Package ec.eval

Class Slave

java.lang.Object
ec.eval.Slave
public class Slave extends Object
Slave.java

Slave is the main entry point for a slave evaluation process. The slave works with a master process, receiving individuals from the master, evaluating them, and reporting the results back to the master, thus enabling distributed evolution.

Slave replicates most of the functionality of the ec.Evolve class, for example in terms of parameters and checkpointing. This is mostly because it needs to bootstrap and set up the EvolutionState in much the same way that ec.Evolve does. Additionally, depending on settings below, the Slave may act like a mini-evolver on the individuals it receives from the master.

Like ec.Evolve, Slave is run with like this:

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

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

Slaves need to know some things in order to run: the master's IP address and socket port number, whether to do compression, and whether or not to return individuals or just fitnesses. Unfortunately, Sun's CompressedInputStream/CompressedOutputStream is broken (it doesn't allow partial flushes, which is critical for doing compressed network streams). In order to do compression, you need to download the JZLIB library from the ECJ website or from http://www.jcraft.com/jzlib/ . ECJ will detect and use it automatically.

Slaves presently always run in single-threaded mode and receive their random number generator seed from the master. Thus they ignore any seed parameters given to them.

Slaves run in one of three modes:

  • "Regular" mode, which does a loop where it receives N individuals, evaluates them, and returns either the individuals or their new fitnesses.

  • "Regular Coevolutionary" mode, which does a loop where it receives N individuals to assess together in a single coevolutionary evaluation, evaluates them, and returns either the individuals or their new fitnesses (or only some fitnesses if only some are requested).

  • "Evolve" mode, which does a loop where it receives N individuals, evaluates them, and if there's some more time left, does a little evolution on those individuals as if they were a population, then when the time is up, the current individuals in the population are returned in lieu of the original individuals. In this second form, individuals MUST be returned, not fitnesses. This mode is not available if you're doing coevolution.

Parameters

eval.slave-name
String 		(the slave's name, only for debugging purposes. If not specified, the slave makes one up.)
eval.master.host
String 		(the IP Address of the master.)
eval.master.port
integer >= 1024 		(the socket port number of the master.)
eval.compression
bool = true or false (default) 		(should we use compressed streams in communicating with the master?)
eval.run-evolve
bool = true or false (default) 		(should we immediately evaluate the individuals and return them (or their fitnesses), or if we have extra time (defined by eval.runtime), should we do a little evolution on our individuals first?)
eval.runtime
integer > 0 		(if eval.run-evolve is true, how long (in milliseconds wall-clock time) should we allow the individuals to evolve?)
eval.return-inds
bool = true or false (default) 		(should we return whole individuals or (if false) just the fitnesses of the individuals? This must be TRUE if eval.run-evolve is true.)
eval.one-shot
bool = true (default) or false		 (Should the slave quit when the master quits, or loop continuously in the background processing new masters?)