v2.2.4 (Efficacy Sampling - Last Minute Changes)

This is the version of the software used to run the efficacy sampling experiments. Previously this was reported to be version 2.2.3, but there were in fact last minute changes made; this release thus captures the state of the software used to generate the data in that efficacy sampling post.

Changes include a code refactor of the generative function task, and tweaks to the generative function task fitness evaluation scheme.

v2.2.3 (Efficacy Sampling)

This version is primarily a marker for the state of the software that efficacy sampling experiments have been performed against. Future changes can be tested with efficacy sampling on the same tasks in order to determine if those changes improve the efficacy of SharpNEAT, or otherwise.

Other changes:

• Function regression task: Implementation changes. Evaluation scheme changes.
• New Generative Sinewave task.

v2.2.2

SharpNEAT 2.2.2
2017-01-28
Colin Green

Changes from previous release

Fixes

• Updated/fixed console app config file.
• Genome sorting now uses an unstable sort algorithm that randomly distributes the position of genomes of equal fitness. This removes the previous arbitrary selection of the top N genomes for elitism and selection. This might improve fostering and maintenance of genetic diversity.
• Mutate_AddConnection() for feedforward networks did not evenly distribute the selection of random source and target neurons. This could result in some neurons having zero probability of being selected.

Enhancements

• Use near-zero weight when adding new connections
• Distribution plots changed to histograms (they always were histograms but plotted as lines instead of bars
• Modifications to the Walker2D problem domain in an attempt to improve efficacy.
  -Change to make 50% of new connections very close to zero; in the spirit of introducing connections with minimal impact which are therefore more likely to be accepted.
  -Adjusted default mutation type probabilities. Simplified the set of connection mutation types - mass weight resets are likely not useful, nor mass weight deltas.

Miscellany

• Updated projects to target .Net framework 4.6.1
• Updated log4net nuget references from version 2.0.5 to 2.0.7.
• Some general purpose utility classes are now referred to in their new home, the Redzen project (available as a nuget package).
• Upgraded sharpneat.sln (solution file) to Visual Studio 2015 format.
• Zedgraph dependency now obtained as a nuget package.
• log4net dependency now obtained as a nuget package.
• Housekeeping. Old/dead code removal.
• Clarified licensing info. Removed last trace of GNU license.
• Box2DX dependency is now resolved using nuget.
• Removed unused Tao assemblies for font rendering. N.B. These are 32 bit assemblies therefore cannot be loaded into a 64bit process.

v2.2.1

SharpNEAT 2.2.1
2015-06-15
Colin Green

Changes from previous release

Fixes

• Performance: ZigguratGaussianSampler.cs: Switched SampleTail() to use NextDoubleNonZero() instead of NextDouble(),
  to avoid attempt to compute Log(0). That wasn't a defect per se because Log() returns a NaN instead of throwing an
  exception, but it did cause a slow execution path.
• FIX: FastAcyclicNetworkFactory.cs: Lookup of definition node index from a new/working index was wrong. The lookup
  table used worked the other way around, i.e. it mapped definition indexes to new indexes. At this time this
  affected only the Walker 2D (Hyperneat) problem domain.
• FIX: Box2dDomainView.Designer.cs: Added openGlControl.DestroyContexts() based on report of a blue screen failure
  on one box being remedied by this (almost certainly a video driver issue).
• FIX: WalkerBox2dExperimentHyperNeat.cs: Nasty bug in setting of substrate node coordinates.
• FIX: FastRandom.cs: Seeds are now hashed. Without this the first random sample for nearby seeds (1,2,3, etc.) are
  very similar (have a similar bit pattern). Thanks to Francois Guibert for identifying this problem.

v2.2.0

SharpNEAT 2.2.0
2012-04-02
Colin Green

Changes from previous release

New Features

• Walker BOX2D Problem Domain.
• Support for multiple auxiliary fitness values per genome (plotted on graphs).
• Acyclic networks as HyperNEAT CPPNs.

