0.4.0

Hello everyone!

This is a time for a 0.4.0 release of GOLEM coming with several major features and fixes.

First of all, the new ecosystem for setting up experiments has been implemented. Now, it is easy to set up experiments and analyze the results. The following PRs contributed to this most of all:

• ExperimentAnalyzer to analyze the results of experiments on the go (#181).
• ExperimentLauncher to launch experiments and save the result in the format required by ExperimentAnalyzer (#213).
• HistoryCollector and AgentTrainer to collect histories and use it to pretrain agent for further experiments (#138).

Then several improvements concerned with MABs and molecules usage example:

• Stable reward for MABs which allowed not to use the raw fitness improvement since it can deteriorate the algorithm’s
  robustness (#148).
• Molecule encodings for contextual MABs were added, so now it is possible to experiment with different encodings for molecules (#179).
• Save and Load options for MABs to plot the mutation probabilities throughout the evolution process (#185).

Also, a couple of fixes were made:

• #227
• #212
• #223
• #195
• #182

0.3.3

Minor release with several bugfixes to simplify installation.

Highlights:

• #156
• #163
• #164

0.3.2

Hi, AI folk! We're releasing new minor version of GOLEM with several important improvements to the core algorithm, visualizations and experimental features.

New Use-case:

• Added Molecule Search example for evolutionary drug design. It uses Guacamole Benchmark for comparison with other SoTA methods. PR #83

Core evolutionary algorithm enhancements:

• Implement checks during evolutionary cycle that help supporting structural diversity of graphs in the population. See PR #131 for details.
• Implement ReproductionController for more stable and diverse populations. PR #104

New Features:

• Added 2 new alternative integrations for hyperparameter optimization, IOpt and OpTuna
• Added new plots for vizualising population diversity! Now you have 2 more kinds of vizualisations. Details are in description of PR #113
• Added new plot MultiLineFitnessPlot for comparing results of multiple experiments where each has a number of trials. As part of #102
• Added possibility to load optimization histories from dependent projects in GOLEM. This enables you to use Fedot.Web project for visualising optimization results in the web. PR #120
• Extended experimental Adaptive Evolution Agent by adding Neural Contextual Bandits. They help to choose best mutation given a context (i.e. graph embedding). This feature is yet experimental and not optimized. PRs #102, #129

Various bug fixes and minor improvements not counting.

Check it out!

Release 0.3.1

Hello, folks! This is a time for a minor release of GOLEM coming with several new features and some incremental improvements.

Highlights are:

• Added several crossovers that exchange edges between graphs, useful for Bayesian network evolution (#58)
• Added Sensitivity Analysis for graphs that check how important certain nodes and edges are for the fitness of the graph (#62)
• Graph visualisations now include label viz (used in particular in Sensitiviy Analysis) #76
• Add new example that implements Tree evolution using fast Edit Distance metric (#60), check it out in tree-search.py
  

Fixes:

• #67
• #75
• #84
• Usability fixes from #86

Release 0.2.0

This release marks several enhancements:

• #14 -- Added integration test
• #21 -- Added Tuner interface for tuning arbitrary graphs with a help of HyperOpt
• #30 -- Fix several problems with adaptive population size
• Several bug & dependency fixes
• Merge latest changes from FEDOT core: add MemoryAnalytics for debugging eval process & OptGraphBuilder for convenient buliding of small directed acyclic graphs)

First release

This is a first release after separation from FEDOT framework