Hal-cgp

This is the third major release of the hal-cgp library.
It contains the following changes:

New features
- Introduce callback to initialize individuals of the initial parent population 253
- Support custom atomic operators 254
- Provide support for hurdles during evolution 259 , 280
- Local search that does not require differential objectives -- evolution strategies 262
- Improved caching decorator 263
- Multiversion doc via readthedocs 315

API changes
- Make levels_back parameter optional 281
- Allow custom functions to be used for computing cache keys 284
- Redesign signatures of compilation targets 294
- Replace min_fitness by termination_fitness 295

Bugfixes
- Reorder changes phenotype for individuals with parameters 265
- Cache decorator: use all args and kwargs when computing fec cache key 272

Maintenance
- Use a persistent process pool instead of a context manager for multiprocessing 260
- Reset parameters when the corresponding node is created by mutation of a genome 264
- Rename input -> address for address genes in hidden/outp… 267
- Check tournament_size <= n_parents in hl_api 268
- Move mutation_rate into EA 269
- Remove support for calling graph directly 270
- Fix error message wrong size passed to to_torch/to_numpy graphs 273
- Remove outdated n_processes argument in hl_api 275
- Add annotation in Individual 293
- Bump pillow version 291
- Improve description of design/use cases 292
- Port CI from travis to github actions 296
- Add test for n_columns=0 299

Contributors (in alphabetical order)

Jakob Jordan
Henrik Mettler
Maximilian Schmidt

This is the second major release of the hal-cgp library.

It contains the following changes:

New features
- Provide possibility to restrict local search to `k` best individuals 160
- Track parent id in offsprings, i.e., cloned individuals, to allow the reconstruction of full lineages 168
- New logic for mutation of genes: instead of mutating the same number of genes in each offspring, the number of mutations is now sampled from a distribution 180
- Allow multiple parameters for parametrized nodes 191
- Sphinx-powered documentation 181 , 188 , 189 , 200 , 205 , 218 , 220 , 223 , 234
- Extend cache decorator with functional equivalence checking 207
- Use norm clipping instead of value clipping to avoid changing the direction of the gradient in local search 210
- Make sure user-set attributes of individuals are copied to offspring to allow consistent recording 222
- New node: if/else operator to support, for example, piecewise target functions 231
- New Examples: 227 (minimal example), 230 (using multiple genomes), 231 (using if/else node)
- Add option to limit number of fitness evaluations instead of number of generations 238
- Add functionality to reorder the nodes in the parents genome(s) before creating offspring 239

Bugfixes
- Avoid parameters being reininitialized in `Genome.mutate()` 213
- Fix broken `Node.repr` 233

Maintenance
- Improvements in test suite 161 , 171 , 176 , 179 , 190 , 193
- Simplification of internal routines 169, 170
- Simplification of the `Node` class 178
- Removal of crossover operation and breeding pool 195
- Remove unused function `_is_active_input_gene` 211
- Use "logical and", and reuse existing functions in helper functions of `Genome` 212
- Sort imports consistently 219

Contributors (in alphabetical order)

Jakob Jordan
Henrik Mettler
Maximilian Schmidt

This release fixes a bug which led to wrong numbers of mutations in mutated genomes (see 156 ). The bug was introduced in https://github.com/Happy-Algorithms-League/hal-cgp/commit/91b62d5e678917a0b61ed8eda4b7dd869da48ed2 and fixed via #157 .

Furthermore, this release contains the following changes:
- Simplify installation of individual extra requirements (145 )
- Refactoring: the `mutate` function was moved from the `IndividualBase` class to the `Genome` class. (153 )
- New example solving the `MountainCarContinuous` environment in the `OpenAI` gym (149 )

Contributors (in alphabetical order)

Jakob Jordan
Hendrik Mettler
Maximilian Schmidt

This is the initial release of the hal-cgp library.