Fixes

• FIX: Prey Capture problem domain: Only one of the four ANN output signals was being read.
  This effectively completely broke the prey capture domain.
• IntPoint: Fixed equality and inequality operators and CalculateDistance().
  In the released code these defects severely affected the prey capture domain.
• FIX: Config loading: Relaxing network delta threshold setting was being parsed as an Int32 instead of a Double.
• Fix to RandomClusteringStrategy.cs. Genome.SpecieIdx was not being set upon allocation.
  Added debug assertion to check specieIdx is correctly set following speciation.
• FIX/MOD: Network visualisation: Layout logic failed when there were large numbers of neurons in
  a layout layer such that the gap between them was less than 0.5 of a pixel, and thus got rounded down to 0,
  positioning all nodes in that layer at the same coordinates.
• FIX/MOD: Ensure ID generators are set accordingly when loading genomes using a pre-existing genome factory.

v2.1.0

SharpNEAT 2.1
2011-09-16
Colin Green

Changes from previous release

Major New Features

• Support for evolution of acyclic networks. This includes:
  • Development of NeatGenome and associated classes to support efficient
    evolution of acyclic networks.
  • Development of neural network classes to efficiently execute neural
    networks.
  • Development of network visualization to layout nodes by layer. This also
    affects how cylcic networks are laid out because the algorithm that
    determines which layer a node is in for acylcic nets was extended to also
    calculate a sensible layer number for cyclic networks, This greatly
    improves how the networks are laid out with respect to gaining
    understanding of the network architecture from its visual rendering.
  • Modified all function regression and binary logic themed experiments to
    use acyclic networks.
• Added support for Box2D (2D physics engine) based problem domains and
  visualisation of Box2D worlds with OpenGL.
  • New experiment - Single pole balancing using Box2D, including
    visualization.
  • New experiment - Inverted double pendulum using Box2D, including
    visualization.

Other Developments

• Improved assertions when running in debug mode. This improves code quailty
  by reducing the number of undetected defects in released code.
• New ZigguratGaussianSampler class for generating Gaussian noise for
  mutations and simulations (where required). This approach has a much reduced
  requirement for calls to expensive floating point operations such as
  Math.Sqr() and Math.Log(). Typically this is about 2x faster than sampling
  by using the previously used and simpler Box-Muller method.
• FastRandom now seeds using random numbers from a global FastRandom. This
  prevents multiple instances from obtaining the same seed from the system
  tick count when initialising within the same clock tick.
• Log(n) function regression experiment.
• XOR and binary multiplexer experiments modified to use fitness score based
  on squared error. This change improves search efficiency.
• Function regression experiments changed to have peaks and troughs at y = 0.9
  and 0.1 respectively. This avoids requiring activation functions to output
  values at the extremes of their ranges.
• CPPNs modified to use a Gaussian activation function instead of
  BipolarGaussian. My opinion is that BipolarGaussian isn't directly useful
  and the equivalent functionality can be achieved if necessary by combining
  Gaussian with Linear.
• Refactoring of mutation type selection logic/code.

Fixes

• Fix to crossover logic whereby connection genes where not copied and thus
  shared between parent and child genomes.
• RelaxingCyclicNetwork and FastRelaxingCyclicNetwork: IsStateValid property
  return value was defined the wrong way around for relaxed networks. It
  returned false when the networks were relaxed (which is the valid state).
  Note - none of the current experiments shipped with SharpNEAT use relaxing
  networks.
• XML I/O was not culture neutral. RBF-NEAT uses comma separated numbers
  within genome XML which conflicted with use of commas as the numeric decimal
  separator in come cultures.
• Fix to genome loading. Now uses genome factory from the current experiment;
  previously it was hard coded to NeatGenomeFactory which was incorrect when
  using sub-classes such as CppnGenomeFactory.
• Genetic crossover of CPPN genomes randomly regenerated the node activation
  functions on each node of a child genome instead of taking the activation
  functions from the parent genome.