Skip to content

Latest commit

 

History

History
356 lines (244 loc) · 11.2 KB

NEWS.md

File metadata and controls

356 lines (244 loc) · 11.2 KB
Raw

News

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog.

[3.0.0] - 2024-04-05

Project is now in maintenance mode. Occasional bug fixes and security patches will still be issued.

Development efforts will concentrate on Ultra which is our new evolutionary algorithms framework with all the wonderful bells and whistles you could only hope for...

So far it's in (pre) alpha state but please take a look at it, I'd love your feedback on it.

Added

  • Support for concurrent access to the fitness cache (hash table). The current solution uses a global (shared) lock.

    This approach is simple and allows preliminary parallel search extensions; furthermore performance of single-thread code isn't affected. It lacks granularity.

    In future an array of mutexes could be used to minimize contention.

Changed

  • BREAKING CHANGE. hits and probes statistical information were hardly used and have been removed. The removal implies a (minor) change in the saving format of the cache.

  • BREAKING CHANGE. gene::arg_locus has been renamed to locus_of_argument. This should only affect secondarly the end user since the function is mainly used in the meanderings of the i_mep class.

  • BREAKING CHANGE. Implementing user defined functions is simpler. Instead of:

    value_t eval(core_interpreter *i) const final
{
  const auto arg(static_cast<interpreter<i_mep> *>(i)->fetch_arg(0));
  if (has_value(a)) { /* do something with `arg` */ }
  // ...
}

    you can simply write:

    value_t eval(symbol_params &args) const final
{
  const auto arg(args[0]);
  if (has_value(a)) { /* do something with `arg` */ }
  // ...
}

    Note:

    • symbol_params gives access to a subset of the interface of a core_interpreter so the user doesn't have to know unnecessary details about interpreters;
    • symbol_params::operator[](unsigned) is equivalent to symbol_params::fetch_arg(unsigned) (syntactic sugar).

2.1.0 - 2020-11-23

Added

  • Base support for user-provided datasets. Part of the classes/functions for symbolic regression can now work with a generic STL container. The new features are described in the extended symbolic regression example.

Changed

  • sum_of_errors_evaluator now uses a running average technique (numerically more stable). This could produce minor differences in a specific test case (because of floating point arithmetic) but the general behavior doesn't change.

Deprecated

  • Support for Genetic Programming with Adaptive representations will be removed and replaced with (a sort of) Straight Line Program structure. The new approach is simpler and fits easily in the current architecture.

2.0.0 - 2020-11-02

Added

  • New dataframe::front() public member function which allows simple access to the first element of a dataframe.

  • dataframe::read_csv now supports a header row. Also the output column can be specified (before it always was the first column):

    dataframe d;

dataframe::params p;
p.has_header = true;
p.output_index = 2;

d.read_csv("filename.csv", p);

Changed

  • BREAKING CHANGE. Sources require a C++17 compatible compiler.

  • BREAKING CHANGE. The interpreter class performs calculation using std::variant insted of std::any.

    This gives a performance improvements without important limitations of use cases.

  • BREAKING CHANGE. dataframe::get_column and dataframe::columns have been removed / merged. The dataframe::columns public data member contains information about all the columns of the dataframe.

    So:

    1. d.columns.size() replaces d.columns()
    2. d.columns[3].name replaces d.get_column(3).name
    3. d.columns[3].category_id replaces d.get_column(3).category_id

1.1.0 - 2019-12-11

This is the last C++14 compatible version. Next versions of the framework will assume a C++17 compiler.

Changed

  • BREAKING CHANGE. Fluent interface for the search / src_search / ga_search classes.

    Insted of:

    search<...> engine(...);
engine.set_training_evaluator<...>(...);
engine.set_validation_evaluator<...>(...);
engine.set_validator<...>(...);

    you should write:

    search<...> engine(...)
            .training_evaluator<...>(...)
  		  .validation_evaluator<...>(...)
  		  .validation_strategy<...>(...);

    This benefits the class user by reducing the amount he has to type and makes the interface more natural.

  • BREAKING CHANGE. Removed read_csv(filename) / read_xrff(filename) from the interface of the dataframe class.

    Both the member functions can be replaced with the already existing dataframe::read(filename) which uses the filename extension to identify the data format.

  • Clearer format when exporting individuals.

    Part of the redundant parentheses has been removed.

    So, instead of ((sin(X1)) + (ln(((X1)+(X2))) you get (sin(X1) + ln(X1+X2))

1.0.0 - 2019-03-22

Added

  • Simpler way to setup Genetic Algorithm / Differential Evolution tasks. We now have the ga_problem and de_problem facades. Their constructors cover the standard situation of a sequence of uniform parameters (possibly in distinct numeric intervals).

    Examples are been rewritten to take advantage of the new classes.

