Psh

Psh is a Java implementation of the Push programming language and of PushGP. Push is a stack-based language designed for evolutionary computation, specifically genetic programming. PushGP is a genetic programming system that evolves programs in Push. More information about Push and PushGP can be found here.

This is v1.0 of Psh.

Getting Started with Psh

Building Psh

Use the next Maven command to build the package:

$ mvn clean compile assembly:single

The resulting jar file will be in the target folder.

Using PshGP

To run PshGP on a sample problem:

$ java -jar psh.jar gpsamples/intreg1.pushgp

This problem uses integer symbolic regression to solve the equation y = 12x^2 + 5. Other sample problems are available, with descriptions, in gpsamples/. For example, intreg2.pushgp uses integer symbolic regression to solve the factorial function, and regression1.pushgp uses float symbolic regression to solve y = 12x^2 + 5.

Using PshInspector

PshInspector allows you to examine every step of a Psh program as it executes. To run PshInspector on a sample psh program:

$ java -jar psh.jar pushsamples/exampleProgram1.push

This push file runs the psh program (2994 5 integer.+) for 100 steps after pushing the inputs 44, 22, true, 17.76. Other sample psh programs are available in pushsamples/.

Psh In More Detail

Configuration Files

PshGP runs are setup using configuration files which have the extension .pushgp. These files contain a list of parameters in the form of

param-name = value

The following parameters must be defined in the configuration file, given with example values:

problem-class = org.spiderland.Psh.IntSymbolicRegression

max-generations = 200
population-size = 1000
execution-limit = 150
max-points-in-program = 100
max-random-code-size = 40

tournament-size = 7
mutation-percent = 30
crossover-percent = 55
simplification-percent = 5

reproduction-simplifications = 25
report-simplifications = 100
final-simplifications = 1000

test-cases = ((1 1) (2 2) (3 6) (4 24) (5 120) (6 720))
instruction-set = (registered.exec registered.boolean integer.% integer.* integer.+ integer.- integer./ integer.dup)

The following parameters are optional. If not specified, the default values below will be used for these parameters, except for the parameters mutation-mode, output-file, and push-frame-mode, which significantly change the run when specified. Also, target-function-string defaults to not displaying a string, but a representative example is given below.

trivial-geography-radius = 10
simplify-flatten-percent = 20
mutation-mode = fair
fair-mutation-range = .3

node-selection-mode = unbiased  (others available are leaf-probability and size-tournament)
node-selection-leaf-probability = 10  (only used if node-selection-mode = leaf-probability)
node-selection-tournament-size = 2  (only used if node-selection-mode = size-tournament)

min-random-integer = -10
max-random-integer = 10
random-integer-resolution = 1
min-random-float = -10.0
max-random-float = 10.0
random-float-resolution = 0.01

target-function-string = "y = x^4 - 2x + 7"

interpreter-class = org.spiderland.Psh.Interpreter
individual-class = org.spiderland.Psh.PushGPIndividual
inputpusher-class = org.spiderland.Psh.InputPusher

output-file = out.txt
push-frame-mode = pushstacks

PshInspector Files

In order to inspect the execution of a program, PshInspector takes a push program file with the extension .push. After every step of the program, the stacks of the interpreter are displayed. The input file contains the following, separated by new lines:

• Program: The Psh program to run
• ExecutionLimit: Maximum execution steps
• Input(optional): Any inputs to be pushed before execution, separated by spaces. The inputs are pushed in the order in which they are given. Note: Only int, float, and boolean inputs are accepted.

Problem Classes

PshGP uses problem classes, implemented as Java classes, to determine certain aspects of the run, such as how to compute fitness values. The choice of problem class determines how test case data is interpreted, and which stacks are used for test case input and output. In addition, certain inherited methods in both GA.java and PushGP.java may be overwritten for further customization.

