of initialize(...) primarily to share code with ec.eval.MasterProblem.

    • buildOutput

      public static Output buildOutput()
      Constructs and sets up an Output object which always issues errors in the traditional fashion rather than throwing them.

    • buildOutput

      public static Output buildOutput(boolean throwsErrors)
      Constructs and sets up an Output object.

    • initialize

      public static EvolutionState initialize(ParameterDatabase parameters, int randomSeedOffset)
      Initializes an evolutionary run given the parameters and a random seed adjustment (added to each random seed). The adjustment offers a convenient way to change the seeds of the random number generators each time you do a new evolutionary run. You are of course welcome to replace the random number generators after initialize(...) but before startFresh(...)

      This method works by first setting up an Output (using buildOutput), then calling initialize(ParameterDatabase, seed, output)

    • initialize

      public static EvolutionState initialize(ParameterDatabase parameters, int randomSeedOffset, Output output)
      Initializes an evolutionary run given the parameters and a random seed adjustment (added to each random seed), with the Output pre-constructed. The adjustment offers a convenient way to change the seeds of the random number generators each time you do a new evolutionary run. You are of course welcome to replace the random number generators after initialize(...) but before startFresh(...)

    • cleanup

      public static void cleanup(EvolutionState state)
      Begins a fresh evolutionary run with a given state. The state should have been provided by initialize(...). The jobPrefix is added to the front of output and checkpoint filenames. If it's null, nothing is added to the front.

    • main

      public static void main(String[] args)
      Top-level evolutionary loop.