Removed

  • problem::chromosome help function. It was a bit of a stretch, now instead of:

    class ad_hoc_integer_like_class { /* ... */ };

vita::problem prob;
prob.chromosome<ad_hoc_integer_like_class>(nr_of_parameters);

    you should write:

    vita::ga_problem prob(nr_of_parameters, {min, max});

  • standalone Makefile-based build system. Now the only supported build system is CMake. This greatly simplify multi-platform testing and documentation.

Changed

  • BREAKING CHANGE. GA-individuals match integer vectors. This allow a faster management, a simpler access to the genome and simpler code.

    Where we wrote:

    genome[i].as<int>()

    we now have:

    genome[i]

  • BREAKING CHANGE. Completely revised validation strategies. The search class now allows specifying two evaluation (a.k.a fitness) functions via the set_training_evaluator and set_validation_evaluator methods.

    This:

    • gives users the option to adopt distinct training / validation strategies;
    • allows a simpler implementation.

    More details are contained in the the validation.md wiki page.

  • BREAKING CHANGE. Removed switch to default initialize an environment object.

    So, instead of:

    environment e(initialization::standard);

    you should write:

    environment e;
e.init();  // initialization

    Usually the initialization is only needed in debug code. The common approach is:

    • left parameters in the environment struct unitialized (auto-tune);
    • leave to the search::run() function the task of identifying the "best" values.

  • BREAKING CHANGE. dataframe::load... renamed to dataframe::read... for uniformity with Pands' dataframe.

  • BREAKING CHANGE. example::tag method has been renamed to label; example::cast_output has been renamed to label_as. Both are now free functions.

    The new naming improves consistency and adopts standard terminology.

  • BREAKING CHANGE. Simpler method to specify GAs parameters.

    Instead of:

    for (unsigned i(0); i < n_jobs; ++i)
  prob.sset.insert(ga::parameter(i, {-0.5, 23.5}));

    use:

    // problem prob;
for (unsigned i(0); i < n_jobs; ++i)
  prob.sset.insert<ga::real>( range(-0.5, 23.5) );

    or equivalently:

    // de_problem prob;
for (unsigned i(0); i < n_jobs; ++i)
  prob.insert( range(-0.5, 23.5) );

    or, even better:

    de_problem prob(n_jobs, {-0.5, 23.5});

    Note that:

    • the parameter index doesn't appear anymore;
    • the range is clearly stated.

    The same approach applies to GAs.

    The ga::parameter help function has been removed.

0.9.14 - 2017-12-11

Added

  • Code of conduct for contributors.
  • Complete control of the generational-gap between subset-selection (DSS algorithm).
  • Simpler setup for GA problems.

Changed

  • symbol_set::roulette() function has the same probability of extracting a terminal / a function regardeless of the specific symbols' weights.

    If all symbols have the same probability to appear into a chromosome, there could be some problems. For instance, if our problem has many variables (let's say 100) and the function set has only 4 symbols we cannot get too complex trees because the functions have a reduced chance to appear in the chromosome (e.g. it happens in the Forex example).

  • BREAKING CHANGE. Enable/disable logging of specific information is a bit simpler.

    Instead of

    env.dynamic = true;
env_dyn_name = "dynamic.txt";

    this is enough:

    env.dynamic_file = "dynamic.txt"

  • BREAKING CHANGE. i_mep::compress function renamed to i_mep::cse (Common Subexpression Elimination).

  • BREAKING CHANGE. Serialization file format for cache and fitness has changed.

    The new format produces files up to 20% smaller omitting:

    1. the seal value which is assumed equal to the cache::seal value
    2. the fitness vector size (and assuming that all the components of the vector are on the same line)

  • BREAKING CHANGE. Instead of:

    symbol_set sset;
environment e(&sset, initialization::standard);

evolution<i_mep, std_es> evo(env, *eva);

    you should write:

    problem p(initialization::standard);

evolution<i_mep, std_es> evo(p, *eva);

    The problem object aggregates an environment and a symbol_set (which are almost always used together).

    This also permits to remove the pointer from the environment to the symbol_set.

  • BREAKING CHANGE. Replaced the search::set_evaluator method: Instead of:

    std::unique_ptr<E> eva(f);
std::search engine(prob);
engine.set_evaluator(std::move(eva));

    you can write:

    std::search engine(prob);
engine.set_evaluator<E>(f);

  • Uniform output results for make / cmake build systems.

0.9.13 - 2017-09-15

Added

  • New printing system for i_mep individuals. You can now export C / C++ / Python / MQL code.
  • Self-adaptive crossover operator.
  • Forex example.

Changed

  • More robust implementation for real functions.
  • New uniform crossover implementation.
  • ALPS can remove multiple identical layers.

Fixed

  • Various compilation warning with recent version of gcc (7.x).