The GP.Lab Genetic Programming Kernel

GP.Lab uses a proprietary Genetic Programming implementation written in C++. The representation of individuals is accomplished in a rather 'verbal' way with dedicated program and node classes.

Though systems like TinyGP demonstrate that it's perfectly feasible to implement tree-based GP systems with minimal amounts of code as well as minimal memory footprint at runtime using flattened (linear) representation for trees we prefer the OO approach: Common techniques for object-oriented code allow us to structure the kernel in a way that's easy to understand and to extend (e.g. to adapt to specific problems, add analytics/statistics, etc.).

The GP kernel implements the basic mechanics of the Genetic Programming system, i.e. representation of individual programs, algorithms for initialization and selection as well as the genetic operators.

At this point the three standard initialization mechanisms for new programs are supported:
  • Grow, 
  • Full and 
  • Ramped Half&Half.
It's probably worth noting that the system uses the actual number of instructions in a program (not just the tree depth) when applying size limits during program initialization and creation of new programs via the genetic operators.

The kernel also provides a range of common instructions:

Program Flow

  • IFGTZ:  If greater than zero (3 parameters: condition, if, else)
  • WHILEGTZ:  A Do-While-Greater-Than-Zero loop

Logic

  • AND:  Logical AND
  • OR:  Logical OR

Math

  • PLUS, MINUS, TIMES, DIV:  Elementary arithmetic (with protected division)
  • SIN, COS, TAN:  Trigonometric functions

Indexed Memory

  • MEMREAD:  Reads the indexed memory content at a specific index
  • MEMWRITE:  Writes to the indexed memory at a specific index
  • MEMSWAP:  Swaps the contents of two indexed memory locations

Specific tasks (problems) require the definition of the function and terminal set for the problem to be solved and subclassing of the program class. The latter contains the logic to initiate evaluation of individual programs and in particular the (entirely problem specific) calculation of fitness values.

Problem specific instructions can be added with a few lines of code by specifying the number of parameters (node fan-in or arity) and logic for evaluating the instruction.
Future posts will cover in more detail the genetic operators and selection mechanisms used by this implementation - both are currently either being refactored or rewritten from scratch as we dust off the project..


Share and Enjoy!

Comments

Post a Comment

Popular posts from this blog

GP.Lab: Dark Silicon Update

Image
It's been a while since my last update. However, it appears GP.Lab is coming out of hibernation these days.. For now there's a new preview to share featuring a moderately modernized app with support for macOS' new(ish) Dark Mode as well as native support for Apple Silicon Macs . Naturally Intel Macs are still supported, as are macOS versions equal to or newer than macOS 11 ("Big Sur"). The latest preview also features a new Sorting module for (well..) sorting an array of 10 integer numbers in ascending order. There's more work to be done on this one so please treat it kindly as an early preview. The Collector module has been improved and bugs impeding the sensory input of our feeble robot were removed - resulting in quite some improvement in its ability to actually collect items and deliver them to the home base. See it in action here: https://youtu.be/0WSjwM38i2Q And one of these days I've got to actually describe all of the modules (tasks) included
Read more

About GP.Lab

Hi there and welcome to the GP.Lab blog. So what's "GP.Lab"? It's a Genetic Programming Workbench and a work in progress. Besides being a fun, non-profit software research project its main goal is to investigate machine learning ( Genetic Programming in particular) and its applications to automatically solve a range of problems. GP.Lab implements a modular architecture: Individual tasks or problems are essentially plug-ins to the core app containing everything related to the specific problem to be solved such as Specific operators  (Robot steering commands, access to data sources) A visualization appropriate for the task Additional logic required to evaluate programs (e.g. a simulation environment) The core app implements basic functionality such as the core GP kernel algorithm, methods for managing the GP population, as well as a bunch of statistical tools to provide insight into the simulated evolutionary process. Initial Set of Tasks At this p
Read more

GP.Lab: Public Preview

Image
As I mentioned earlier GP.Lab is freeware - assuming that term still makes sense in today's diverse landscape of monetizing software products.. Anyhow: A preview release of GP.Lab is now available for download here . A Mac running macOS 10.11 or newer is required. The current implementation has modest memory requirements but you'll want to run it on a machine with plenty cores to speed up the evaluation of programs in the population. Quick Guide When the app is launched the Project Selector is presented. Select the task you would like the machine learning system to tackle. A brief description is provided for each task and additional details can be found in earlier blog posts. Customize Run Parameters such as program and population size, selection mode and genetic operator probabilities in the next window. The tournament size and elitism parameters should be handled with extra care: both result in quick loss of diversity in the simulated gene pool when chosen too
Read more