Psh comes with a few standard problem classes. The following problem classes are currently implemented, and are in the ProbClass subpackage:

• FloatSymbolicRegression.java: Maps an input floating point value to an output floating point value. Error value is computed as the difference between the desired output value and the top value on the float stack.
• IntSymbolicRegression.java: Maps an input integer value to an output integer value. Error value is computed as the difference between the desired output value and the top value on the integer stack.
• CartCentering.java: Maps two input floats (position and velocity) to a boolean value that represents a forward or backward force applied to a cart. The error is the amount of time required to stop the cart at the origin. For more information, see the problem class file.

In order to perform runs for other types of problems, you can implement your own custom problem classes. Please note the following:

• You will likely want to implement the InitFromParamenters method, which can be used to set up test cases. If so, make sure to also call its parent method.
• In PshGP, the term fitness actually refers to error values, which means that lower values are considered more fit and that 0.0 represents no error. The EvaluateTestCase method must be implemented by any problem class, and should compute an individual's fitness, with lower values being better.
• The InitInterpreter method must be implemented by all problem classes though many times this method is simply left empty.
• There are other optional methods that can be overwritten or extended in the GA.java and PushGP.java classes. For example, the CartCentering.java problem class implements the Success method in order to override the conditions that GA uses to identify a successful run.

Changelog

Major Changes since v1.0:

• The parameters that affect Ephemeral Random Constant creation, such as the minimum random integer, are now available as optional configuration parameters. See Configuration Files above for more details.
• Implement new instructions: integer.pow, integer.min, integer.max, float.exp, float.pow. Also, fixed a bug in float.max.
• Moved problem classes and test cases to their own packages to reduce clutter.
• Fixed holes in many integer and float instructions that could cause underflow, overflow, or NaN errors.
• Made FloatSymbolicRegression and IntSymbolicRegression as well as co-evolved FloatSymbolicRegression work with test-case generators.
• PshGP now primarily uses the mean of test case errors for an individual's error instead of the total sum of the errors.
• Added optional parameter target-function-string, which specifies a human-readable version of the target function, which is only used in I/O.
• Added many instructions that were missing from the Push 3.0 specification.
• Added node-selection-mode as an optional parameter, as well as node size tournaments for node selection.

Major Changes since v0.3:

• Added new integer and float instructions: abs, neg, sin, cos, tan, max, min.
• Added new boolean instructions: and, or, xor, not.
• Added problem class for the cart centering problem (CartCentering.java), an optimal control problem.
• Made many parameters of .pushgp files optional. This should make creating .pushgp files for new users much simpler, as many parameters are rarely (if ever) changed. Optional parameters are listed in this readme.
• Change Psh over to Apache 2.0 license.
• The number of fitness evaluations is now displayed during reports.

Major Changes since v0.2:

• All instructions have been converted into lower case to match Schush and other implementations.
• An input stack was added, which holds all inputs. It has the following instructions:
  1. input.index - Pops n off of the integer stack and pushes input[n] onto corresponding stack. If integer stack is empty, acts as a no-op.
  2. input.makeinputsN - Creates N instructions called 'input.in0', 'input.in1', ..., 'input.in(N-1)'
  3. input.inall - For all n in 0 to input.size, push input[n] onto the corresponding stack.
  4. input.inallrev - For all n in input.size to 9, push input[n] onto the corresponding stack.
  5. input.stackdepth - Puts size of stack on integer stack.
• In config files, you can now include all instructions for a certain type using 'registered.type' (e.g. 'registered.integer' or 'registered.stack').
• Implemented auto-simplification, which is used during generation and final reports. Auto-simplification may also be used as a genetic operator along with mutation and crossover.

Major Changes since v0.1:

• Added problem classes for integer symbolic regression (IntSymbolicRegression.java) and integer symbolic regression without an input instruction (IntSymbolicRegressionNoInput.java).
• Fixed 'code' and 'exec' stack iteration functions, which were not executing correctly according to Push 3.0 standards.
• PshGP now displays the error values for the best program during the generation report.
• PshInspector was created to inspect interpreter stacks of push programs as they execute. This can be used to catch errors and trace executions. To run, see Using PshInspector section above.

Acknowledgement

This material is based upon work supported by the National Science Foundation under Grant No. 1017817. Any opinions, findings, and conclusions or recommendations expressed in this publication are those of the authors and do not necessarily reflect the views of the National Science Foundation.